ai-art-embeddings
Cafe's TI Repo Colab and Kaggle Implementation. (6x faster training compared to the webui)
Revamped Cafe's TI repo Colab implementation by Andite.
It should be much more cleaner now, and I included some instructions in it. The rest of it are common sense though.
This training repo is 6x faster compared to the webui since the TI code from the webui is bad compared to the original TI code.
Colab:
https://colab.research.google.com/drive/1bbtGmH0XfQWzVKROhiIP8x5EAv6XuohJ?usp=sharing
Kaggle Implementation by <>.
Kaggle:
https://www.kaggle.com/code/xynonners/stable-textual-inversion-cafe-kaggle-public/notebook
Original repo by <> if you want to train locally:
https://github.com/cafeai/stable-textual-inversion-cafe
maybe pin? <>
i can't see it..
<:ReisenBlank:958503452047196191>
Ohh maybe lacks of permission?
seems like it
i'll ask.
there we go.
Hello, I have a question to restore the training in the "last_embed" part, do I have to put the last one, for example, embeddings_gs-8900.pt..?
yes
its gonna retrain starting from 0 steps, but dont worry about that. its not gonna literally start training from **zero**
just take it as `steps of ur last embedding + steps u retrained`
I mean, it's being released again, and how can I use the style? Can I choose the various embeddings_gs that it generates and rename it and use it?
Yes, you can rename it and use it.
you can use any of the embeddings it generated, pick the best one by testing them out
Thanks for the colab, but why no xformers?
cafe will add xformers soon
also limit the minimum vram requirement from 12 gb to 8 gb
Cool thx, I was just curious.
I just ran
%pip install -q https://github.com/metrolobo/xformers_wheels/releases/download/1d31a3ac_various_6/xformers-0.0.14.dev0-cp37-cp37m-linux_x86_64.whl
did it work?
yes
oh, i might implement that in the colab if that's the case.
the non-pro colabs run on T4s
i'll credit you for it
no need
It's from shivs repo
precompiled for T4 and P100
wait, does this mean I can make TIs locally (not colab) with only 8gb vram?
soonβ’?
yeah, if u have 8 gb vram.
yeaaaaah
so i imagine this will be done without Automatic1111's web ui
ye, but cafe alr did a pull request iirc
not merged yet tho
Wow, this works super well. I'm training an OC and getting way better results than I had with automatic1111 TI. <:AunnHob:977014561188229191>
Arigatou~ Cafe-sama!
Hmm how is this different from webui
Because I made some pr changes for webui and I want to fix as much as I can
wait 225 tokens per step??
rip auto
I think cafe already made a pr in webui, dunno
another question in "initializer_words" what would be the specific list to put? is there any limit?
no limit
anything really, since initializer words are something you pull from the model, in my understanding
as long as its trained in the model
anyone know what the difference is between these "initalizer_words"
and the "initalizer_word" that goes beneath the project name?
Tried this colab, looks like it stops training at 9700 steps, but still keeps working and generates some images
colab has a runtime limit
What should I doοΌ Restart or something else?
use another email
I think you didn't get me correctly, colab keeps working (I'm not out of limit yet), just instead of generating new checkpoints I constantly see message about DDIM SAMPLER progress bar, that resets every 50 steps
usually happens when ur in a near breakthrough to a new epoch
keep in mind not to overtrain
Ah, got it, now I see a message about "Validation dataloader" progressing (currently at sixty percents)
By the way, I multiple times saw word "epoch" when people talk about ai, what does epoch means?
it was not understood very well, You must follow the instructions, if it does not mention it, it is because it does not have to be modified, you could try.
hello, getting this type of errors in kaggle repo when trying to start training, any ideas?
```RuntimeError: Error(s) in loading state_dict for LatentDiffusion:
size mismatch for betas: copying a param with shape torch.Size([1000]) from checkpoint, the shape in current model is torch.Size([4000]).
size mismatch for alphas_cumprod: copying a param with shape torch.Size([1000]) from checkpoint, the shape in current model is torch.Size([4000]).
```
<>
ping<:NotLikeKogasa:996413933449650218>
he needs help with da kaggle
idk how kaggle works so
are you trying to resume training or train new?
also what model are you using to train?
trying new training, custom merge model combined of gape+sd1.5+f222
seems like a model error to me
agreed, doesn't seem to be a kaggle issue
does it work with any other model?
and does the model work somewhere else (e.g. colab)?
gonna try that
well, on google colab it is working
I wish auto1111 made the train tab an extension and included that
welp pretty odd
you sure you selected the correct model path and stuff
(though you prob already just using colab at this point lol)
it would still suck
lol
Why?
hope so <:MomijiStare:1037138362248728576> , now it is giving those errors on kaggle
```RuntimeError: Error(s) in loading state_dict for LatentDiffusion:
size mismatch for betas: copying a param with shape torch.Size([1000]) from checkpoint, the shape in current model is torch.Size([2000]).
size mismatch for alphas_cumprod: copying a param with shape torch.Size([1000]) from checkpoint, the shape in current model is torch.Size([2000]).
size mismatch for alphas_cumprod_prev: copying a param with shape torch.Size([1000]) from checkpoint, the shape in current model is torch.Size([2000]).```
either it is something with model, or with config, or i dunno
so, finally managed to launch it on kaggle
just dont touch this `timesteps: 1000` and name dataset without spaces
maybe a problem with the environment
oh, you got it running?
yes, just right now
huh
changing the timesteps in config works in colab but not kaggle?
did you try cat'ing the file after it was written
maybe something weird to do with %%writefile
and great that you got it running π
Cat-ing? Wdym? <>
cat
linux command
dunno what to do with it <:dhob:977013890745499648>
and it stops training when it hits step counter, not continuing with next epoch, same with continuing training
but overwriting file goes without any problems i guess, it throws this in output log on config rewrite block
```Overwriting /kaggle/working/stable-textual-inversion-cafe/configs/stable-diffusion/artstyle.yaml```
yeah that's expected
no idea what the issue could be tbh π€·
the commands are literally mostly the same as the colab so
so it stopped at 1000?
i changed steps per image (from 100 to 500), and cap is 10020
this one line
so it's working?
and if you change that line it fails only in kaggle
if i change this line, then yes, it does brake
this works perfectly rn, attempted to continue training after finishing first
well as long as it works lol
not at my computer so can't test anything
Timesteps should always be 1000
That's a fixed hyperparameter used in ddpm. You shouldn't change it
what are you trying to do?
are you trying to change preview and checkpoint generation from 100 to 500?
that's the one you're supposed to edit
if you want to change it
every_n_train_steps for checkpoint
batch_frequency for preview gens
the repeats should be at 100, don't change that.
I want to increase steps per image
Has anyone tried the sd 2.0 version?
It's going to need modification to work with SD 2.0, it's not configured for it.
Got a question about training, it is normal when you continue training it is resetting epoch count from any to 0? Using kaggle
yes
it starts from 0, but not literally
it continues where it left from
<> From that question earlier, do you think you'll be making a version of this for SD 2.0?
I think cafe would have to implement that
^
i don't make the repo, i only implement or make them useable in colab
<:KoishiSip:1018748040200331264>
Ah ok fair enough
I found this it looks very new https://github.com/giannisdaras/multires_textual_inversion
Is it possible to disable dataset mirroring in the colab? I'm trying to train a character that isn't symmetric
it shouldnt matter
<:thonk:365312925789061120> i have all the reqs and can run the original textual inversion repo but not this one, i get this message at the validation sanity check
`python main.py --base configs/stable-diffusion/v1-finetune_lowmemory.yaml -t --no-test --actual_resume "C:/Users/ctd/stable-diffusion-webui/models/Stable-diffusion/model.ckpt" --gpus 0 --data_root "C:\sdtraining\test\2\3" --init_word "illustration" -n "style3"`
I think it did for me, I was training a character with a hairclip on the right side, but the trained output have 2 hairclips on both sides
Do I need to stay in colab page during training ?
yes
just watch out for the runtime limit
It's because of symmetry
When should I stop training, are there some specific indicators ?
when it seems like its degrading further every steps
or when it doesn't improve anymore
u need to do alot of testing
by prompting them.
6k step embed could be better than 7k. or the opposite
handpick the embeds according to which image preview looks the best
i'd pick 2-3 embeds that looks the best and compare them to each other
I am using lr 0.001
and it's 35k step now
it's honestly up to the user when to stop, just don't overcook it
Β―\_(γ)_/Β―
35k step is overload
idk how to measure 0.001 so, that's up to you
I don't know, it still looks good
ya better test the embeds while its training, they may be overfit
how to know it's overfit
big change in low denoise strength ?
if its hard to edit it
via prompts
not that flexible or smth
have to use alot of weights to edit the prompt with the TI
if it takes 3-4 sets of parentheses to change smth, thats overfit
π₯² I'll stop it after the epoth (9)
epoch 9, damn
u dont even need epochs for this repo
lol
u do u i guess, just don't overcook it. test em out
use a separate gmail and tab
for the webui colab
emm makes sense
what do you think is the minimal number for training a style
artists usually have different style in different periods .The one I am training, I can only find around 15 images with very similar style, and add another 30 images which are kind of similar.
i recall having a step mini guide
in the colab
few datasets require lower steps though
where can I find prompts used for scaled images ?
i keep getting `Error: credential propagation was unsuccessful` and this screen when trying to allow access to my drive, anyone got a fix?
oh yeah, same here.
are you using an extension such as clearurl
it removes the required login tags from the url when redirected to account page
have to disable it if u are
oh right, im not at my desktop anymore but ill try removing it tomorrow, thanks
honestly I can't believe I tried lots of shit like disabling strict tracking protections and several addons but never thought about switching browsers or turning every addon off
Which folder contains the training log images, `logs/.../images/train` or `logs/.../images/val`?
train
val is after training finished
both folders contain images during training tho
val generates some images after each epoch
uuuuhh so I tried it out and it didn't seem to increase the training speed; might just be because I'm using camenduru's wheel instead though. Might need further testing...
^^ I got about 1.00s/it as for speed both with and without xformers on T4 Colab.
It won't do anything until Cafe doesn't integrate xformers itself into the trainer repo
colab broken
run it locally for now
or use kaggle
huh how did you break it
i guess it's because of the new pytorch version
I've made quickly a new codeblock, which I've run after the "crash the colab" block.
```
!pip install torch==1.12.1+cu113 torchvision==0.13.1+cu113 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu113
!pip install torchtext==0.13.1
```
It runs fine after this.
I'm sure there's a better fix than this but it does the job.
<>
i see, let me try
colab works fine now
thank you! <>
https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/3958
seems like the training from the webui is also broken, i guess i wasnt alone lol
haha
oh nice
what's the "nai-wd" model preloaded in the colab?
a merge of nai and wd at 0.2
its my personal training mix
> personal training mix
so do i actually need to replace this with a link to whichever model i use or can i leave it at this? does not using the target model the TI will be used on have any visible effects?
figured as much; hoping to see xformers so that we can get even faster training speeds soon
also why does this Colab work so much better than the one on the webui? the webui textual inversion gave me unusable results lol pmuch
leave it to that. it works on any model. works the best on an anime model tho ofc
when sample_scaled_gs got generated, is it normal for it to look squished?
no <:restrain:1006306907994202132>
π
is it safe to assume this still doesnt work with the anything model?
not really a problem with the repo, anything just sucks for training for some reason
cool
I'm trying to run the colab version of this and i'm using the latest version 12/8/22. I'm getting this error on the training step. Anyone suggestions on fixing it?
what model are you using for training?
I was using a custom blend. I tried again using a base model and that's working π€
if it had anything from anything then it's probably better that it didn't start
I kinda need help with this Kaputii style TI
the samples in google drive at 8000 steps looks way better than when I use the 8000 step checkpoint for some reason, is there something I'm missing?
TI also attached so it can be tested
model used for training was the default one in Cafe's repo colab
https://huggingface.co/andite/models/resolve/main/nai-wd.ckpt
what are ur prompts
`kaputii`
Negative prompt: `lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry`
yeah u gotta use painting of
and the keywords used for those sample scaled images
I have no idea what the keywords are lol
theyre in the yaml you used for training i believe
^
the prompts used
to generate the previews
oh shit thanks
oh wow
nope still getting that deep fried / overexposed look
`a cool painting in the style of kaputii`
do I really need to use the model used to train to get the same results <:NotLikeKogasa:996413933449650218>
testing various checkpoints of the TI with Anything v3 just to see if it was overtrained
also tested clip skip 1-3
still no idea why i'm not getting the same colors as the samples on google drive
after some more testing, DDIM looks significantly better than Euler and dpm++2m, why?
<:NotLikeKogasa:996413933449650218>
ok looks like the real problem was my local install for some reason, at least that means I can probably release the TI now
same seed, same vae, same clip skip 2, but the google colab one is using a 4gb anything v3.safetensors file while my local is using a 2gb one
[update] realized this TI is extremely sensitive to CFG scale, the best colors happens at less than 8
<>tried reducing weight of your TI in prompt?
might help with overcooked look
Yeah already tested that, didn't help
The real solution was to use lower cfg to get better colors, I really like the results with cfg 6
is there any way to disable mirroring at all?
my trained character is still getting duplicate (mirrored) accessories
Can I ask quickly if there's a method to batch replace a transparency layer of several with a single colour like white? I've tried turning it into a jpg and that didn't work as intended. I want it so that's it's similar to the action of putting a white colour layer under a transparent image in photoshop
^and also if the TI can deal with transparent images well compared to webui training (which is poopoo)
how many images used to train do you recommend?
at least 50, ideally 100
I'm trying to train a TI on my system, but every time I try to run the command, I get an error. I am using a RTX 3090 and this is the error I get when using `...v1-finetune.yaml`:
```
RuntimeError: CUDA out of memory. Tried to allocate 3.21 GiB (GPU 0; 23.68 GiB total capacity; 16.63 GiB
already allocated; 2.49 GiB free; 18.66 GiB reserved in total by PyTorch) If reserved memory is >> allocated
memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and
PYTORCH_CUDA_ALLOC_CONF
```
I am running this command:
`python main.py --base configs/stable-diffusion/v1-finetune_lowmemory.yaml --train --no-test --gpus 0, --data_root "/mnt/Database/ganyutrain/" --init_word "ganyus" --name "ganyu_set_testX"`
following that, I set `max_split_size_mb` to both 512 and 1024, but still get the same crash
Is there anything I can do to fix that?
this is the error when using `v1-finetune_lowmemory.yaml`:
```
IndexError: Caught IndexError in DataLoader worker process 0.
Original Traceback (most recent call last):
File "/home/alexander/anaconda3/envs/ldm/lib/python3.8/site-packages/torch/utils/data/_utils/worker.py",
line 302, in _worker_loop
data = fetcher.fetch(index)
File "/home/alexander/anaconda3/envs/ldm/lib/python3.8/site-packages/torch/utils/data/_utils/fetch.py", line
49, in fetch
data = [self.dataset[idx] for idx in possibly_batched_index]
File "/home/alexander/anaconda3/envs/ldm/lib/python3.8/site-packages/torch/utils/data/_utils/fetch.py", line
49, in
data = [self.dataset[idx] for idx in possibly_batched_index]
File "/mnt/Database/stablediffusion/stable-textual-inversion-cafe/ldm/data/personalized.py", line 84, in
__getitem__
text = random.choice(self.templates).format(placeholder_string)
File "/home/alexander/anaconda3/envs/ldm/lib/python3.8/random.py", line 290, in choice
raise IndexError('Cannot choose from an empty sequence') from None
IndexError: Cannot choose from an empty sequence
```
batch size problem in config file
set it to 1
is there any way to merge multiple .pt's into one file?
there's a py file where u can merge embeddings in the repo
no cell for it though, you'll have to use that manually
cause it's a niche option
Thats all good, thx!
in the google drive, where do you find the model?
do you have to upload it?
any way to confirm a .pt file is safe to run?
how did everyone gather enough dataset to train with? Ive searched through danbooru to collect but the style for each pic was so vary to be able to gather 100+ pics with consistent style
Why just danbooru?
Revamped Cafe's TI repo Colab implementation by Andite.
It should be much more cleaner now, and I included some instructions in it. The rest of it are common sense though.
This training repo is 6x faster compared to the webui since the TI code from the webui is bad compared to the original TI code.
Colab:
https://colab.research.google.com/drive/1bbtGmH0XfQWzVKROhiIP8x5EAv6XuohJ?usp=sharing
Kaggle Implementation by <>.
Kaggle:
https://www.kaggle.com/code/xynonners/stable-textual-inversion-cafe-kaggle-public/notebook
Original repo by @Cafe - γγ΅γ if you want to train locally:
https://github.com/cafeai/stable-textual-inversion-cafe
Reposting since it got buried.
pin it
cant
working on a new colab which automates a lot of stuff. basically after this it would train the TI automatically, stops when finished, etc
it calculates the token count, max steps based on the image count
will be good for beginners and lazy people
from someone who belong to the latter
i approv
<:Dumbass:1052329551486787594>
<>
I'm done, but I haven't tested everything yet if it works correctly. If you have some time can you please check it out? Feel free to modify anything if it's needed.
https://colab.research.google.com/drive/1ouTImTpYkBrX5hiVWrzJeFtOyKES92uV?usp=sharing
I've changed some folders, so you have to clone your repo again. Before that, make sure you backup your logs folder. I didn't include it because I'm using a more persistent location for it.
i'm in mobile rn unfortunately, but ill test it out later in the morning <:AYAYAYA:978879333248667718>
i also made a colab, mainly to add support for rclone so i can use whatever cloud provider i want
https://colab.research.google.com/github/jamesWalker55/cafe-ti/blob/main/cafe-ti_jamm.ipynb
rclone gud
it's great for when you reached colab's limit and switch to a different account then resume training
ikr, i was using it already
also, good if you dont want google to watch your images
think ill have to update my training mix
since wd 1.3.5 came out
wait, with this repo I don't have to install the TI thing in my drive?
yeah, i moved the TI thing to the temporary storage of colab; your drive should only store training images and training outputs
okok, just doublechecking
and how does rclone work? never heard of it before
it basically lets you manage files on cloud services
and it comes with a `mount` command that mounts a cloud service locally
but setup is slightly complicated, have a look at their docs
https://rclone.org/
Is there a way to avoid having to reinstall python packages everytime?
for the collab.
yes. packaged it, and ship it somewhere else for future use. Like what I did with my one-click colab:
https://colab.research.google.com/drive/1ptUE6EB01UkONCczGqMdlzmk2K3Z2XWI
oh, tomorrow I'll have a better look to this.
it would be nice to have everything in one gdrive even when switching account
thank you
In the output folder for images there is reconstruction, sample and samples_scaled, how should I be interpreting the results from these images?
the gamples ones look extremely bad compared to samples_scaled
my g
this thing perfect for me
what does it do?
has anyone been able to convert .pt's to huggingface, i've got a script that i've made that should theoretically work but doesnt, i keep getting torch.expand errors when assigning the embeddings to the token
i think i might've gotten it
nvm
did the long overdue update for kaggle notebook
Does the repo train on images one at a time? Can it train on several at once? I saw some in <#1044620915377451058> that trains on batches
You can change the batch size but then you'll need colab pro and an A100
~~I see, but is that configurable in the `.yaml`?~~
Nvm I saw it instantly, line 82
anyone recommended settings to minimize compute token loss and training time to ~10k steps?
https://colab.research.google.com/drive/1ouTImTpYkBrX5hiVWrzJeFtOyKES92uV?usp=sharing
added some more optimizations to the config.yaml creation
should be much-much faster than the previous version
for this one I just need place the link of the folder with the images in the data set line, choose between artstyle or character and room every other cell as default?
Nevermind did not worked
Probably you need to redownload the repo by running that block again. Make sure you backup your current logs folder
No, it doesn't support it yet. I'll do it later tho, but don't want to make it too difficult. It has to be a Google drive link or a hugging face url like in andite's Colab when you download the dataset
<> would you mind if I stole the code from your colab? just took a look and would possibly be a nice thing to have on kaggle instead of the weird %%writefile stuff it has now
How dare you
Ofc you can
haha okie
christmas update incoming, complete with mashiro
Added support for using path instead of drive/huggingface url
I was able to start training the AI yesterday
And I was really happy about! But end up being not a hypernetwork but UI training
https://tenor.com/view/tsugumi-nisekoi-cry-sad-gif-7395114
kisamala!
Also Idk if a UI training is better then a hypernetwork but was just trying to have the ai to replicate both the specific artstyle with the good results that comes from anythingV3
this colab trains Textual Inversion embeddings, not hypernetworks
hope there are plans to update this repo for v2 π (and safetensors too bc i can't for the life of me find a script to convert it *to* ckpt lol)
We need more <:5Head:945662780852297768> people
hmm
Any video tutorial to create embedding from that colab? I got crashed
Crashed on this step
I thought it's supposed to crash there
it was supposed to do that
i have no clue how python works so i dont know how to fix this, it might be some stupid simple error though :(
Read the cell's title please
It is because after some package install, you have to restart the jupyter kernel but there's no other way on Colab to do it, only just by crashing it
syntax error in this cell, i think you mixed up another language on this `if` block
Faking python
Ty I'll fix it
There should be a :
Also after the elses
also your yaml loader is really broken
I changed your `yaml.load(f)` to `yaml.safe_load(f)`
Thanks, yeah I was using an old method
Worked fine on paperspace
hm, the training cell's `if` blocks has the same problems too haha
Ye already fixed
aight pog
Thanks for the help
npnp
So I am getting this error now and idk exactly why was able to train without any issues a while ago but now it just gives this no matter what I do anything to fix this or what's the issue here
Probably that an image has some funky transparency?
~~I found a bug on the first run of your colab, <>, `zip` doesn't create the `/outputs/` folder, I changed the last lines of the train cell to this:~~
```
%cd {logs_folder}/{newest_folder}
!mkdir -p /content/drive/MyDrive/sd_text_inversion/output/
!zip /content/drive/MyDrive/sd_text_inversion/output/{project_name}.zip checkpoints/*
print("Done! The checkpoints is saved in the output folder.")
```
actually, nvm i didnt run the "Get the repo for textual inversion." cell... my bad
ye maybe I should highlight that this is not the same repo structure
and running that cell is recommended even if you used andite's colab before
thanks for the notice
guys i think im stupid
im trying to make the colab work on like colab storage but this keeps on showing up
the error message doesnt tell me where its happening as well so that i can fix it :(
π
(the reason why im trying in the first place is because creating an account needs a phone number now and mine is unfortunately tied to 2FA π and im out of storage)
ok it just suddenly works i just stopped the runtime and reran it
sorry guys
Is that a thing?
Neither of the images were transparent
Do I need to upgrade my Google Collab or I can use the free version of the Google Collab for this?
Works on free ofc
Nice π
I run it on a laptop. Then, i just left it for about 5 hours. When I come back, the colab is timeout & nothing on the output folder.
check the logs folder, it should have `embdeddings_gs` files on it
I deleted all my files in Google Drive for fresh restart π
At 83% now <:Agony:996544918610784296>
Maybe the cause is the model? What models are you guys using?
default
I made the last cell for this case
Dang, safetensor is still not compatible...
use nai-wd.ckpt
most models still based on novelai so the TIs will work fine
where can i find it?
just a few cells up
there's a block which downloads it from huggingface
https://huggingface.co/andite/training_models/tree/main/nai-wd.ckpt
There's two.
meant this
oh, i changed the link in the cell so i didn't know.
use andite's link
Make training cell ignore directories?
It errors when it detects .ipynb_checkpoints folder that generates inside `/content/drive/MyDrive/sd_text_inversion/Imagesfortraining`
``!rm -r /content/drive/MyDrive/sd_text_inversion/Imagesfortraining/.ipynb_checkpoints``
create new cell and paste that code in. run it
then rerun the training cell
Already did that, thanks though.
Is it possible to run this repo locally? I have a i7-11800H 3060 6gb laptop version.
i'm afraid not, the minimum vram requirement is like 12 GB or somethings
Awww I am a potato then
Question, should I put booru tags for characters in initializer words? Like: Pony tail, Red Hair, Long Hair? Without underscores.
On the same note, I may also have fucked up the training by using 10 vector tokens in 84 images dataset.
I used 140 images dataset for 16 vector. Is that a bad thing?
It depends on what your embedding outputs.
Trying to create an embed for an artist style, using approx 1k images that were filtered by myself. Was curious if this is overkill and will ruin the training? Also do I need to flip the images as that is not something I have done?
It's overkill for TI
You'd better be making dreambooth with those images
I trained TI using 140 images vector 16 & got 13000 steps. All the 140 images is face only. The problem is, I cannot generate full body. Only can when img2img
whats dreambooth lol
Im using the webui to make the style
currently at like 28k steps
There's a dreambooth plugin for webui
huh
got a link for it?
It should be in the available extensions
I dont see it?
https://github.com/d8ahazard/sd_dreambooth_extension
https://rentry.org/textard
TI is not really for artstyles
Pardon me, wrong link
hmm
The original author of TI didn't even consider using multiple tokens
so does dreambooth do it better or something?
dreambooth is making your own model
Im making an embed
tho?
TI will get worse with multiple images
https://rentry.org/simplified-embed-training
Oh that guide is good
says who
you could use around 300 and more
<:cirnoWide:930962869401571348>
multiple as in more than 1k
my arknights is 1.4k datasets
doesnt really matter as long as u know what ur doing
just fuck around with ur settings
But TI tries to interpolate every image from the dataset
arknightsstyle?
original author didnt consider using multiple tokens cause it was the first textual inversion implementation
the script for it improved
And then there's also multiple token count that changes a lot
or is it a different one?
arknightsstyle yes
though, if you have that kind of amount, you better just use dreambooth or arknights
i mean finetune
fuck lol
hahaha
always use arknights
hmmm
goldenglow TI?
you'd want that though, pr sure no one likes rigid or oddly coherent generations. flexibility is also an important factor
cafe's ti repo is much better than the original TI trainer in webui anyway
Well I'm the one who fixed WebUI TI trainer
Haven't tried cafe's code yet
Whats a good token range? and vector range?
Is WD1.4 okay to use with this colab/repo?
May have issues with vram, and the yaml needs to be set to a 768/768 resolution π€ but should work
Which file to set resolution
The yaml
There's a link to the file in the colab's instructions
line 57 to be precise
ok
Would free colab be enough vram ?
Yes l, probably
i've trained with 768x768 with the free colab, so probably yes
What how
On a T4?
ye
When I tried it definitely wasn't enough
I trained this with 768x768 on a T4, it just took like 3 hours to do so
https://discord.com/channels/930499730843250783/1057106334698446859/1057271806400135168
Have you modify anything else in the config?
nop
maybe that's because?
Nah, me neither when I last tried
It just went out of memory instantly
Interesting
That colab donβt have some memory optimization
Mm which
And how should I optimize it
Are there a better ti trainer repo out there?
No
i don't understand then
11k steps takes 3 hours but on 512x512 (T4)
That is normal, I trained dreambooth before and it still takes about 1s/it just like TI
i mean, 4093 just said that he trained it on 768x768 on a T4 for roughly 3 hours to get 11k steps
i can't even get the T4 start training on 768x768 because CUDA runs out of memory
Maybe he used T4 that has larger VRAM size
<:thonk:876908278125330463>
You can train TI faster if the card has more VRAM
I guess he used a T4 with 24gb vram
but T4 has only 16 GB
there's no other variants
Colab Pro has A100s
with 48GB if I remember correctly
it's possible to train on that but that's definitely not the free version
also, it should be much faster than 3 hours
I'm guessing that he was training on 512x512 unconsciously
Oh
I guess he used 768x768 images to train
But he didn't set the resolution in the yaml file to 768x768
yeah i checked the dataset, it has 768x768 images
So the trainer resized it automaticlly
can be, i've never tested it in that way
<:Sadge:945662781938606151> I was getting hopeful
You can upscale 1x 512x512 images
The images will be denoised
I think even if 768x768 images can be used to train TI, the speed will be very slow
I feel like for wd1.4 you wont need 768x768 images to create TIs
I've checked the first epoch and it can create very good images on 512x512
so free Colab training won't die
No, even if you use wd1.4 to create TIs, the vram usage will still be the same as when you use other models
i know, but current TIs not work on wd1.4 and we have to later retrain all of them
also the cafe repo will need an update to support sd2.0 models, i'm sure
Cafe will add some optimization to the colab I think
he's very busy
I can't count on that
also he just made the repo, not the colab
I forgot π
if the Colab could be somehow optimized more, I think I've already tried to do it
Is current WD1.4 still based on 512x512?
That's what I'm trying to confirm but I found nothing
but I remember that I saw that somewhere
if that's the case, free Colab TI training would be still viable
I think we can use NAI-based models to generate images using the current TIs, then use img2img with wd1.4 to get better result
welll yeah
Or vice versa?
mmm
in theory it could work
or just inpaint the hands and bad parts if wd1.4 becomes that good
for now it still definitely needs some cooking but seems promising
The collab doesn't have xformers
cafe's repo doesn't support it yet
even if you install it, it wont do anything
I see. . .
nope! as far as im aware there are only 16Gbs T4s available in the free tier of colab
yeah but with andite's colab it is not possible to train 768x768 images with T4
i see
also the 3 hours for 11k steps checks out perfectly with 512x512 images on T4
so it indeed got resized to 512
it must have been
tbf at least we now that it can resize images by itself
but probably it's not ideal
Also I think it would cost 9 hours to train on 768x768 images with Colab T4 if it is possible
well.. in my experience it wont make too much difference anyways
(training on 768 instead of 512)
hmm I'm trying TI with diffusers, results are not good...
Not sure if the problem comes from diffusers or the model
What exactly to put in the "initializer_words"? I want to train a style. The images for training i have is only the face part.
Do I need to delete the default words at the "initializer_words" like the Illustration & Clothes?
face, shading, lighting, sketch(if it's like a drawing), illustraion, etc.
My image dataset don't have clothes in it, only face. Thus, does that mean I have to retrain my embedding?
well, it won't make it worse in any ways if you are using unnecessary initializer words
But it helps if you use words that are relevant
Thank you, i will retrain the embedding with initializer words that is relevant to the dataset π
How to use batch with A100?
Last time I used A100, I only used less than 20% vram
I think by changing the config `.yaml` file at `data.target.params.batch_size` (line 82)
Do I still need to run the cell if I am using local dataset ?
Why ?????
no
u can ignore that
that cell is for ppl with slow upload speed like me
guess I have to go back to the old repo
wait , the old repo gives the same error
should I use wd-1-4-anime_e1.ckpt or wd-1-3-5_80000-fp32.ckpt
in training ?
No, old repo and new one didn't support sd 2.x models yet
So I can't train ti with wd1.4 for now ?
No. You can do it through the webui but it is really slow
then I'll just wait
maybe it's not a good idea to train embeds on the first epoch of wd1.4
I just purchased 200 compute units.
Do we have a repo for dreambooth ?
Try this one
https://colab.research.google.com/github/Linaqruf/kohya-trainer/blob/main/kohya-dreambooth.ipynb
You can finetune wd 1.4 using it iirc
does it work with SD2.1 based model ?
Yes
Oh also you will need to add --v2 to the training cell iirc
Just realized I wanted to train 768 model
768 in most cases can contain character's face, I don't need to downscale anymore
Nahida TI revamped
Nahida from Genshin Impact retrained on the Cafe TI repo Colab.
looking forward to a Klee embed with similar quality π
Actually...
maybe tomorrow
holy cow that's way better than mine, what kind of training settings/dataset size did you use?
~120 images, ~9K steps. I've cropped them myself instead of using the autocropper thingy
Roxy Migurdia TI
It has been trained for 4700 steps with the collab on 60 fanarts of Roxy.
Oh wow!
Any chance of Eris or Sylphy?
I want to do Eris, Rudeus, Sylphy, Rujierd, I wanna do everyone xD
but I also want to dodge any potential spoilers, so I'm afraid to lookup fanarts that may portray moments beyond the anime
I already spoiled myself a bit, so I'm scared to do this any further
it's relatively easy to make a TI with the collab andite made once you get the hang of it, I encourage anyone who's interested to try it out
Yeah Fan art all over the place ATM. They just released the Illustrations for the light novel Volume 26. Lots of people are recoloring them and posting them right now. Best to avoid it if you have not read it. Although officially, you can only read to Volume 19 ATM. The anime goes up to Volume 6, and next season will only go up to Volume 12.
please consider making an Eris!
I'll think about it lol
I'm really determined to dodge any potential spoilers
Klee TI
Trained on Colab using 120 images, 11k steps (It's over 9 thousand!)
You have to use pretty strong negative prompts to be able to remove that hat.
next? π
candace
dehya
yeaaaaa thanks π
I'm pretty much out of Genshin character requests
Raiden Shogun is can be generated without TIs so that's pretty much everyone in my my main party
dunno if you're up for embedding recent animu characters
If it's not bocchi, then maybe
https://tenor.com/view/reincarnated-as-a-sword-tensei-shitara-ken-deshita-fran-calm-ten-ken-gif-26921712
how does Fran sound
and OH MY GOD AI generated Klee looks extremely legit
bless you π
just checked, there aren't too many good images of her
true, pretty new animu so the only source is the episodes themselves which is a pain
srsly thanks for the Klee and Nahida TIs tho π
blessed TI π
Alice real
How do I start training on colab?
click on this link, then follow the instructions on the google collab:
https://discord.com/channels/930499730843250783/1044623252326195311/1044623252326195311
if you're having trouble with it, let us know in <#1019446913268973689>! More people will see and may be able to help
Klee alternate outfit chibi version
how do I put Klee.pt into my google colab code? Like, which section and line should I put this .pt file?
you should upload it in the sdwebui's embeddings folder. idk if you are using your Google Drive or the Colab as storage
ohh. gotcha
Kyaru TI
Layla TI
A TI trained on roughly 150 images for 8000 steps featuring Layla from Genshin Impact!
Ryosios TI
Fischl TI
I've created a TI for the Genshin Impact character 'Fischl'. It was trained with 58 pictures in 9200 steps.
The eyepatch still isn't perfect, same for fingers.
I hope that you enjoy it β€οΈ
download the pt file. Throw it into your 'stable-diffusion-webui\embeddings' folder. Launch your web UI and put in the name of the .pt file. In this case 'fischl-9200'. You can change the name of the file to your liking. (You shouldn't use words that could be already registered in the database, or so I have heard. So add a number or use a unique name)
Cool!
Love to see more character based embeds.
Please leave feedback and share results. π₯°
Pretty good, but I miss the green eye color. Did you use the Cafe TI Colab?
I did, yes
try to specify the eye color through the prompt. That should work
Elysium_anime2 Model :3
Ralmia TI
TI for Ralmia - Sonic Racer from Shadowverse. Trained on 50 images or so with Cafe's repo colab for 7200 steps. Requires about 1.6 weights to change outfits and has a heavy bias for including skies.
Would help if I actually posted the embedding file.
KOS-MOS Version 1 - 12 Tokens, 50 images, 5000 steps
It's not terrific, but it has her likeness down pretty well.
not all blossoms spark
Lamia Ti
Quasarcake Style TI
This is a TI embedding trained on Qusarcake style (an artist that draw too much Hololive)
Trained on 191 upscaled and then downscaled images to remove jpg artifacts <:Uhh:1037138364056481793>
Trained for 100k steps on 20 vector per token on my laptop RTX 3070 (yes wtf is wrong with me... <:Agony:996544918610784296> )
Don't forget to check out some Qusarcake art at https://twitter.com/QUASARCAKE
Add some HoloEN members (like Kronii or Amelia Watson) to the negative, if you don't want it to generate Kronii or Amelia all the time.
Here are some versions with less steps π₯΄
on your laptop for 100k steps. Impressive my dear
<:pAwoo:266366834587074571>
Kuuko W TI
Its based on a Cosplayer trained with 299 Images using the collab
Its hard to get SFW tho so add negatives for NSFW stuff if u want SFW results, im using it with the Elysium Anime V2 Model
Prompts i used: (kuukow-445:1.1)
fully clothes, full body, (sitting on a bench), hair ornament
masterpiece, best quality, 1girl, ((kuuko)) in (winter clothes) during a (snowstorm) in the (Bamboo Forest), ((Adapted Costume)), Solo
Negative prompt: (((mutated hands and fingers))), ((poorly drawn hands)), ((poorly drawn face)), (((mutation))), (((deformed face))), ((ugly)), ((bad anatomy)), (((bad proportions))), (((extra limbs))), extra face, ((double head)), ((extra head)), (big breast), (((extra feet))), monster, (text), (logo), [blurry], (penis), lowres,bad anatomy,bad hands, text, error, missing fingers,extra digit, fewer digits, cropped, worstquality, low quality, normal quality,jpegartifacts,signature, watermark, username,blurry,bad feet, cencored deformed, blurry, (((mutated hands and fingers))), ((poorly drawn hands)), ((poorly drawn face)), (((mutation))), (((deformed face))), ((ugly)), ((bad anatomy)), (((bad proportions))), (((extra limbs))), extra face, ((double head)), ((extra head)), (big breast), (((extra feet))), monster, (text), (logo), [blurry], (penis), lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, lowres, bad anatomy, (bad hands), (text), error, (missing fingers), (extra digit), fewer digits, cropped, worst quality, low quality, normal quality, (username), blurry, (multiple girls), (2girls), (3girls), long legs, (long legs) (bra)
Steps: 40, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 3281597649, Size: 1024x1024, Model hash: 6b508e59, Denoising strength: 0.7, Clip skip: 2, First pass size: 0x0
very nice
trained with nai-wd?
ye
Sitora TI
It has been trained for 12.5K steps on 60 images from the artist.
This was meant to be an artstyle TI, but the composition doesn't vary much in the dataset, and the characters look a lot alike. This is why this TI can feel like it's ~70% a character TI and 30% an artstyle TI.
I suggest weighting it down to something like `(sitora:0.7)`, because it's pretty strong. You can also separate it from the rest of your prompt with the `AND` keyword if you want, this can help to merge it with any other concept you're trying to do.
Ms. Kobayashi TI
Trained for ~16k steps on 35 images, due to the community being degenerates there wasn't much good material.
While it does generate Ms. Kobayashi, it's not that consistent (mainly her eyes, and getting merged with Tooru). Still better than what the base AI can manage.
Credits: Help and assistance from: <>(Dataset and Advice) <>(Technical Help and Advice) <> (Advice)
β€οΈ
Congrats on getting this done
Some tips for use: scrape the pngs I sent for their data, the way I prompted may be useful for crafting what you need.
Especially the negatives that come after: "amputee"
For some reason high resolution really helped. I was generating on 1024x1536 with highres fix on.
And of course: Kobayashi is one of two good charachters in that show, along with fafnir. Yeah I said it, fight me!
Shitty meme to go along with my TI
> due to the community being degenerates there wasn't much good material.
?
bro...
I'm working on a v2 of this embedding.
it did not go well π
how
crow what happened
It has a better understanding of the eyes, but it drops the quality of the images
idk, let me keep testing
It's pretty much getting the eyebrows right every time now
but uh, still too big of eyes
that's probably just anything 3.0 though.
so yeah, unless someone makes a real sanpaku embed/hypernetwork I don't think it'll ever look right.
There are times when the shows takes itself more seriously and that is pretty much what kobayashi looks like. So atleast it can do that.
Attached is the new version, with proper init words, don't really know if it's that much better. Can't tell.
so after some more testing "blank stare" can get a little closer, still not ideal since it ruins any other expression
I'll mess with it more when I wake up
Scaramouche TI
This TI has been trained for **11k steps on 360 images**, this was supposed to be posted like yesterday but I was lazy and made it train for more steps. <:AYAYAYA:978879333248667718> . It generates Scaramouche in his default outfit (in-game one), So you **will have to use really strong negatives** to give him other clothes. Negatives to use : ((((hat)))), ((((necklace)))), ((((pendant)))), ((((symbol)))), ((((umbrella)))).
Also Boys aren't in the dataset of models as much as girls are, **you will need to use negatives to avoid it generating feminine versions**, avoid using ((masterpiece)) prompt since it biases towards girls. Enjoy! <:AYAYAYA:978879333248667718>
not much of a scaramouche fan myself, but I am glad we're seeing more Genshin characters
scaramid
Curss Style TI
Trained on Cafe's TI repo colab by me. Slime girls.
*trying to upload more images but it's a hard fight with clyde*
I like it I'll try that wfter work, thanks for your work! π
Finally good TI
π€
extremely based
I love the TI, thanks!
How can I start training on colab? I never done waifu or stable diffusion before. My computer ain't strong enough for those programs, but it can handle the demo version and the free Google version on the web!
use a google collab and participate with us in <#1019446913268973689>
π we can help
dont be afraid to ask questions
aww thank you!!
lets hope we can help you π
<> How can I participate in the google collab? I went through the whole process, but can't seem to find a link for the group. Is it a link? I don't know.
Delutaya TI
It has been trained for 11K steps on 60+ fanarts and some screenshots from her streams.
It can generate red hair if used as is, so I recommend adding `green hair` in the positive prompt, and something like `red hair` in the negative.
The twin tails don't always generate properly because of the training data, you can add something like `spiral twin drill hair` to make it more consistent.
LETS GOOOOOOOOOOOOOO
https://tenor.com/bYrt6.gif
<3
cute!
*artstyle* S4l (panamaman) TI
TI made on panamaman's art
(trained on 49 pictures, 8 vectors)
Model used: https://huggingface.co/Rasgeath/Kani
12500 steps and
16500 steps
Hitori Gotoh or Bocchi (not the brown hair one) TI
trained on 80 images, 9824 steps
Probably should redo with images that show less guitar and backpack straps cause its kinda frequent with this ti but it does generate her recognizably.....albeit it gets the dices weird in some of them
YES
thanks
i fucking love you π
<> THEY BEAT US
ily
omg
andite TI???!?!?
fool
don't let this make u giv up
do better
π
ok
they made a TI of andite naowaowe
do one with less straps so i dont have to do it myself
truly a cinematic experience when bocchi said "its bocchi the rocking time" and bocchied all of them
tfw andite TI
yooooo thanks a lot
FINALLY A BOCCHI TI
`masterpiece, best quality, Bocchi, , High Quality, yoji shinkawa style`
I'm getting weird artifacting, wut do?
Shamiko > Bocchi
<:sunglas:938888180311162950>
cannot assert nor deny
based
egh
its hard with a TI
also artifacting
getting there
Kyouka TI
Arona TI
Senko TI
Rotte (Touhou Lost Word) Style TI
Trained with Cafe's Repo at 6700 steps
Copies the overall feel of the Touhou Lost Word gacha game art style, minus the fancy rainbow hair highlights that I really, really wanted π’
Even copies their poses
You'll need to put high emphasis if you want to generate someone specific in this style or else it generates combinations of Touhou characters
prompt used in examples:
`masterpiece, best quality, (((Alice Margatroid))), rotte, blue eyes, (red hairband), blonde hair, watercolor`
Okayu TI
Nekomata Okayu, hololive. Trained on Cafe TI Colab.
LETS GOOOOOOOOOOOOOOO
OKAYU
do kson
now
aaaaaaaaaaaaaaaaaaaaaaaaa
Shirakami Fubuki | Nagishiro Mito TI
Shirakami Fubuki/Nagishiro Mito's style.
V1.1: scroll down to download, trained with cafe's TI repo colab. 3800-5600 steps. https://discord.com/channels/930499730843250783/1048266381264683098/1048799140466933790
V1: Trained on Waifu Diffusion 1.3, mixed with Anything V3, 15k steps, generated with Anything V3.
First upload. I have no idea how AI works so if you have questions I don't know if I can really answer them.
```Prompts for image 1: Nagishiro Mito 15k, masterpiece, absurdres, (fake cover:1.4)
1girl, fox girl, white hair,(fox ears:1.3), (animal ear fluff:1.1), (Shirakami Fubuki:1.4), portrait, (full body:1.2), green eyes, intricate eyes, (looking at viewer:1.2), standing, (facing viewer:1.2), detatched sleeves, (ass visible through thighs:1.3), (small ass:1.3), standing, small breasts, (midriff:0.9), skinny, tall girl
low ponytail held with a black hair ribbon and sidelocks and a side braid, strapless hooded top with a front slit and a (blue neckerchief:1.2), white detached wide sleeves, (black shorts:1.2), short shorts, a single thighhigh, thigh strap and black boots,
(morning:1.1), (beach:1.1), (palm trees:1.1), (tree shadows:1.1), (ocean:1.1), clouds
Negative prompt: nsfw, gun, weapon, holding weapon, sword, extra hands, extra arms, extra limbs, extra fox ears, bad anatomy, fat, thicc, loli, panties, panty peek
Steps: 100, Sampler: Euler a, CFG scale: 7, Seed: 2964469581, Size: 640x960, Model hash: 2700c435```
Oh, how did you manage to train on Anything?
actually
yeah
For me it always became a nightmare
Like this process, but with embedding instead, 0.05 https://youtu.be/1mEggRgRgfg
I initially trained Fubuki/Mito on hypernetworks, then I decided to try embeds.
thank you so much for this!
Added extra 0 when it starts to look good
Mh I tried this but didn't worked on Anything for me
oh u didnt use the ti repo
or did u
I didn't hear about it until like 10 minutes ago lol
0.05 learning rate?
anything doesn't usually work, even in the webui
that's weird, u sure its anything lol
Maybe its a miracle
or 0.005 initially, whatever is the default
I also put 20 vectors per token, I think
Is it working, or is it generating incoherent garbage? Can I see?
anything doesn't support training though?
It's either Anything V3 or Waifu Diffusion, I lost track of things, sorry.
or mixed both up by accident while training
Ah, should be WD then
I couldn't generate coherently or high quality in WD, so I switched to Anything V3.
trained in WD, used the embed in ANY
got it
here i thought someone figured out how to train in any
sadge
Is that normally how it works?
Cause I am really new.
π
You can tho use the trained TIs with anything
and have no idea what I am talking about.
^
Mikeneko TI when
Sadge, I hoped someone solved the mistery
Not enough high quality artworks to train.
The lower quality ones have too much variation in style.
nuuu
maybe in a year or so, artists will have made more for us
sadge
Nice to see Hololive fans here.
https://tenor.com/view/vtuber-hololive-%E3%83%9B%E3%83%AD%E3%83%A9%E3%82%A4%E3%83%96-shirakami-fubuki-fubuki-gif-26172434
https://tenor.com/view/uhum-fubuki-gif-26031953
working on sakamata TI right now, can't make a single preview because of hands so far haha
I trained a Rushia one a few days ago and it was a failure, I'll perhaps train that again and see how it goes.
yeah, in my experience the collab gives better results somehow
and it's faster (when you have access to a gpu)
2080 ti
Try the colab, or running the repo locally, it's really good
I am right now
So you're telling me I can do 6x the work in 1x the time
<:Agony:996544918610784296>
It's free real estate (if you have enough vram)
yeah that's what intelligent people have said
it's apparently 6x faster
you can go even faster if you have more vram
with batch size
Any help on the yaml file editing?
what do you recommend?
Editing the 27. And 29. Line
Also if you want, you can change at the bottom the preview image generation frequency
Idk exactly what lines but close to each other, modify the 100 to 200 or bigger
Two entries
training now
I leave for work in 20 minutes, I guess I will train it more when I come back.
Okay
Where do I see the training output images?
this doesn't work
go to your drive, and follow that hierarchy
they are in the logs
Generated with Anything V3, with Nagishiro Mito V1.1, upscaled
Sakamata Chloe TI
Trained for 7.5K steps on fanarts from danbooru.
It loves the policeman cap because I might have added too many of these in the training data, but you can mostly get rid of it if you add `policeman cap` in the negatives.
I'd suggest also adding `leather straps` in the positive as well, it can help to generate her normal outfit.
wait so are we gonna do the whole agency xD
tbh, I'd be down to do it
if anyone wanna help, let me know so we don't duplicate a TI
Nice
Sure, LOL
I'm probably gonna do iroha or koyori next
shinystyle2
trained for 5k steps on webui, adds glitter/shiny stuff to the eyes and saturated colors. a little strong so i usually use one [
Looks beautiful!
kokomi kokomi kokomi TI
Trained by over 70k steps and 100+ HQ hand-picked kokomi artworks
u guys need to try it
this has to be the easiest to use character TI i've ever come across, thanks!
video: https://youtu.be/nMEn_2H_Dbk
<> do you have a less trained version?
Shirakami Fubuki | Nagishiro Mito V1.1 TI
Shirakami Fubuki / Nagishiro Mito's art style. Lots of white hair fox girls (besides Fubuki) in the training images, you'll need stronger prompts to change hair colour.
Trained with Cafe's TI Repo Colab, 3800-5600 steps, 110-ish images, improved from my initial release.
Images generated with Anything V3 and upscaled, no hypernetwork/VAE used in generation.
NagishiroMito-and-ShirakamiFubuki-embed-V1.1-.zip
Title: Isekai Empress
Made with the v1.1 embed (mixed steps), 11 hours of inpainting, 2100+ image re-generations later.
Pretty fun.
Unicorn TI
Keqing TI
Shenhe TI
My very first TI! 134 images, 9k steps.
My first attempt only had training images of her in her full outfit, so changing her clothes (or removing them) was very difficult. This TI is a lot better than that first attempt.
Let me know if you want a more/less trained version.
Many thanks to <> for helping me!
incredibly consistent TI, nice job π
Anything 3.0? That model responds really well to this TI!
yes, it was with anything
is there a slightly less trained version of this? like say 6k to 7.5k steps?
<> is this less trained enough? Or would you prefer even less
<> lesser like 6k or 7.5k
8k is still very hard to modify
here's a 6 and a 7.5 - could you let me know which is best ?
7.5 worked well
btw, I found a way to add comments to your prompt in webui without modifying the code lol
you just add your comment in `[this thing::-1]` and it's gonna ignore it
oh wait, I thought I was in the chat lol
why am I writing under someone's embedding lol
BOCCHI TI
made by BASED DEPARTMENT!!! [peanuts, and BESTIE <> ILY]
TRAINED IN 376 IMAGES, 400,000 STEPS!!!!:https://cdn.discordapp.com/attachments/1046410688521846834/1049172464153935962/nutbocchi40000.pt
5000 steps:https://cdn.discordapp.com/attachments/1046410688521846834/1049172462975328317/nutbocchi5000.pt
10,000 steps:https://cdn.discordapp.com/attachments/1046410688521846834/1049172463264743434/nutbocchi10000.pt
20,000 steps:https://cdn.discordapp.com/attachments/1046410688521846834/1049172463558328480/nutbocchi20000.pt
30,000 steps:https://cdn.discordapp.com/attachments/1046410688521846834/1049172463843549234/nutbocchi30000.pt
40,000 steps:https://cdn.discordapp.com/attachments/1046410688521846834/1049172464153935962/nutbocchi40000.pt
yeay another bocchi TI!
have you tried both at the same time lol
I have no idea
what that would do
Generating Bocchi using a Bocchi Dreambooth model and a Bocchi TI on the prompt with Bocchi Hypernetwork on
Even though I worked on this, once again...
Shamiko >> Bocchi
we could make a shamiko ti after our game of rivals if u want
why not.
maximum bocchi
40,000 steps
It does the guitar so put "holding, instrument, holding instrument, guitar," in the negative prompts
if you don't want a guitar
gonna take me a while to get the dataset first tho
seems a tad overbaked
Maybe 20-30k is the sweet spot
but what do I know
30k keeps making a guitar
I prefer 40k ngl
alr i give up here's 30k
then uh
maybe I was wrong
I think we had too many guitar bocchis
20k
10k
Maybe should just do same prompt and seed for a proper comparison?
Oh well
oh its the same prompt
Looks good
Oh well, guess more steps just does that
wait wtf i didnt see this
lol
<> here
oh
Shamiko TI
MADE BY [based department] <> trained on 153 images
https://cdn.discordapp.com/attachments/1046410688521846834/1049490165380415518/shamicrow30000.pt
Lower step count
https://cdn.discordapp.com/attachments/1046410688521846834/1049490165086830632/shamicrow20000.pt
https://cdn.discordapp.com/attachments/1046410688521846834/1049490164797427783/shamicrow10000.pt
https://cdn.discordapp.com/attachments/1046410688521846834/1049490165736951818/shamicrow5000.pt
was this trained in the webui?
with higher batch count? (higher than 1 I mean)
Nope
cafe ti
30K steps <:monkaS:447131036900392960>
yep
768x768
how long did it take to train?
2 days?
4 hours
i believe
A100?
yep
that's cool. very nice TI
Thank you gekka
Momo soon!
<:AunnCool_:1006253939190153296> based ti
I owe you my life thank you crow
Yoisaki Kanade TI
Yoisaki Kanade from Project SEKAI: Colorful Stage. Trained for about 3500 steps or so, with dataset of 60 images, comprised of both official in-game card images and handpicked fanart.
>For some reason sometimes has a tendency to turn her hair blonde.
Also, I didn't train with her signature outfit (please show me how to i don't know how to do this aaaaaaaaaaaaaaaaaaaa)
also kanade touching grass in the sun no way
oh and as stated above if someone can show me how to properly train an embedding so that it includes a character's outfit with it that'd be great i really wanna make it so that it includes her jacket too
put hair on initializer words next time
so it could avoid the tendency to change color. it usually works for me
put attire on initializer words too if you want to train her signature outfit. make sure you have it in the dataset as well
thank you so much! although the images are only supposed to 512x512, i wonder how you're supposed to fit her entire outfit into it? its gonna be like less detail because its zoomed out in the same resolution, so idk how thats gonna affect the end results?
^^ for context im using the colab in the pinned forum section
in my case, i split each image to 3 via webui
preprocess images, split threshold to 1
you don't have to worry about cropped heads, it's barely an issue as long as u have datasets that includes faces
tysm for doing a kanade TI
tried doing it as a lucky charm for the portray yourself banner on jp
didnt work
im now sad
Prinz Eugen TI
Prinz Eugen from Azur Lane, trained on ~90 image, ~6k steps with Cafe's TI Colab.
Now do kancolle prinz
Roboco-san (Jacket) TI
Trained with NAI and 118 images for 10k steps using Cafe's TI Colab. This is also my first TI
Note: You may often encounter cat ears in each generation because majority of the arts trained with has her with cat ears. Adding *animal ears* in negative prompt should avoid it.
Trying so hard to get examples of her approved because of her lewd design <:DrollHell:1037148937888485427>
yoooo I love youuuu
thanks for making roboco
Ayanami TI
Yami TI
Konjiki no Yami from To Love-Ru, trained on many images and many steps, but you won't guess where.
sus
π€¨
> trained on many images and many steps, but you won't guess where.
great work thought π
π
thank you
Did not work for me. From what model is made?
NAI
the TI works fine with any model
haha
bruh, use anything
for training?
<:MomijiWide:1028509125044359259>
yes
No ty
why?
i dont want to see those preview images ever again
Centaur TI
A centaur TI I trained for ~10000 steps on 120 curated images from Danbooru.
Initializer is **"centaur-120"**. In addition to *"centaur"*, I recommend *"human torso"*, *"monster girl"*, and *"four legs"* in your prompt, and *"horse head"* and *"halter"* in your negatives.
(Removed first version of TI from OP)
Download v2 here:
https://discord.com/channels/930499730843250783/1049694491738648636/1050022356363513927
Better examples from v2 here:
https://discord.com/channels/930499730843250783/1049694491738648636/1050311975202197524
https://discord.com/channels/930499730843250783/1049694491738648636/1058703772722925629
Why does a TI need a specific initializer? It's not like dreambooth...
Are you sure you're using the TI for these images? You have to put the TI's file name in the prompt for it to do anything.
Because I screwed up and didn't make it the same as the filename, because I thought you weren't supposed to. Sorry about that. Like I said this is my first TI and I'm learning this through trial and error. It will be fixed in version 2.
I could just make a new thread and rename the file to centaur-50.pt
Not sure if someone told you in another chat already, but TIs don't need to be called through the initializer, all you need is the filename in the prompt, regardless of what name it has.
I see
The hit rate seems to have improved when I put the filename instead of the initializer in the prompt
Or maybe not, hard to tell
It was probably assuming it was the embedding sometimes.. Not sure how it interprets tokens exactly
CENTAURS YESSSS
Let me know if you fix it up, this looks great so far!
learn to make x/y plot and you can answer these questions. these should be standard for all posts in here.
The dash is important, it should be centaur-50
my mistake. but it still isn't doing anything. its not picking up the embedding unless you use the filename, its just emphasizing the word "centaur"
you could also test this by removing the embedding from the folder and I think you would get the same result.
might want repost these picture with censer breasts or a mod will delete it lol
yea my bad I copied the prompt from their png. u can just ban me I dont really post here.
nah it all good Lol
this one is sus
π€¨
π
cannot get any good images
Here's centaur TI v2, I haven't made demo shots yet
Mind if I ask what model you are using? Didn't try v2 yet, but v1 was giving me weird results with AnyV3 and ElysiumV2/V3...
Merged Model, Anything3 70%, Elysium 30%
These images are actually something of a lie, I didn't know how to call TIs
Version 2 is much better, I just haven't gotten around to really taking demo shots
I'm testing it out but still not much luck... I actually managed to get some nice centaurs without the TI, so my guess is that I need to figure more negatives.. Here is a sample: ||https://imgur.com/a/sZ6vRVA||
Imgur link cause even censored Clyde isn't letting it through
Gonna give that Anything/Elysium mix a try later
Made this with v2
π
Prompt is
```masterpiece, best quality, 1girl, (centaur-120:.80),(Centaur:Mature Female:.55)```
π€
Negative:
```Nsfw, (((deformed))), [blurry], bad anatomy, disfigured, poorly drawn face, mutation, mutated, (extra limb), anatomical nonsense, (extra legs), (ugly), messy drawing, text, halter, (horse head:.50), (horse neck), (Human Legs), Bipedal,```
finnaly, thank you π
Nice!
What's your model?
it is a mix
anythingv3 pruned32float-gyokai 50% (1)
+
elysiumanimev2-gyokai 50% (2)
and then
(1)+(2)50%= this mix
βοΈ
I see
ye i lerned it from Dr.God(Dr.bob)
nice
not so great, I don't know what prompt you are using to get so clean images
`masterpiece, best quality, 1girl, (centaur-120:.80),(Centaur:Mature Female:.55)
Negative prompt: Nsfw, (((deformed))), [blurry], bad anatomy, disfigured, poorly drawn face, mutation, mutated, (extra limb), anatomical nonsense, (extra legs), (ugly), messy drawing, text, halter, (horse head:.50), (horse neck), (Human Legs), Bipedal,`
this is the creators prompt
ty
the models they used whery anythingv3 70% merged with Elysium 30%
Gives pretty similar results from what I tried..
||I probably forgot restore faces enabled and it screwed things up, sorry about that lol||
xd
just realized that my 3200 embed is named as `centaur50-3200`, so those 4 on the right are without it <:kek:821047103451037778>
sorry about that
yeah i was wonding
cherry picked from Elysium Anime V2
π€
These are very nice. I should learn to make grids sometime
it is pretty simple to use, the only "complex" features are prompt SR and ranges. Both are on the webui wiki tho
Sorry if these embeds are confusingly named, centaur-120 is the newer and bigger one (v2) . I might close this thread and start a new one for easing confusion
no need to close it. Just remove the old files, upload new ones and pin it.~~ I think you can pin them at least~~
Especially because I might have not even been using the embed right in the OP preview images
Are you using the same prompt I did? Just curious
```masterpiece, best quality, 1girl, (((Centaur))), Equine Taur, (Human Torso) on Equine Body, Hooves, Blonde Ponytail, Blue Eyes, Perfect Face, Arms, Large Breasts, (Intricate), Toned, Good Anatomy, Four Legs, Full Body, Solo, Monster Girl, {INSERT EMBED NAME HERE}
Negative prompt: Nsfw, (((deformed))), [blurry], bad anatomy, disfigured, poorly drawn face, mutation, mutated, (extra limb), (ugly), messy drawing, text, halter, horse head, legs, (human legs:1.3)```
last row is without embeds btw, but keep in mind those were picked from a 100x3 grid <:kek:821047103451037778>
I see
Does it have to be anything float 32? Does float 16 work?
uh idk, i used the 32pruned one, but if your pc or what ever you are using cannot handle 32, maybe try 16float
I don't think it makes a perceivable difference in quality
π€
lets hope
Is gyokai a hypernetwork?
Or a model?
a model
idk where it came from but
uh
https://huggingface.co/NoCrypt/gyokai/tree/main
you can get it hee
Thanks
π
I succeeded in replicating your "DrBob Mix", our model hashes match up exactly according to PNG info. Here's your image with a different VAE
Neigh
Img2img is so powerful
Was able to fix the perspective issues
Fixed hands, kinda
Alright, that's enough for now
Works with realistic styles, too
img2img is sorcery
π€ weird
~~ill pass you my model once i get it to work
:/~~
(edit)
OOPS i has miss read your words π
i like that it can still make flat chest, if prompted well, well mine is a boy, but π€·ββοΈ
i like it
<> did you train this embedding as a style, or as a character?
Artstyle
Did you add any additional initializer words in the yaml file?
Centaur, Hooves, Horse Legs, and Monster Girl, IIRC
Whoa
π
Trying out your prompt as well but still having issues. Tried to copy your set up from here but the result is still bad. I'm using AnythingV3.
Ok I am in the situation where in a batch the first one turned out well but the other 3 did not. Interesting
π€
Can I try and make a request for harpies? As it turned out I can't do textual inversion because I don't have enough VRAM.
i rember tamzly tying that
Person or model?
person in this discord
You sure? Because I can't find him.
no, they said they are tyying to make it
Ok in that case I just wait for it to come out and focuse on other things at the mean time
oh thats me, yeah im sorting through 14000 harpy images I downloaded from pixiv atm but I was considering waiting on WD 1.4 to use that for my training base instead of NAI+WD1.3
you can have the old harpy embed I made
Its not great atm, seems to only be able to do frontal shots
βοΈ
Thank you. I will do some texting in the upcoming days and give you the results.
Also. What pictures did you use? As in characters
there were a couple Papis but i tried not to focus on a specific character
most images were OCs from Gelbooru
also I only used pictures where the harpies didnt have arms
but i know some people draw harpies with arms + talons, i might make a separate TI for that specifically
the problem is that getting good quality harpie pictures to use for a TI has been difficult so far
less than 0.01% of the pictures ive sorted from pixiv are usable
at least from the first 1000 i looked at
That is good since the AI usualy can't comprehand that part
it still makes arms sometimes, on the other hand it does well with the bird feet
That's good. than I can use is
Tryed to make some Harpie Ladies but it just didn't worked
ahh harpie ladies are quite a bit different than what i trained this on. they have arms and talons and iirc weird colored skin
Indeed. So If I get results that are decent with your embeding I will give you the info on it
alright sounds good, do try to "break" it as well so i know what to focus on for v2 of the TI
i had one report that it doesnt do back shots well at all
Oky. If there is anything that won't work or you want something specific to be tested just let me know
Updated OP, removed version 1 for being obsolete/causing confusion
Some recent generations of mine using the v2 TI
(This one below made by <> )
i should have posted some myself
unorthodox centaurs
ok i see
Man these looks so good!
Why can't mine work this well!? <:Agony:996544918610784296>
try using clip skip 2, hires fix on 640x960 and study my prompt
Shinon's TI
[Textual Inversion List]:
**Shimada Humikane Test v2 (6200):**
-
Barkhorn-inversion [Modified Andite colab notebook with additional features]:
-
Going to try out a few training a few Textual Inversions. I didn't like flooding <#930499731451428926> with WIP images while it is training.
1600 Steps in.
2100. TI appears to really like cat ears. [Well the dataset is kinda foucssed on it.]
Text Steps:
2200: Lewd, so not showing. But the style is being captured decently.
2300: AI actually posted NSFW lol. not showing fr obvious reasons. I think the AI got the hair down right. still needs work on the eyes though.
2400. The hair is there I think. Not the rest though.
It's going fast ish. So I'll post probably every 1000 steps
3000
Cropped because the bottom is kinda... well...
42nd post >_>
can I ask what character you are training here?
or is it an art style?
oh is it your art?
It's an artstyle
And no it isn't my art. I just want to train it to see where it will lead to
the artist himself used AI previously for a couple of 1 off tries.
6200 steps in. I think it managed to get the closeups about what I expected it to be. but thet eyes still needs improvement.
The arms are just [No]
Faces are getting there I think.
4300+4400 steps
4300+6000
**TI Test release:**
Textual Inversion trained on Shimada Humikane artstyle. (10900 Steps)
**Usage warning:**
This model is likely to generate NSFW artwork. So your milage may vary.
Dataset is biased toward World Witches Series as the creator himself draws a bunch of those.
**Data Source:**
Fantia images. Which has censored nudity.
**Keyword:**
`humikane`
Sample look
14200 Steps
Just had a look at the TI's trained and IMO I think I need to redo the dataset.
dataset redone. Will be retraining once the colab thingy works again lol
Dataset changes:
- Reduced the total images to 240ish
- Picked images with 2000px > in both dimensions
- Cropped them at 512px
- quality filter them to have a more standard look
- Removed tan variants.
**Barkhorn Inversion Notes:**
In general I probably will only do additional bugfixing when it breaks on my side. So avoid any feature requests, etc.
**TI Test release v2:**
Textual Inversion trained on Shimada Humikane artstyle. (6200 Steps)
Seems to be decent.
Biases:
- Seems to be biased on hair color and eye colors. you might want to reduce it with `[]` and `()` respectively.
- Attaches cat ears/fox/whatever ears to characters. Could be an advantage.
- May be too blurry at times.
MOMO TI
BY <> AND <>
https://cdn.discordapp.com/attachments/1046410688521846834/1049850868700299364/cromo40000.pt
https://cdn.discordapp.com/attachments/1046410688521846834/1049850868700299364/cromo40000.pt
https://cdn.discordapp.com/attachments/1046410688521846834/1049850868377321495/cromo30000.pt
https://cdn.discordapp.com/attachments/1046410688521846834/1049850868066947112/cromo20000.pt
https://cdn.discordapp.com/attachments/1046410688521846834/1049850869258125432/cromo10000.pt
https://cdn.discordapp.com/attachments/1046410688521846834/1049850868956139610/cromo5000.pt
30k
I uhh
Am not a big fan of 30k
20k
40k is personally my favorite
maybe I should have put hair ornament in the init words
seems to be free styling with it.
maybe this one is the "ornamentless" version and you gotta put "hair ornament" in negatives
I CHANGED MY MIND
30K IS GOOD
hm
wrong image mb
try putting hair ornament in the negatives
Can it generate her without it?
alr
huh
this looks like someone forgot hi-res fix lol
lonk
Khyle Style TI
Trained on the entire portfolio of notorious meme artist Khyle.
https://twitter.com/khyleri
16 vectors, 6000 steps (Cafe trainer, batch size 2).
For best results, *remove* `masterpiece, high quality` and any other highres related tags from your prompt.
You may get film dust aka "dirty photo" look in your images. This is normal and part of the artist's style. Lowering TI strength to 0.85 or running the image through an upscaler may reduce its visibility.
Works great! <> also curious if the 4th image of what appears to be asuka in the baseball cap is img2img? If it's a 1 off txt2img i'd love to learn from the prompt if you are ok with sharing :)
It's all txt2img. Asuka prompt was `1girl, solo, :o, simple background, red eyes, ((souryuu asuka langley)),
so simple, but i love it. reminder that smileys and emojis can be used was abit of a wakeup though cheers :o
emojis do work. `:o` especially works because its a valid Danbooru tag. Some of the emojis work better than others however.
Style TI
Fran TI V0.5
I'm gonna post this now before I get distracted and forget.
Uhhh Fran TI trained on like 63 images from the anime up to 40000 Steps
you forgot the prompt recommendations
I'm workin' on it!
I'd reccomend using either the 20,000 or 30,000 step variations.
But it's your life do what you want.
Here's just some of the tags I used to try to fine tune the results. Mix and match them as needed (for example, for shots that aren't full body don't include stuff like boots, or for shots with the hands out of view don't describe the gloves Use common sense, and experiment!)
```
(dark blue animal ear fluff:1.0), cat girl, (cat ears:1.2), short hair, short hair with long locks, crossed bangs, blue eyes, detailed eyes, detailed face, anime coloring, (black thighhighs:1.2), (white skirt:1.1), (one white stitch:1.0), short sleeves, (fingerless gloves), (white armored dress:1.0), (breastplate:1.05), fantasy, leather ankle boots, (ankle boots:1.15), (brown footwear:1.0)
```
I also had these negatives for most of images, but I have no clue how much they were actually helping I did not test them as extensively as my positive tags
```
big breasts, large breasts, (sword:1.2), (holding weapon:1.2), (holding sword:1.2), (((cat tail))), multiple girls, multiple views, bow, maid, yellow, magic, button, long hair, ponytail, sweat, (blue cat ears:1.0), (dark blue cat ears:1.0), (buttons), long skirt, long dress, (white thighhighs:1.0), ((tail)), (((multiple tails))), bare legs, (more than one hair ornament:1.0), (more than one ring around cat ear:1.0), reflective, miniskirt, reflection
```
**EDIT: If you don't really care about the outfit, obviously don't include those tags. just the cat and hair stuff (maybe anime coloring) will help with the blue ear issue**
I will be also working on a V1 of this once the show is fully out, and we get a full fanart dump. As while the anime screencaps *do* produce results, they also make the images look pretty low res and anime-y.
Have fun and good luck!
GOOD JOB!
so cute
what model is this for?
using the tested positive prompt above, this is what you can expect on the TI files when using Anything v3
<> let's goooo
wait it's not recommended it just kinda worked for me
I don't want to get anyone's hopes up
the grid does show that it works better than just masterpiece, best quality though
I was using anything 3.0, never tried anything else.
I know it works with other embeds decently, but requires more prompt work.
but uhhh it should maybe work with other models
fixed some issues π
Glad she took off her breastplate while eating, it could restrict swallowing.
Getting back on topic, my actual prompt included things like solo and 1girl in positives
and multiple views, multiple girls, and 2girls in negatives
to prevent multi frans.
Maybe I should have included those, but I thought I kinda thought it just part of everyone's base toolkit
alright, changed recommended -> tested
Well, you didn't have to do that. It's your message after all. :)
but uhh thanks.
Here's a decent one from elysium. Massively increases the fidelity of the backgrounds, as usual.
It might also have slighter better quality? but that might be placebo.
(also rip ankle, you will be missed π )
<> My hero
all I did was take screenshots, Crow did all the training lol
<> <> My Heroes
When do i become a herp
hero
https://tenor.com/view/kratos-gif-26454802
You need to slay the demon king
this looks like tag from "anime screencap"
because it's basically that.
But you can use it to make non Anime screencap pictures
Dataset curators get no recognition. π
I should have put a credits
Fucking amazing work.
π
itβs fine iβm not in it for the recognition
Iβm in it to spread joy
Your embedding is amazing i can finally make accurate Fran pictures. I have to go to sleep now though i stayed up an hour longer than i initially wanted because of your generous donation to the community
Thank you all of you that participated and are still participating in this project
Good Night
yep π
I just found out that this TI is extremely high quality
just need to drop the cfg scale and we're good to go lmao <>
Outfit consistency
I might be a bit obsessive over the outfit consistency
not my problem, I attempted to make the TI 5 times so I can dress her up and put her in other settings lol
Then maybe I shouldn't have put outfit in the Init words
maybe a v1 thiing
nah i set my goal really low because we currently don't have any good way of generating sword shishou guy
depending on how good WD 1.4 is at drawing hands and swords, I might actually want to put her in accurate outfits
I think it is still possible to use other outfits though
maybe they will still be similar to her default outfit but with enough prompting and use of negative prompts ypu might be able to archieve it
Fran's black cat armor set actually made it to the anime π
yeah that's definitely needed for V1 of this TI
fran loves pancakes π
cat girl
bad-artist negative embedding
negative embedding (embedding for use in negative prompt)
BTFOs your 50 token negative prompt in only 2 tokens
bad-artist negative embedding
The prompt for the images above:
```
solo
Negative prompt: sketch by bad-artist
Steps: 15, Sampler: DPM++ 2M Karras, CFG scale: 4, Seed: 1476197242, Size: 512x640, Clip skip: 2
```
Positive prompt is literally **just 'solo'**.
**Only 2 vectors** per token. Trained for 15000 (1875x8) steps, on 8 images.
It is meant to replace copy-paste negative prompts that are over 50+ tokens long in just 2 tokens. In fact, it produces better quality generations.
I recommend to use it with 'by', so for example 'sketch __by bad-artist__', or 'painting __by bad-artist__' .
Trained on blossom-extract [add_difference(A=AnythingV3, B=F222, C=SD1.4, M=1.0)], so it should work the best on that, however it should still work well with similar models.
How was this trained?
I took the 'schizo negative prompt' (https://rentry.co/nuwvd), generated 8 images with it in the positives and trained on that (rest of the settings are as posted above).
Surely, more than 8 images should be better... right?
Also, can I see said 8 images?
they're nsfw
and i'm currently not on my pc, i'm ssh-d via my laptop
well, actually
i think i can get them
Which model did you use to generate those images?
I tried both full and pruned didn't get quite near those
None of the above. It's AnythingV3 based. Specifically blossom-extract : add_difference(A=AnythingV3, B=F222, C=SD1.4, M=1.0)
Ah okay thanks
although it should still work pretty well
strange that it's not
what VAE do you have loaded?
Just AnythingV3 VAE
it's the same as NAI VAE
so I renamed it
yeah I suppose it works less-well with Anything
i've just tested, give me a second
Okay thanks so much for the info
looks alright (pure anything)
(this was with CLIP skip off)
however combined with an actual prompt it still looks pretty good
<> pic above was
```
solo, (absurdres:1.2) landscape wallpaper, portrait, intricate high detail
Negative prompt: sketch by bad-artist
Steps: 9, Sampler: DPM++ 2M Karras, CFG scale: 4, Seed: 2520566801, Size: 448x576, Model hash: 7ab762a7, Model: anything, Batch size: 16, Batch pos: 0, Eta: 0.69
```
still pretty minimal prompt
Some are good some are bad I would say lol
it's quite random
much better on blossom-extract
what is your CFG?
4
pretty much this param
i'd say it did pretty well with just 'solo'
yeah
that should give you a good result
9 steps is all you need
with my usual prompt tho
I've gone from this
to this
I censored cleavage a lil bit lol
they look pretty similar
kinda good ngl
also try cranking down CFG
to like 4
see if it does something nice
(made with Anything v3)
*first one is long negative, second one is by bad-artist, (1920x1024 each)
I wanted to check how well this does against a 50 token negative prompt so I did tests using the same seeds with highres fix
turns out this TI changes images a lot than just improve things so what I did isn't the right way to test lol
I imagine we'll have to go with statistics - whether we like more images made with long ass negative prompts vs by bad-artist
with around 100 test samples
if it doesn't then just turn it back
the 1st one is definitely 'cleaner' per-se.
the 2nd one looks like it would be awesome if you gave it 2x the pixels to work with
my vram can't handle 3,840 x 2048 <:NotLikeKogasa:996413933449650218>
fucking 4k
here's my result
PROMPT:
```
((masterpiece)), ((best quality)), (an extremely delicate and beautiful), (beautiful detailed eyes), ((very detailed face)), ((lustrous skin)), ray tracing, finely detail,
(breasts:0.997), (1girl:0.992), (blush:0.970), (pink hair:1.5), (sailor collar), (long hair:0.958), (blue eyes:0.912), (huge breasts:0.897), (school uniform), (serafuku), (solo:0.873), (bow:0.825), (gradient:0.807), (gradient background:0.805), (nose blush:0.762), (hair ornament:0.753), (hair flower), (grey background:0.747), (sweat:0.728), (heart:0.636), (very long hair:0.634), (eyebrows visible through hair:0.626), (red bow:0.571), (hair between eyes:0.523),
(cowboy shot:1.2)
```
first one is my normal (ultra long negative prompt) with `(buttons:2.5), belt, (open clothes), (garter straps), (sleeveless), (detached collar)` at the end
second one negative prompt is same as the first one but add `sketch by bad-artist` at the top
third one is `sketch by bad-artist, (buttons:2.5), belt, (open clothes), (garter straps), (sleeveless), (detached collar)`
Sorry for my bad English, I can't speak properly lol
it looks like it's made the contrast too high
what was the CFG?
4
wow
can you try (bad artist:0.8)?
i've managed to (sorta) replicate it
I was using Euler (Whatever the correct way to say it is lol)
somehow it produce hands and fingers better than DPM++ 2M Karras
now this was with (bad-artist:0.8)
```
parameters
((masterpiece)), ((best quality)), (an extremely delicate and beautiful), (beautiful detailed eyes), ((very detailed face)), ((lustrous skin)), ray tracing, finely detail,
(breasts:0.997), (1girl:0.992), (blush:0.970), (pink hair:1.5), (sailor collar), (long hair:0.958), (blue eyes:0.912), (huge breasts:0.897), (school uniform), (serafuku), (solo:0.873), (bow:0.825), (gradient:0.807), (gradient background:0.805), (nose blush:0.762), (hair ornament:0.753), (hair flower), (grey background:0.747), (sweat:0.728), (heart:0.636), (very long hair:0.634), (eyebrows visible through hair:0.626), (red bow:0.571), (hair between eyes:0.523),
(cowboy shot:1.2)
Negative prompt: sketch by (bad-artist:0.8), (buttons:2.5), belt, (open clothes), (garter straps), (sleeveless), (detached collar)
Steps: 28, Sampler: Euler, CFG scale: 4, Seed: 1251940333, Size: 576x768, Model hash: 2700c435, Batch size: 2, Batch pos: 0, Clip skip: 2, ENSD: 31337
```
i suppose there's something in your normal negative prompt that makes the image softer
which adds a nice aesthetic
because removing 'bad-artist' also gives a too high contrast image
I never ever read that
too long for discord too lmao
anyway that's what blossom-extract gives
with bad-artist
nice
guess it really depends on the prompt
for example, i re-tried the replied prompt with clip skip on
^ that was with the long negative prompt
i fucked up the seeds, give me a sec
^ that was with bad-artist embedding, which i think looks much cleaner
still anything-v3
yo I think that's a brilliant idea, maybe I'll try doing a negative embed at some point
not sure why we haven't done this sooner
I might just yoink the idea as well.
same
<> https://media.discordapp.net/attachments/1041850339034415124/1051876303931965501/00304-1476197242-solo.png tried your blossom extract with a replication, and it is what i got
whats the prompt
The same as yours
the first one
solo
only
clip skip?
same seed
2
it changed the clip too
the seed in that post is garbage
i added it afterwards
Oh
but this is what i got anyway
Oh yeah i see, but do you think you can share your own blossom extract ?
nvm thanks clyde
what i got for that image
yeah i was wondering, maybe your blossom extract is really good than compared to us
do you have vae
anything vae yes
the original vae
can you do a quick
```
masterpiece, best quality, solo, 1girl
Steps: 19, Sampler: Euler, CFG scale: 4, Seed: 1, Size: 512x512, Model hash: 7ab762a7, Model: blossom-extract, Eta: 0.69, Clip skip: 2
```
sure
what i get btw
oh wait
the sketch thing popped off
wait a bit
yeah, this is a control test (without the embedding)
your model looks different to mine
maybe you mixed something wrong
you think you can share yours ?
it's 2g
i have 10mbps upload
crappy crap
are you (sure) you mixed it correctly
that is really really sad
what model you used, the original anything or the unpickled
for the merge
add_difference(A=AnythingV3, B=F222, C=SD1.4, M=1.0)
i converted them all from .ckpt, it will not matter
was it 1 or 0.1?
1.0
oh so 1 then
did you do 0.1
my friend did yeah lmao
gj
have you remerged it ?
it's on going
damn
low ram i take it?
thanks π
<> the model is so strong that even an A100 cant support high resolutions
do you have xformers
activated it just rn
have you tried again with xformers
yeah i'm going to test rn
but it's really good
the model really is good
yes
i think the xformers is broken on the colab <>
looks like it
do you pay for colab?
sometimes yeah
are you using https://github.com/TheLastBen/fast-stable-diffusion
i'm using nocrypt colab
try this one?
xformers is a must
I am using nocrypt colab, it's easier and also it has much more extensions and stuff
sure
i am also using it, and i get 0 errors π€
probably do not have a100
anyway, have you tried the embedding?
im about to
Yeah
is it okay?
seems fine
trying my ice queen prompt wait
how to do x/y script with prompts π€
prompt s/r
ah thanks
https://www.reddit.com/r/StableDiffusion/comments/zk8y50/badartist_negative_embedding/
shill my post to the algorithm
is it even visible
it is.
i have escaped karma hell.
gg I am happy for you π
Will this embedding work with other style embeddings?
(like "bad_prompt" embedding?)
sure, it should work fine
yes
What terminal is that!?!?! <:YukarE:964342205269565510>
the new windows one (windows terminal)
Only for windows 11?
no
https://learn.microsoft.com/en-us/windows/terminal/install
Wow thanks!
Hmm thanks for reply instead of me
Feels like people keep asking me that
Hope that helps
BASED
Hi, where can I learn more about this so called "blossom-extract"? Is it a model? If yes, how to obtain it?
(A=AnythingV3, B=F222, C=SD1.4, M=1.0)
you have to make it your self
Why SD 1.4, is it bad to use 1.5?
π€·ββοΈ
The way add-difference merge works, it works best if model B is a finetune of model C.
https://huggingface.co/NUROISEA/anything-mix/blob/main/blossom-extract.safetensors
You can increase maximal output in width and height or is it upscaled?
what exactly?
to reach 3840 x 2048 stable diffusion images on regular consumer cards, you definitely need upscaling
1920x1088 AI images only need 8gb vram + medvram
Yeah thought so
Oh interesting
What gpu do you have?
rtx 2060 super 8gb
I have 1080Ti and 11GB of vram
nice, wallpaper size should be no problem
Thing is I struggle with generating 1024x1536
oh wait gtx
Hell 896x1280 eats 10GB vram
yeah I think that's because gtx cards don't have support for half-precision (FP16 instead of FP32)
Aaa sadge
So Iβm screwed?
not really - I'm pretty sure 1080TIs can be resold for a lot of money
can probably buy a new card with it
Yeah Iβm hoping to buy a new one
I sold my gtx1070 to buy my current (used) rtx 2060 super lol
Does nvidia tesla have support for half precision?
every rtx card has it π
that's why people could get away with their really old rtx 2060 non-super 6gb cards
Nice
So i guess βno-half-vae is eating my VRAM
on 8gb vram + medvram
adding no-half-vae definitely prevents me from making 1920x1088 images
Mhm
I had to enable it since black outputs sometimes
same, but as long as I need to work on wallpaper size images I literally have to live with that problem lol
Ah Understandable
I wanna buy this gpu purely for genning NVIDIA Tesla K80 (24GB Gpu GDDR5)
But the question is does it have half precision like RTX does xF
oh checking for that is pretty ez
Yeah?
How can I check?
looks like the card is from 2014 which explains why it doesn't have FP16
it...doesn't have FP16 lol
Hmm
Is that bad?
we should probably move this discussion to DMs since we're not talking about the TI lol
Yeah we should
https://cdn.discordapp.com/attachments/1010577750077210726/1054277230466973726/00946-4066035123-masterpiece_best_quality_traditional_Japanese_painting_style_1.43_traditional_Chinese_painting_style_1.43_divine_heavenly.png
Kaputii Style TI
Trained on Cafe's colab at 8000 steps and 82 images
artist: https://www.pixiv.net/en/users/245287
Known for drawing cute Yuyuko regular and chibi pics, I really wanted to see Alice drawn in this artist's style so yeah
This style TI doesn't really capture the artist's chibi version of Yuyuko though
What it does: make girls pretty and cute at the same time
Recommended CFG: less than 8
Prompts in samples
Regular Yuyuko: `solo, kaputii`
Chibi Yuyuko: `solo, ((chibi)), kaputii`
Samples were made with Anything v3
Drawing other characters in this style requires a lot of emphasis:
`((masterpiece)),((best quality)), ((((solo)))), (((Alice Margatroid))), ((((blonde hair)))), (((yellow hair))), (((blue eyes))), kaputii`
negative: `((((saigyouji_yuyuko)))), ((((kimono)))), (((((pink hair))))), (((pink eyes))), (((two-tone_hair))), (((multicolored_hair ))), (((mobcap))), (((green dress)))`
Another example of drawing non-Yuyuko in this style:
`((masterpiece)),((best quality)), ((((solo)))), (((Cirno))), ((((blue hair)))), (((blue eyes))), kaputii`
negative: `((((saigyouji_yuyuko)))), ((((kimono)))), (((((pink hair))))), (((pink eyes))), (((two-tone_hair))), (((multicolored_hair ))), (((mobcap)))`
The TI does help in making chibi Yuyuko so that's something lol
wow it's really cute but i felt bad for the artist itself lmao
That definitely sounds like this TI was a success lol
I just *really* wanted to see Alice drawn in this artist's style π
Yuuji Style TI
Decided to train a TI because I've been meaning to train a TI for a while now, pretty happy with how it came out. Example generations are non-cherry picked 768x768 generations as they are only to show off the style. In the comparison image, top is without the TI, bottom is with. Attatched is a comparison with and without the TI. Full res images can be found here: https://imgur.com/a/JKQFubm
I trained against 45 images with Anything V3 for 5k steps, 4k steps is looked the best, so thats what is included.
Example prompt:
1girl :< air conditioner alley animal animal ears bag bandaid bandaid on knee bandaid on leg beer crate black cat black footwear black hair black jacket black leg warmers black shirt black skirt bottle broken broom broom cat cat ears chromatic aberration closed mouth crate door frilled shirt frilled skirt frills grey leg warmers hand in pocket highres holding holding bag industrial pipe jacket ladder light particles looking at viewer medium hair neck ribbon neoki ohae off shoulder original outdoors paper bag ribbon shirt shoes skirt solo standing striped leg warmers twitter username watermark yellow eyes yellow ribbon yuuji4k yuuji4k
neg:
lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, censored, 2girls, 2009, 2010, 2011, 2012, 2013, 2014, 2015
45 steps euler a
>trained on Anythingv3
<:HUH:1002459549615271936>
<:reimuWHAT:975538133468839966>
>trained on Anythingv3
tell me the secret master
I didn't do anything special lol
when I used Anything for training it always produced nightmare results
Preview images are completely broken but the embedding itself works
how, mine was always bad
Idk lol
i tried it on the webui training, also on colab, both were bad
I'll do another one tomorrow
I used cafe's repo on a rented gpu because I didn't feel like waiting for slow colab GPUs
mmm would using a different GPU other than colab's affect the result I wonder
being able to train on anything would be good
Traptrix Sera TI
TI Trained on 100 or so pictures of Traptrix Sera from Yugioh for 4800 steps using Cafe's TI repo. Mixture of SFW and NSFW with a few group pictures to make it easier to include other characters when using her TI. Solo and SFW tags are recommended as using the TI by itself can generate group shots.
I didn't realize the training had continued after I thought it stopped. Here is a higher step count embedding that might work better.
I-19 (kancolle) TI
TI by me.
The text on the swimsuit is almost always borked but that's what the current models are capable of.
She likes to play with her torpedo
Genshin style TI
Trained on Cafe's colab at 10,000 steps and 178 images
This handles both regular and chibi versions of waifus quite well
very recommended negative prompt: `(logo:3), (text:3)`
Just like using the ordinary genshin impact tag, this also combines clothes from every region so you can't get specific region clothes without specifying with prompts.
comparison between using this style TI vs simply adding Genshin Impact to the prompt
logo too strong , even negatives cant kick it out
You could probably just like, use photoshop and paint out the logos, even if it's shitty it being the right color should just blend it in the background
Would have to manually do it for each image, then remake the TI
So not very practical.
nah it's significantly easier to just put `(logo:3), (text:3)` in the negative prompts than to modify the dataset images lol
even works on low cfg scales
Nice, I love the Genshin Reimu
Subaru - Oozora Subaru (Hololive)
Paimon TI (Genshin Impact)
Barbara - Genshin Impact
Prompt `white pantyhose` for best output
this is so good
anything, or a blend?
anything
full version
Ganyu - Genshin Impact TI (Holy Cocogoat)
This is a very horny Ganyu Embedding enjoy my homies out there.
Pretrained Model Nai +WD
20k steps Batch Size 2 90 images for training etc...
You can pull/summon her with "Ganyu-I" for free i promise β€οΈ
π€¨
Huh
let me fix that real quick
working now?
<:NotLikeKogasa:996413933449650218>
realism version
Not bad, im curious
Should embeds be traines with 2d and 3d gals?
Kanade - Tachibana Kanade (Angel Beats!)
Oekaki negative TI
This TI was intended to reproduce the Oekaki artstyle, but it didn't work. But when used as a negative prompt it improves quality in a weird way. Only seems to work on Anything, might work on others by messing with the weights.
Trained on 90 images, for 9000 steps in total.
I guess this works because the pics i used were low resolution lineart, so it pushes the model away from those. Pretty interesting that it works like that.
When i say "Oekaki artstyle" i'm talking about this. I just call them like that for not knowing a better name.
I was messing with the Embedding Inspector extension, to see if i could make the embedding better. And i tried making a new one using similar tokens. It kinda works.
So, what i did was i picked the 6 first tokens from the first 3 vectors, then i made 3 embeddings with concat marked, then i combined them into a 1 vector embedding.
The result seems to be an different new embedding that does a similar job, but as a prompt instead of a negative.
I guess someone might make some good embeddings by mixing random crap.
Rime [Hypernetwork]
A Hypernetwork of Rime from Cevio AI. Trained on 10 pics. Enjoy!
astolfo
mordred
I really like the style options it gives
chainsawman TI
trained on 36 images from part 1 & 2 of the chainsaw man manga for 1.5k steps on the webui
example prompt `(best quality:1.4), (ultra-detailed:1.3), (illustration:1.3), (greyscale, monochrome:1.2), (1girl, solo, very long hair, messy hair, disheveled hair, hair between eyes, black eyes:1.0), (sitting, looking at viewer:1.2), outdoors, snow, trees, falling snow, windy, japanese clothes, white hair, art by csmstyle1`
How long did it take on the webui
I think an hour or so with batch size 2 and gradient accumulation 4 & a RTX 3060
i assume u have more than 8gb vram
yep, I have 12gb
I can't get cafe's repo to work so I'm just using the webui π«‘
fuck that turned out good
Haru [Hypernetwork]
Faruzan - Faruzan (Genshin Impact)
KurumiTI
Trained with 50 Images of Kurumi
Trained on nai-wd using collab
for some reason she always Holds her dress xD
If eyes are off prompting "heterochromia" Sometimes works sometimes doesnt
Prompts in png info
new TI with more Pictures feeded, 14 vectors
Ah that's actually 9400 steps not 7300 BTW xD
<:cirnoWide:930962869401571348>
Kafu [Hypernetwork]
A Hypernetwork made of Kafu from Cevio AI. Made from 12 images. Enjoy! Trained with AnythingV3.
π
what settings did u use? Im trying to learn myself
I mean training settings
https://i.imgur.com/e7U73sj.png
https://i.imgur.com/PF8tLbQ.png
you actually did 100k steps?
Oh heck no, I stopped it at 9000 steps.
Sekai [Hypernetwork]
A Hypernetwork made of Sekai from Cevio AI. Made from 12 images. Enjoy! Trained with AnythingV3.
[CLOSED]
Example Request:
Character: Chisato Nishikigi
Portfolio: https://danbooru.donmai.us/posts?tags=nishikigi_chisato
π€
im actually requesting that one, too lazy
u could prompt her without a TI
she's popular enough
True, maybe I'll ask for someone else π€
Milf bronya gambatte
https://danbooru.donmai.us/posts?tags=bronya_zaychik_%28silverwing%3A_n-ex%29&z=5
Requesting training of Noi TI here. (giant woman from Dorohedoro)
Should we pin this too?
Request:
Character: Special Week (Umamusume)
Portfolio: https://danbooru.donmai.us/posts?tags=special_week_%28umamusume%29+&z=5
Thanks π
Request:
Style: ASK/ Askzy
Portfolio: https://danbooru.donmai.us/posts?page=2&tags=ask_%28askzy%29
Request:
character: Vermeil from vermeil in gold anime, alternate name in jp: kinsou no vermeil
Portfolio: https://aniyuki.com/vermeil-in-gold-images/
https://danbooru.donmai.us/posts?tags=vermeil_%28kinsou_no_vermeil%29+&z=5
Aight Now gimme **Bao **the whale
Request: Eris Greyrat
There's actually a hypernetwork on Chisato in the rentry, wasn't sure if you knew. I know it's not as embed, but yeah.
π€
Request:
Character: kotatsu tamaki (Fire Force)
Character: Shirai Kuroko
Portfolio: https://danbooru.donmai.us/posts?tags=shirai_kuroko
Character name : **Mash Kyrielight (Fate/Grand Order)**
https://danbooru.donmai.us/posts?page=1&tags=mash_kyrielight+
https://id.pinterest.com/search/pins/?rs=ac&len=2&q=mash%20kyrielight&eq=mash%20ky&etslf=6811
https://wallpaperaccess.com/mash-kyrielight
https://www.zerochan.net/Mash+Kyrielight,Mobile+Wallpaper
https://www.pixiv.net/en/tags/%E3%83%9E%E3%82%B7%E3%83%A5%E3%83%BB%E3%82%AD%E3%83%AA%E3%82%A8%E3%83%A9%E3%82%A4%E3%83%88
https://yande.re/post?tags=mash_kyrielight
https://www.reddit.com/r/MashuKyrielight/
https://anime-pictures.net/pictures/view_posts/0?search_tag=mash+kyrielight&lang=en
https://wallpapercave.com/mash-kyrielight-wallpapers
i already have seen mash pics on aibooru
with her armor?
no just dangerous beast and a santa dress
Artist request: Marota (warning, lots of lewd stuff)
https://danbooru.donmai.us/posts?tags=marota+&z=5
oo this is a nice place
does anyone have a request for a TI?
do you?
no
im just bored and have no idea what to do
making ti's are a bit fun for me
well look up
just do a TI for of the ones we want a request of then
Character : **Ushiwakamaru (Fate/Grand Order)**
https://danbooru.donmai.us/posts?page=6&tags=ushiwakamaru_%28fate%29+
https://wallpaperaccess.com/ushiwakamaru
https://id.pinterest.com/pg2125/ushiwakamaru-fate-grand-order/
https://id.pinterest.com/search/pins/?q=ushiwakamaru&rs=typed
https://www.pixiv.net/en/tags/Ushiwakamaru
https://yande.re/post?tags=ushiwakamaru_%28fate%2Fgrand_order%29+
https://www.reddit.com/r/Ushiwakamaru/
https://anime-pictures.net/pictures/view_posts/0?search_tag=ushiwakamaru+%28fate%29&lang=en
https://wallpapercave.com/ushiwakamaru-wallpapers
mobile legeneds π
hehe
guess what monstrosity im making today
popstars?
that's a lot of dataset π
guess which one
`kama_(second_ascension)_(fate)` outfit. https://danbooru.donmai.us/posts?page=1&tags=kama_%28second_ascension%29_%28fate%29+
uh idk xd
its rose from blackpink
its
r o s e .
Would be cute to see Lo-fi girl but there isnβt too much art of her.
https://danbooru.donmai.us/posts?tags=lofi_girl
Thereβs some nsfw on the booru, so I guess I should add a warning for that
https://tenor.com/view/kanye-west-stare-staring-funny-gif-13590085
Arknight lovers lol
Character : **Ch'en (Arknights)**
https://danbooru.donmai.us/posts?page=134&tags=ch%27en_%28arknights%29
https://www.pixiv.net/en/tags/%E3%83%81%E3%82%A7%E3%83%B3(%E3%82%A2%E3%83%BC%E3%82%AF%E3%83%8A%E3%82%A4%E3%83%84)/artworks
https://id.pinterest.com/search/pins/?q=Ch%E2%80%99en%20arknights&rs=typed
https://www.zerochan.net/Ch%E2%80%99en
https://yande.re/post?tags=ch%27en_%28arknights%29+
https://www.tumblr.com/search/ch'en%20arknights
https://anime-pictures.net/pictures/view_posts/0?search_tag=ch%27en+%28arknights%29&lang=es
Reminder, if you're interested or bored enough to pick up and train the requests here if you want to, you could do it, since it's a public post. help da community winky wink
I don't mind training her if someone posts a dataset of 50 to 100 images, but I do not wish to build the dataset myself
I wouldn't mind doing it but what and where is the database where u save em
you gotta search the internet, then add that in a gdrive folder or something, and share it here
Ahh kk
you can print screen or download the image, downloading works better
This gonna be a pain, her images are scattered, alot of adult fanart of her
yes, feel our pain
<:Dumbass:1052329551486787594>
I think a couple of nsfw is fine, as long as the majority of the images are her wearing her outfit lol
Fair enough I should get it done Soon
yeay, I'll try to do it tomorrow if you do it by then, else I have the holidays, and I won't be home lol
I will just make a folder in my Google drive and when done send u the link
sure, sounds good
I can try to do it from my parent's house but it's more pain
I'll see how it goes
hanyuu dreambooth please
https://tenor.com/view/hanyuu-higurashi-no-naku-koro-ni-watching-blushing-anime-gif-16502165
Do they all gotta be the same resolution?
no
Also PNG or jpg, or it don't matter?
doesnt matter
doesnt matter
<:cirnoWide:930962869401571348>
unless you'll preprocess it urself
to 512x512
read the thread again
cant see where "dreambooth" is
<:AunnThink:993596417488728174>
hanyuu TI please
kokkoro
I can pre process them myself, you don't have to do it
unless you don't trust me and really want specific parts of the image xD
I'll just use the webui pre processing tab with my prefered settings if you send them raw
ohh kk
https://drive.google.com/drive/folders/1d3j8qo-Oo9R6iP5Y27_9f76dGOSMxhKf?usp=sharing
does this work
so, as it is right now, I need to request access to be able to use it
https://drive.google.com/drive/folders/1d3j8qo-Oo9R6iP5Y27_9f76dGOSMxhKf?usp=sharing#
try now
nice, that's gonna work!
gg
<:MomijiWide:1028509125044359259>
I'll try to finish it before next week xD
good luck
I mean next weekend
yeah
thx
has anyone started Hanyuu? If not I'll do it
yea u da first one
man π΄
Requesting kanye west TI
ill do her, i like her alot
gimme maybe 3 days, working in second kurumi TI rn
i'm requesting a beatrice Ti https://rezero.fandom.com/fr/wiki/BΓ©atrice
<:MomijiWide:1028509125044359259> take your time
<:02hug:806525393535369276>
this is in my to do list
on it
ah nice, i belibe astorx wanted it aswell
and ther he is
LOL
if u do it then <> dont need to do it
it is my request
ah
I was going to do eris today even before I saw this thread
gathered the images last night π
let's goooo also i love you pfp
<3
having a look now at beatrice. Not that many images found so far :/
π
Ill give it a go πͺ
It will probably force generate books/cats, and I doubt it will get the eyes (TIs never do with special eyes), but you never know
i love you
I can try and gather some if i remember later today, but I'm a little bussy now
I think collab is done training lol
<>
if you wanna check what my embedd do before doing your own
for Eris
I DM'd you
can someone do one of rina tennoji https://love-live.fandom.com/wiki/Rina_Tennoji
based
https://tenor.com/view/nico-lil-pogchamps-urination-nation-rina-tennoji-love-live-gif-21711225
ShyLily TI, providing dataset as for one reason or another I cant train locally, so hoping someone can get it to work https://drive.google.com/drive/folders/10qL7DHOqMRkURGF3D0NXekuBCBo4o0dq?usp=sharing
shylily :o
I can do it for you
thank you chief, for some reason I cant do TI on my own, hypernetworks fine but not embeddings
would anyone be willing to make one with this character, All in panties
i will get the pics might be around 30 or soo
uhhhh
dont blame me lol, all the art from this one artist is all in panties
you could of just crop the image
oh hey kether
hello
tru
I don't mind doing it, can you confirm the artist name is `sabamen`?
Btw I don't think you I'll be able to get a specific outfit out of it...
If you are willing to cherry pick the images of her using the same outfit (or similar ones at least) that would be fine tho, I just don't wanna go through 368 imgs trying to find it
i want the art style/ character
idk tbf, i see this one account with 144k followers on twitter post it
@e7CdUVTndAtBvP6 ?
ah yeah that is him, first pic is the one you sent lol
ye
do u want me to get pics, or u alr? i assume u know more about the artist than i do
don't worry, already downloading it
i actually feel bad asking for a t, the art is too good
at least in my opinion
ayyooo!
would like to request an aoi ogata art style TI
https://www.deviantart.com/aoiogataartist
irisviel von einzbern
suprised no ones done kanna
I think there is a HN for her, not entirely sure tho
whats that
hypernetwork.. But nvm, it isn't on sdgoldmine at least
If it exists idk where I saw it
There is one for Ilulu tho
i've already made both
<>
ohh nice
ilulu is definitely better than that hypernet
Most models have a good idea of illya already u can just prompt her especially in nai based models
but she's nsfw by just existing so I can't make a post
Just use the danboru tag from the series and emphesise that along wint illyasviel von einsbern
That is her mother tho
But you can probably get it by prompting illya with big boobs <:MomijiWide:1028509125044359259>
Illya, (70yo girl:1.6)
Yeah or add (adult:1.5) to it
I got some not-so-nsfw images of her with the TI <:thonk:705424275196084344> gotta cherry pick but definitely possible if you wanna make a post
didn't try the kanna one yet tho
i mean
those honkers are not safe
too lewd for me to post
<> what model did u use it aint wokring well for me
How many images would you like for there to be in the TI
have you tried both TIs? sushi's my work better
for your model
I'm using a mix of AnyV3 and Bondage8
ye i have
hmm not sure which model sushi used for the last 3 images, you can ask him if you want
I think his model works well with the TI as well
model used is a blend
but its anything-based
funny coz the ones i use are mixed with anything
Are 320 images enough for an ArtStyle TI?
more than enough
the more the better tho
<:troll_doughey:971633334964846642>
i havent been able to get raw anything to work for embeddings, howd u do it
however many you would like, dont know thaat much about TI still tampering around with it
not trained with anythingV3, it's trained with another model, if that's your question
the model used for training is the old mix andite was using
ahh
and I don't remember what it is exactly since I'm dumb and I renamed it to NAI-WD
for reference, this is the kind of renders I'm getting with anythingV3 alone
you can try to add meat around it. The TI likes cat ears, so maybe add that to negative. Also, for my TI, it generates her with blush and open mouth a lot, so maybe add that to the negative as well
sorta works
it hates the realism promt
looking good to me!
stepbro im stuck
Waifu thru a microscope
has anyone tried a horror embed
with anyv3
<:troll_titan:971634013917818940>
wdym horror
I could try some junji ito shit
Lmk when its done or dm me it if it dont get posted in the Embeds
I was just testing the results lol
not perfect but it is looking interesting so far
<:MomijiWide:1028509125044359259>
Feel free to dm, since I know u prob can't post it here
pretty sure that is fine
That's interesting
considering what is often posted on <#1019446913268973689>
I wonder how it would look trained on yukiko or anything
ok, that is pretty much what he draws but lacking his artstyle <:kek:821047103451037778>
just gotta add some prompts like `flat chest, (hoodie:1.3), plain panties, (standing:1.1), short hair`
ohh there
i like the simple style for this
i'll let it train some more before making a post, maybe it will get better, but here is the latest embed if you want
thx
may have added some realism but, its getting baked with realims. gonna try with normal promts
looks pretty good, what model is that? <:thonk:705424275196084344>
Got NSFW with like 5 promts
not intentional but shes naked
<:NotLikeKogasa:996413933449650218>
good
<:YamaKKHTA:966212243022835712>
`meltryllis_(swimsuit_lancer)_(first_ascension)_(fate)` https://danbooru.donmai.us/posts?page=1&tags=meltryllis_%28swimsuit_lancer%29_%28first_ascension%29_%28fate%29
Would anyone be interested in making a TI on Takehito Harada's art style? He's the artist for the disgaea series. I already have a ton of images and can zip and upload them to save you time if you're interested. Let me know!
anything/nai already does that with the `(disgaea:1.2)` tag
Oh neat! Thanks! I don't know how I didn't think to try that honestly
cheat sheet. https://zele.st/NovelAI/?Danbooru%20-%20Copyright
Yo ill make upload now, cant get uniform tho because i used too many lewd images
Character: Ms. Fortune (Skullgirls)
https://danbooru.donmai.us/posts?tags=ms._fortune_%28skullgirls%29+&z=5
This one can be interesting since she can detach her limbs
imagine if I crawled all your public discord/reddit/facebook posts and made a bot that talks exactly like you, and attached your username to the project. wouldnt that be creepy? you shouldn't train models or embeddings on a specific artist and post it online without their permision.
https://tenor.com/view/guy-arguing-guy-talking-to-wall-talking-brick-wall-gif-18667615
https://tenor.com/view/doja-cat-star-wars-gif-25078126
Wdym creepy if someone made a bot based on my likeness ill gladly support that dude
is this a request?I think its possible using gpt
id feel like a data worth using for if its possible
id love it
<:AunnThink:993596417488728174>
id even use it myself
character.ai moment
yeah its def possible if you wanted to do this for some reason. I think people use it for twitter bots.
theres a thing where they make bot versions of subreddits as well. idk about scraping discord or how you would do that.
Pls replace me
It would be extremely flattering if someone were to do that
Request for an Artist by the name of Atelier GONS
Idk if it makes a huge difference but you can download several albums or whatever of their works
https://hitomi.la/group/atelier%20gons-all.html
That is, if im understanding the pain of building a database/scraping correctly
A mele weapon ti
Wait thats actually a good idea but for the main SD
Isnt this just a porn account
When i opened the link i found just CG Hentai
One of my friend agreed to me doing this just so he can have a bit of himself
I mean was it not made by someone?
Tbh I might do that yo myself. Talking to people is tiring
Wait is there no megumin ti
or am i blind
most models knows her pretty well
maybe but
suprising
pretty https://danbooru.donmai.us/posts?tags=rioka_%28southern_blue_sky%29&z=2
there is. Just not posted here
There actually is a bot like that online right now, also nice bait
I kinda need one, can u hook me up
if it wasn't mentioned yet, then project sekai art style ti https://projectsekai.fandom.com/wiki/Project_SEKAI_COLORFUL_STAGE!/Card_List
Would someone make Makise Kurisu TI?
<:gigachad:962417952290922586>
i'm already on it, but i dont guarantee if it will be done today
day 2 of asking for someone to do Rina tennoji https://danbooru.donmai.us/posts?tags=tennouji_rina+rating%3Ageneral+&z=5
gl soldier
if i were to do it myself would i get better result using embeds or hypernetwork?
I've been trying to figure this out doing both and I still can't figure out which is "better." both work perfectly fine. The Cafe embedding Collab is easier and quicker tho. I train hypernetworks on Paperspace and it always slows down.
not mine so not making a thread
thx
rip my request
Just going to throw my request into the ring, but maybe a Nikke TI? Sources can be found at: https://danbooru.donmai.us/posts?tags=goddess_of_victory%3A_nikke While characters do vary, maybe focusing on Rapi or Anis would work.
Donβt worry take your time , youβre a legend!
you can make very accurate pics of Mash already because she has a lot of fanart. but I might make an embedding for her in a couple of days if nobody else does it. I got some other things to train first.
I'd like to request Suletta Mercury from Gundam The Witch From Mercury. I'd happily train a TI myself if I had any idea how. If more images are needed for a data please let me know and I can start downloading a bunch.
Source: https://danbooru.donmai.us/posts?tags=suletta_mercury+&z=5
Sidenote: I'd also be interested in learning how to make a TI if anyone is interested in teaching.
https://static.wikia.nocookie.net/gundam/images/3/3a/Suletta_Mercury_Front.png/revision/latest?cb=20220806220201
Mikeneko and Uruha Rushia TI
ok, thanks anyway
walzrj erorider
https://hentaizap.com/gallery/535083/
https://danbooru.donmai.us/posts?tags=walzrj
https://twitter.com/SimpNeelix/status/1604304063226691585 this made me curious on how a chris chan ti would look like
ti isnβt enough
chris chan model
but also youβd have to pay someone for big stuff like that cuz thereβs no sane person willing to do it for free
I would never give up my sanity to make this
did someone say chris chan
i think a hypernetwork trained off of stable diffusion could work
hi <>
hello!
I am done with your thing
ahhhh!!! awesome! π im so excited
thats def his style holy
hmm
https://discord.com/channels/930499730843250783/1054741992594219089
wait let me see
if you want you can send me all the photos you have and i can blend out the artist logo of all his photos (if thats a problem with the TI) only thing i could see being a problem is his art evolved pretty dramatically over the years
it went from a heavy 2d based to semi 3d
id say its pretty close seems like its missing details im assuming this is on any?
I am not really good with prompts so this must be why
that one is is def his work i know what photos the ai referenced LOL
thank you!!
Would someone make an embed of Hifumi Togo? She is from persona 5 and sadly has very few images on danbooru ;-;
she is not given enough love'
at some point I intend to do all the P5 girls
Hifumi isn't a main char sure, but she is awesome so I'll try get her in
Yooo another legend
Looking forward to your TI's man!
have you made any already?
Yeah, lots
some of which are in this discord π
oh i have to check
are there any persona characters TI?
I hadnt found any yet
elphelt valentine https://danbooru.donmai.us/wiki_pages/elphelt_valentine
Nagao Kagetora https://danbooru.donmai.us/wiki_pages/nagao_kagetora_(fate)
up
Yae Miko
https://danbooru.donmai.us/posts?tags=yae_miko+&z=5
Where the love for Yae Sakura
you can generate very accurate pics of yae miko already, you don't need yae emb to generate it
requesting sorasaki hina from blue archive:
https://danbooru.donmai.us/wiki_pages/hina_(blue_archive)?z=1
requesting azusa from blue archive: https://danbooru.donmai.us/wiki_pages/azusa_(blue_archive)?z=2
Hina π
I'll do it
requesting a style embedding from wei xiao https://danbooru.donmai.us/posts/5589499?q=wei_xiao
https://danbooru.donmai.us/posts?tags=walter_white&z=1 requesting waltuh TI
Character: Houshou Marine
Portfolio: https://danbooru.donmai.us/posts?tags=houshou_marine
https://safebooru.org/index.php?page=post&s=list&tags=houshou_marine
Yor Briar
https://danbooru.donmai.us/posts?tags=yor_briar+&z=5
asagami fujino https://danbooru.donmai.us/wiki_pages/asagami_fujino
~~bocchi's sister TI~~
Sakurajima Mai
https://danbooru.donmai.us/posts?tags=sakurajima_mai
https://id.pinterest.com/search/pins/?q=sakurajima%20mai&rs=typed
https://wall.alphacoders.com/tag/mai-sakurajima-wallpapers
https://www.pixiv.net/en/tags/%E6%A1%9C%E5%B3%B6%E9%BA%BB%E8%A1%A3/artworks
https://yande.re/post?tags=sakurajima_mai
there's already one made by anon
Theres actually a dreambooth model for her lol
The fuck
hmm it seems no one has done any of my suggestions, are they that bad?
No! Just that this format is pretty tricky to find requests in!
Perhaps all top level comments should be just requests, and any conversation should happen in threads.
I'll look above and find one of your requests, see if I can get it done today
yeah requests should be entire forums like channel like embeddings is
<> any thoughts on this
would be messy
<:AunnThink:993596417488728174>
still better than this
right, i misread it
i do not have permissions though
xD
rip
hmm so like can u @ moderator or something
what about posting request posts in <#1044620915377451058> instead
and ppl could just
insert the pt in their post
once its done
but embeds is for downloads
right, would be messy lol
<> can u take a look
moderators dont have channel perms btw
gotta ask a director about this
arbizo made this channel
coz of my request
gg then
at what specifically
well andite clarified that mods don't have channel perms
we were asking for the requests thread to be a standalone forum like embeddings
since it's rly hard to navigate with all the convos in here
ahh, yeah. You'd need dev/director level settings. So like Salt or arbizo
<:AunnThink:993596417488728174>
Today I am doing
1. tougou hifumi (P5, reuqest)
2. Mai Sakurajima (Rascal does not dream of bunny girl sempai, Request, I just wanted to)
3. Lumine (genshin, personal)
4. Open to requests
lumin <:MomijiWide:1028509125044359259>
π corrected
I already seen a lumine embedding before...
Maybe not here
it's not there in embeds
Only TI i requested was a Marota TI because i like her art style
~~and totally not for high quality giant boba art of anyone~~
sure π
My only issue is...
Wait no
Not issue
The thing is, she does a lot of cowkini art
Lumine? Youβre awesome. I was just about to post her for one. Look forward to it
hifumi & mai are training. I don't holdout much hope for hifumi as the source images were pretty sparse... but mai should be great π
Rejoice, your request were answered <#1055460728846426152>
very nice , suppose we'll start using it
it should be renamed to 'requested-embeddings' as a final touch
ahh dope looks like ill throw my request over there
i'm deleting this thread
or just archive it ,
give people who posted earlier and got no response time to make a post in the new request embeds forum
ngl
idek how to delete it
just close the post
<:Agony:996544918610784296>
i just did
came back when azwitcher
chat here
<:AunnThink:993596417488728174>
but yeah ill let it stay for a while
<> could we get tags in the request forum? open/completed?
<>
I don't even know what that means but I'll ask
You can add tags to forum posts, and then filter by those tags π
Neat, didn't know discord did that
I'm sure if I had more than like 5 people and two bots in my discord I'd learn these neat things haha
Make sure to thank Arbizo for being awesome next time you see him
It is done
<> think its open and completed
it should be two tags <:AYAYAYA:978879333248667718>
haha yeah sorry
It has been set with open and completed. Arbizo for PoG champ today/night
[CLOSED]
Coko [Hypernetwork]
A Hypernetwork made of Coko from Cevio AI. Made from 3 images. Enjoy! Trained with AnythingV3.
Kokichi Oma [Hypernetwork]
A Hypernetwork made of Kokichi Oma from Danganronpa. Made from 15 images. Enjoy! Trained with AnythingV3.
Hypernetwork chad
<:gigachad:962417952290922586>
miwano ti
hey goobers i made ti with <>
20k steps is the first image, 2nd is 10k, 3rd is 30k, 4th is 40k, 5th is 50k
fuck i forgot clip skip 2
and the vae
Dats some fluttery art
I USED RANDOM SEEDS OOPS
you goober
Random seeds or not, still shows the difference
a bit
and users can always test
Nice
can I know what model you used to generate the sample?
anything v3
Thanks
Rolua Style TI
Experimental TI trained on a curated dataset of artist Rolua's 66 most stylized images.
https://www.pixiv.net/en/users/4094653/artworks?p=1
9500 steps. Works best with Anything V3.
Yohan Diffusion [Finetune Model]
Finetuned using kohya-trainer.
Used around 226 images of yohan's artwork. -> https://www.pixiv.net/en/users/4446354
Trained with AnythingV3 as base.
20k steps.
**Must do:**
> Set clip skip to 2
> Use Any or WD vae.
https://huggingface.co/andite/yohan-diffusion
More examples in the huggingface link cause clyde wont let me.
bocchi ti (FINAL)
No more bocchi ti, bocchi model is now what I'll use
credits to <>
bocchi diffusion
this still remains my masterpiece
Bocchi Dreammbooth
Hanyuu TI
Hanyuu from Higurashi, trained on Colab.
<>
surprisingly lewd results
no u
this TI works great
<:manofculture:585924636576055296>
a bit of inpainting
Nice π
Poxl?
yes
is that an artist known by the model?
its a different model someone posted in this server
makes images chibi style like that
uh where?
tried to search for it
https://discord.com/channels/930499730843250783/1019446913268973689/1054795370598781028
ah found it
okay
i was trying to find it too
<:MomijiWide:1028509125044359259>
do you need to mix this or it's a standalone?
standalone
dont think mixing gets great results
π ty
cleaner
Owo
Beatrice - Rezero
Made for <>
ππππ
omg
Whelp someone is gonna lewd her, that ain't me tho Beatrice is too pure
<>good job i hope we can get Rem Ram Felix π and emilia too
Why
Why
WHy oh god fucking why
someone is going to lewd her one day and im going to hate everyone here
π
Felix haaaa
You can kind of do rem/ram currently, but it mixes them up and gives you purple hair. Quite difficult to get perfect results
That said, I will ofc do them
<>I was working on felix π I'm waiting the result
nice nice π <> I wont be able to do any new TIs until wednesday
Eris Greyrat TI
These are 2 different TIs trained separately!
eris\_37 is from me, eris\_8 is from <>.
For eris\_37:
Trained for 3.7K steps, 81 images.
For eris\_8:
Trained for 8K, ~149 images.
Both were trained with andite's collab of cafeai's TI repo.
The first 3 images are for eris\_37, the last 3 are for eris\_8.
Jibril TI
Jibril from No Game No Life. Trained on Colab.
would you be so kind and upload your data set?
I'll send you a dm
mind send me your the dataset too, just curious
I'm also curious, can you send me too?
Nanahira TI
Trained on Colab.
don't ask how, somehow it came out good
OMG THANK YOU
God I love nanahira sooooo good music
<:manofculture:585924636576055296>
cultured embed
I love me some Nanahira
Truly, we live in the best timeline
1.048596%
my wife
wow a nanahira TI
I really love Nanahira. Like, a lot. Like, a whole lot. You have no idea. I love her so much that it is inexplicable, and I'm ninety-nine percent sure that I have an unhealthy obsession. I will never get tired of listening that sweet, angelic voice of hers.
god bless
o7
Mayushi TI
Shylily TI
Trained on Colab, Cafe TI repo.
Using `double ears` in negative prompt is recommended.
<>
womp womp
Thank you very much lol
π¦
nice, what model is this?
Anything
Womp Womp
Fireforce Tamaki TI
Trained on 150 imgs, 16 vectors using collab
Dataset included mostly NSFW content so yeah Gl getting her uniform (my bad) XD
<> i wanted that, i hope u wanna lewd her, id make a SFW one aswell if u dont like it xD
<:cirnoWide:930962869401571348>
Collab stopped there, should be good enoth see wich one is better for you :3
<:cirnoWide:930962869401571348>
Elbe TI
Made with Cafe's TI on colab with 35 images and 7k steps
I have no idea what I'm doing but it really resembles her so I think I won. Example images are randomly picked to be Highres fixed
patrician choice
she's my fave boat <:picardia:807578629948047420>
boats are good
based
Satania TI
Fumo TI
trained on merged weights of 50%WD1.3 50%SD1.5
I remember there was a hypernetwork of this. Glad to see theres a ti version too now
cheat sheet:
the prompts I used
((hatsune miku)), blue hair, long hair, twintails, detached sleeves, sleeveless, blue eyes, white shirt, black footwear, (simple background), (white background), art by (fumoWDSD5050-5000:0.8)
negative: lowres, 2girls, 2boys, text, error, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, out of focus, amateur drawing, 4girls, 5girls, 6+girls, multiple girls, multiple heads, head out of frame, extra legs, extra limbs, looking at viewer, fingers
Steps: 30, Sampler: DPM++ 2M Karras, CFG scale: 12, Size: 512x512
made with help of amu_neizan
I found a 2gigabytes ckpt for fumos in chat
Sucrose TI
Trained on Colab, on Cafe TI repo.
I recommend using `(long floppy ears:1.2)` in prompt with `animal ears` to get the correct look.
Many generations always get a bit different than original Sucrose. Especially with the blue hair strand and long hair. Or maybe I'm just using it wrong
Nvm, she's got that strand in game, i just haven't noticed it
Sucrose seems to be a pretty difficult character for the models (with the TI too)
Yeah, her outfit is very complicated and it definitely has troubles with it
without the `long floppy ears` it keeps generating human ears for me too
it sometimes gives her a pink hair strand
ya, bad seed
on how many images did you train?
110
Kurisu TI
Makise Kurisu from Steins;Gate. Trained on Colab, using the Cafe TI repo.
Awesome! I'll be sure to test it soon
It doesn't let me send the pic even though it's sfw
HA it let me send that one
clyde is a dick
yes
he really is
seems like the embed is pretty good , great job!
thanks, have fun with it
thanks π
i want to eventually learn how to make embeds too
Ransu_33 style TI
Trained on Cafe's colab at 6,000 steps and 188 images
You'll need to prompt for the breast size that the artist usually draws at (large)
Main drawback I noticed is short-hair blonde girls will turn into Serval
*adding (Serval:1.5), (animal_ears:1.5) in negatives helps with that issue
Other than short-hair blonde girls this TI has no problem drawing other characters
Wonderful style
Hu Tao TI
Trained on Colab, using Cafe TI repo.
It works best with Anythingv3, but should be fine with other models too.
Thanks
Jebriodo style TI
Trained on Jebriodo's art, recommend to use with words "cartoon, realistic, traditional media". It was trained on nsfw materials, so use negative prompt to avoid nudity if you don't want to see it. Also keep in mind, this artist mostly draws muscular ladies. Tested on Anything V3 model
Flaw: embedding often tries to replicate signature
kokkoro TI
trained on colab
generated with anythingV3
hands are jank
kokkro <:teriderp:959632676174839818>
kokkoro TI
Nezuko (Demon Slayer) TI
TI by <>
I hope you enjoy this TI as much as I do.
Nezuko (Demon Slayer) TI
cool, which model is this?
^^^
what should we write for bamboo?
it just kind of puts it in tbh
more often than not
But if not, try `gag` or `gagged`
the TI should bamboo-ify it
what are the settings for the TI?
To get the gag?
What model are you using? Do you get good results using other the other TIs here?
i mean the settings for Cafe TI to train this character, like steps, how many images, and what model you used for training and etc
okki thanks,
Ooo. 149 images (which is more than I would usually use). Used the NAI-wd model which is default in the Collab.
This TI is 9k steps. I tend to find my best results are 8/9k when an epoch is 12-15k steps. Basically never have I needed to go into a second epoch.
mind send me your dataset just for reference
Sure, I'll DM
me too senpi
how many the token for this emb?
<> 16
what?
16 tokens were used π
Mash Kyrielight TI
trained 67 images 6700 steps
how do i get her original armor? <:Agony:996544918610784296>
It was mainly trained on her white jacket. try emphasizing purple armor, compression sleeve, bare shoulders etc.
nice
What model did you trained it on? Waifu diffusion or Anything?
used Cafe Collab
I never used that one before.
I only use the 2 above so I don't know If it would work for me
it will work fine. that model is merge of WD and NAI.
Oh I see. right now testing it.
pls do morgan
Iahfy style TI
Trained on Iahfy's art (mix of sfw and nsfw), 7500 steps. Tested on Anything V3, CFG Scale 13
Hina TI
Sorasaki Hina from Blue Archive, trained using Cafe's TI repo.
Not a super-easy TI to use, but it's rewarding.
Recommended using `halo` and `horns` for the og look.
Using `forehead` also helps shaping her hair.
<>
God bless
πππππππππ
Gud0c style TI
Trained on Gud0c's art.
Flaw: A lot of pictures of Miruko were used for training, so don't use this embedding at full force, to avoid generating bunny ears.
To get better results type (gud0c-7000:0.85) and put words "animal ears" and "rabbit ears" in the negative prompt
LaPlus TI
La+ Darknesss from Hololive. Trained with Cafe TI repo using 170 images.
manual crop?
as always
thank you π
https://tenor.com/view/vtuber-hololive-%E3%83%9B%E3%83%AD%E3%83%A9%E3%82%A4%E3%83%96-%E3%83%A9%E3%83%97%E3%83%A9%E3%82%B9%E3%83%80%E3%83%BC%E3%82%AF%E3%83%8D%E3%82%B93d-3d-gif-26066854
thank you for doing yamada
Hifumi Togo - Persona 5
9k steps with only 43 images. <>
It works well enough but due to the small dataset can be a bit tempermental and restrictive. Be sure to prompt out `black thighhighs, pointy ears`.
Thank you very much , what do you use to make it?
google colab (for now), crawling a load of sites to get images
A tool to resize them to 512x512
normal stuff π
https://discord.com/channels/930499730843250783/1044623252326195311
Chisato TI
Nishikigi Chisato from Lycoris Recoil, trained with Cafe TI repo, using ~100 images.
based as hell thank you
my savior
Mai Sakurajima TI
~140 images in dataset.
Due to half the images being bunnygirl outfit, and half being the brown blazer you have to use negative prompts to get the output you want: such as negative prompting `(rabbit ears), bunny outift, (bunny ears)` and prompting in `brown blazer` to get her school outfit reliably.
Clyde hates the bunnygirl images, so I haven't added any!
π
Takina TI
Yoko Littner TI
Lumine - Genshin Impact TI
127 images, 8k steps
Thanks so much for this!
nice
Nice
ππ
is this anything3.0 base model, upscaled?
Mai Sakurajima [Hypernetwork]
I made this a while ago but hadn't gotten around to posting it anywhere until now. As I go to post it today I see <> has beaten me to it with with a TI. I really like their TI but I decided to post this anyway since it produces very different results.
Intended for use with Anythingv3 but seems to work ok with other NAI models.
Trained for 5000 steps on 21 cropped screen captures from the anime with a learning rate of "5e-5:100, 5e-6:1500, 5e-7:10000, 5e-8:20000"
Training images contained no bunny girl outfit(you pervs) but can produce something similar if prompted.
I recommend the tags `1girl, black hair, hairclip, ` as a base but most prompts should at least produce the general face shape and hairstyle unless you prompt something else. Character tag is not necessary.
Weird note: because several of the training images were from the shopping scene, it has a tendency to place the character in front of produce <:chenShrug:930630372348289044>
Example 1girl generation(you see what I mean about produce):
Natsume Iroha TI
first time posting, trained using Cafe's TI (50ish images, forgot how many steps taken since it's been a long while)
can't do halo for whatever reason, and sometimes it'll fail to generate her uniform, works best with `nsfw, nipples, nude` with high weighting (1.4 ~ 1.5) on negative. Why, you ask? Well..... let's just say it does funny stuff sometimes.
πππ
Amamiya Kokoro TI
trained using 90 images via Cafe's TI, first iteration stopped at 4.5k steps because I used the colab on one account a bit too much and I decided to just resume the training on different account with 9k steps.
around 70% of the times it'll generate with twintails, if you want the base hair you can trigger it by adding `long hair, two side up` to the prompt. Also, good luck generating her outfit. Also highly recommended to add `nsfw, nipples, nude` to negative to avoid *unwanted things* except if that's what you want
since I'm absolute dumb and can't really tell the difference I'll provide everything in a zip (v2 is the continued training)
Speedl00ver style TI
Tested on Anything V3, Clip setting 2, Guidance Scale 13
awesome π working on something similar
Dark Angel Olivia TI (Shadowverse)
Made using Cafe's TI repo Colab implemented by andite.
If I'm motivated I might retrain it to try to get her wings better.
Dark Angel Olivia TI (Shadowverse)
Lexington TI
Trained on Colab for 8k steps with 16 images (32 after preprocessing)
Only 16 images since she's practically forgotten in Azur Lane now and only having 55 in Danbooru lmao
Her official outfit is white but since my data has a quarter of her wearing black attires it skews it heavily. `white dress` can be used, maybe. Also, slap `hat` in the negative prompt.
Yumeko TI
Ghislaine Dedoldia TI
Uhh Forgot what I tained this with but you wanted Hot Cat mommy so here!
bro shes the reason why i got into ai art cuz she needs more fan art π₯² ilysm
Its okay homie
the embed is a little on the heavy side so uhh use Weights :3
I.e (GD:0.1) etc ++++
0.7? oh its that heavy ahahahah
<:MomijiWide:1028509125044359259>
She still comes out nice tho 90% of the time
Quite a horny girl so you need to wrestle with her attire/ appearence
is there anything that would be recommended for neg prompts with embedding?
<:MomijiWide:1028509125044359259>
None
hahah
<>
Here is a stoopid neg prompt i sometimes employ
((bad anatomy)), bad hands, bad feet, text, error, missing fingers, extra fingers, ((extra digit)), ((fewer digits)), cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, ((((blurry)))), artist name, ((((ugly)))), (((duplicate))), ((morbid)), ((mutilated)), [out of frame], extra fingers, mutated hands, ((poorly drawn hands)), ((poorly drawn face)), (((mutation))), (((deformed))), blurry, (((bad proportions))), ((extra limbs)), cloned face, (((disfigured))), gross proportions, (malformed limbs), ((missing arms)), ((missing legs)), (((extra arms))), (((extra legs))), mutated hands, (fused fingers), (too many fingers), (((long neck))), ((futanari)), ((panties)), ((censored)), ((bar censor)), ((blank censor)), ((blur censor)), ((glitch censor)), ((heart censor)), ((light censor)), ((mosaic censoring)), ((novelty censor)), ((character censor)), ((text censor)), ((flower censor)), ((convenient censoring)), ((covering)), ((out-of-frame censoring)), ((pubic hair)), ((invisible penis)), ((tail censor)), ((hair censor)), ((steam censor)), ((pointless censoring)), ((identity censor)), ((2girls)), ((3girls)), ((4girls)), ((5girls)), ((6+girls)), ((multiple girls)), ((androgynous)), ((crossdressing)), ((bulge)), ((otoko no ko)), ((pasties)), ((maebari)), ((bandaid on pussy)), ((tape on pussy)), ((yaoi)), ((tentacle hair)), ((slime girl)), ((tentacle limbs)), ((imminent penetration)), ((3d)), ((error)), ((ai-generated)), ((long tongue)), ((poorly drawn)), ((clothing aside)), ((disembodied limb)), ((disembodied penis)), ((partially visible vulva)), ((anus peek)), ((pussy peek)), ((clitoris slip)), ((wedgie)), ((fused)), ((extra ears)), ((gross)), ((((multiple views)))), ((((comic)))), ((((elbow gloves)))), ((((detached sleeves)))), ((((garter straps)))), ((((long sleeves)))), ((((cape)))), ((((capelet)))), ((((cloak)))), ((((pleated skirt)))), (multiple tails:1.5), ((((pelvic curtain)))), ((((armor)))),((((uneven gloves)))), ((((belt))))
I'd also just say swap between VAE and CLIP 1,2
<>
ahhh ty! i didnt think about using multiple tails as a neg even tho i get it so often LOL
<:MomijiWide:1028509125044359259>
Korone Inugami TI
Same dig i kinda forgot how much i trained this one
Amazing TI, very nice job!
:>
<>
I can't get her ears to be droopy like this in my gens.
so most of the times it doesn't look like her.
any thing I can do to fix this ?
You can try Masterpiece, Beautiful, A highly detailed Photorealistic illustration of (EmbeddingName:1.0)
Or add (Feral:0.6 Animal Ears)
Or put some "focus" on the hair may help π
Iono TI
Yelan - Genshin Impact TI
Two versions,
yelan, 156 images, 10k steps, randanon method - might be overbaked
yelan2, same images but split 512, ~5k steps, Cafe TI repo colab method - less overbaked but looks like it has a hypernetwork applied to it?
Learned a lot trying to make these, also learned that her dice necklace thing is like impossible to gen consistently, would probably have to create a whole seperate embedding for it π
currently also making one. ill post it if its better.
sounds good, lmk if you want my dataset or anything
i got my own. its fine
this is my take on it
was gonna post a 4x4, but the bot wont let me
it can do nsfw too. trained with a good amount on it
cool
Mike Ehrmantraut Embedding
Who said SD can't render Fingers?
put your dick away walter
i'm not having sex with you right now waltuh
Holy shit thank you
No prob, Bob
this is amazing
Eyafjalla (Arknights) TI
Trained on Cafe's colab at 7000 steps and 50 images
mai waifu in Arknights
v2 but i have no idea if this is better than the first
converting my old simple art of her from 2 years ago
Tennoji Rina (love live) TI
Generated with AnythingV3 prompt: ```(love_live),rinaC-4000, (medium_hair)```
Shirosaki Hana
Textual inversion of Shirosaki Hana. Trained with a batch size of 3 and a max steps of 6500.
batch size of 3? 3 images?
It's trained with 3 images at a time with a total of 210 images in the dataset
holy based
π€¨
Matoi Ryuko - Kill la kill- TI
Training Info:
```Epoch 9 - 15249:S
Init word: Character
Trained Words: Character, Girl, Hair, Face, Attire
31 roughly clear images
8 Vectors
Batch 2```
but, isn't she from kill la kill?
Yeah she's from Kill la kill
Whoooops
Matoi Ryuko - Kill la kill- TI
<:NotLikeKogasa:996413933449650218> <> <> It's been a few years since I watched xd
What model/blend is used in the previews?
https://huggingface.co/DrBob2142/Mix-Models/tree/main
These were made with bobs mixes
Futaba Sakura from Persona 5
Trained on only 28 images, all SFW. Trained for 3 full epochs at a batch of 3 with 944 steps each.
Nice model in the preview image. custom blend?
The mode is on my Huggingface https://huggingface.co/Kutsuya/Yukis-Mixes
Shardanic style TI
Tested on Anything V3 model, CFG scale 13, 29 steps. At full force this embedding for me worked very badly, but when I typed (shardanic-8000:0.9) or (shardanic-8000:0.8) results were very nice
Rice Shower (Uma Musume) TI
Trained on Cafe's TI, around 50 ~ 60 images were used for the training
Result can sometimes be inconsistent, best to add `hair over one eye, horse ears` though it's optional.
Sample (had to delete thumbnail because of a minor mistake):
Izumi Konata from Lucky Star
<>
<>
Konata Izumi from Lucky Star
Trained it for 1 epoch at 12100 steps. There are 363 files in the dataset, all SFW.
<>
What
<:AunnThink:993596417488728174>
sweet. anyone know if the other characters have been done?
not yet, but it might in the future π
<#1056671862261813329>
<#1056703437447889027>
Amau Ako TI
Trained on Colab for 16800 steps with 168 sfw images, examples are using the 14000 steps embedding.
Use `sideboob` with this enbedding to have more accurate uniform. Halos are almost non-existent.
All of the embeddings created from 100 to 16800, go wild lmao
Also grids from 3k to max, 1k increments (ako is 16800)
Amau Ako TI
Bronya Zaychik Silverwing N-EX (Honkai Impact 3)
Two versions. Neither will get her original outfit's shoes correctly.
bronya-9200 retains original outfit details better but more heavily influences outfit changes.
bronya_generic-12400 influences outfits less but is more prone to losing details about the character.
Both are capable of making images that will upside clyde bot.
Bronya Zaychik Silverwing N-EX (Honkai Impact 3)
Rentb style TI
embedding trained on muscle artist Rentb, can generate both "normal" and incredibly buff women. Tested on Anything V3, CFG scale 13
Warning: it was trained on nsfw images, so use negative prompt to avoid generating nudity if you don't want it
More muscular examples
IllyaTI (Illyasviel von Einzbern)
Trained on 900 Images using collab
Id recommend putting:"animal ears, cat ears, dog ears, rabbit ears" in negatives if u dont want them.
Sotcho style TI
Trained on Sotcho's art, tested on Anything V3 model (put elf ears and animal ears in the negative prompt)
Hades style TI
TI trained on the art from Hades game. I recommend to use words "realistic", "cartoon", "comic", "dark outlines" for better results
W ti S tier game
Dumb question: what model do you tend to use your Hades TI with? I've been having trouble getting quite the right look
I used Anything V3 model
might i ask how many images did you train with ?
I don't remember, but a lot
these came out really good, and i'm attempting to do this with `nichijou`, my images have multiple people and objects in it.
Can you recall if for Hades you were able to use images contains multiple people, or was it mostly single person illustrations?
One image had twins (it was a fan art mimicking Hades style), but most of the pictures used for training had only one person
hatsune miku really rocks this style!
Looks amazing!
Hiiragi Kagami from Lucky Star
Trained it for 2 epochs at 4100 steps. There are 64 files in the dataset, all SFW.
Megumin
Just a Megumin embedding. Only did it up to 1k steps.
this is actually so clean for 1000 steps wtf
it's cause she's already really easy to replicate with prompts
Albedo
Just an Albedo TI.
Decided to call it at 3100 steps.
good job!
Hiiragi Tsukasa from Lucky Star
Trained it for 3 epochs at 3000 steps. There are 37 files in the dataset, all SFW.
Houjou Satoko from Higurashi
6000 steps with 44 anime screencaps in the dataset, using Colab.
Please use with lower prompt strength for best results. For example, (satoko-4000:0.7).
Above image made with 4k steps embedding.
(satoko-4000:0.7), solo, 1girl, masterpiece, best quality, (empty background), (white background), yellow hair, short hair, purple eyes, closed mouth, happy, large breasts, gyokai, white collared shirt, pink ribbon, looking at viewer
lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, 3d, nsfw, frown, braids, furrowed brows
What model/blend is used in the trainning?
wd-nai
for this preview?
the images here is generated with anythingv3-gyokai 50/50 merge
<:manofculture:585924636576055296>
could you share prompt for these? I can never get them at a desk like this
(satoko-4000:0.8), solo, 1girl, masterpiece, best quality, large breasts, (breast rest), (breasts on table), white collared shirt, (pink ribbon:1.2), closed eyes, blush, short hair, yellow hair, closed mouth, gyokai, tired, exhausted, arms outstretched, (sleeping on a table), short sleeves, books on table, (drooling:0.6), classroom, wooden table
negatives lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, 3d, shadows, lighting, cleavage, black shirt
π
Suletta Mercury from Gundam Witch From Mercury (G-Witch)
Trained on 56 images at 7000 steps. Not sure what best prompt is atm in terms of weight. If you want to try and add to her outfit or change it you will likely need to give it a high weight such as 2.0 or something.
Looking forward to seeing what people make with this, would love to see what models work best with it.
All my images shown here still have their metadata. Just put them them in PNG info for the prompts
Model used in all images was Yohan Baked
Christmas Suletta prompts: A girl standing under the mistletoe by a christmas tree blushing at the viewer,
Negative prompt: multiple girls, lowres, text, error, missing arms, missing legs, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, jpeg artifacts, signature, watermark, out of frame, extra fingers, mutated hands, (poorly drawn hands), (poorly drawn face), (mutation), (deformed breasts), (ugly), blurry, (bad anatomy), (bad proportions), (extra limbs), cloned face, flat color, (depth of field:1.5), (blurry foreground:1.5), (blurry foreground:1.5)
Steps: 30, Sampler: Euler a, CFG scale: 7, Seed: 2327751118, Size: 768x1200, Model hash: 6569e224, Clip skip: 2, ENSD: 3133
why her skin is white on thefirst and last image? <:cirnoWide:930962869401571348>
the VAE probably
Been having fun with this TI: https://safe.aibooru.online/posts/12639
One thing you might try is `orange tie, yellow tie, gold tie` in the negative prompts.
A wild Smoletta appeares
What do you recommend for negative embeds?
Been experimenting with `naval_uniform, white_shirt, black_shirt, epaulets` but I am not sure that does much.
I think lowering the weights for the embedd and putting more on what outfit u want is ur best bet
Askziye
My attempt at a style embedding for the artist Askziye
https://twitter.com/askziye
Example prompt used:
```best quality, Cirno, skin caustics, button nose, __randomprompt9__, in the style of
Does not appear to really emulate the style, but the results seem rather alright.
omgg
Yua Serufu, Do it Yourself!!!
6900 steps with 50 anime screencaps in the dataset, using Colab.
Please use with lower prompt strength for best results. For example, (yuaserufu-5000:0.8)
Image was created with following prompts
(yuaserufu-5000:0.8), solo, 1girl, gray hair, red eyes, loli, white collared shirt, white background, empty background, masterpiece, best quality, gyokai, closed mouth, medium hair, water color style, happy, ahoge, close to viewer
Negative prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, 3d, glasses, spectacles, black shirt, black straps
https://tenor.com/view/do-it-yourself-anime-clap-yua-serufu-gif-26868268
unfathomably based
serufu burns down the house
Gyate hypernetwork
Here's something literally nobody asked for: https://pixeldrain.com/u/FpdsPsBL
btw the chibi tag with a few brackets is needed before u can even hope to get some half decent results (chances of you getting garbage is still pretty much 100% unfortunately [took me over a hundred generations to get these 4 images])
<> god among men.
And your wrong I wanted this.
Wow 607mb
hypernetwork chad <:gigachad:962417952290922586>
you bitch, this was a hypernetwork
fucking list it properly
its art embeddings
not hypernetworks
but.....a hypernetwork is a type of embed
and there's no thread for HN
lol
not bad it worked first try for me.
prompt? I had no luck no matter how much chibi I gave it
Looks amazing haha, thanks for the share!
Fumio style TI
TI trained on Fumio's art. Tested on Anything V3
Warning: this embedding is mildly nsfw
Kuki Shinobu Genshin Impact
3200 Steps, ~500 images from danbooru. Feel free to improve upon it π
500 images wow
Eula (Genshin Impact)
45 images, IDR the number of steps
looks a bit off to me. will be doing one in the next day or so. will post when its ready.
a lot of the renders are from the instagram merged model, can you compare to how it looks in your models? what about it looks off? Just kinda learning to improve that in whatever embeddings I make next
yes. hair color for some of my models are off. for some models too, output is blurry. lots of model are custom and aimed at semi-real. plus i want to make it more nsfw versatile.
and my dataset i have is around 300 or so.
will start on it tmr during work
Where did you find those eula images from?
Yeah I used like 45 images or so
danbooru, sakaku, gelbooru
haha, sure thanks very much though
gonna deliver better embeddings now on myself too πͺ
<:AunnCool_:1006253939190153296>
here are some samples with your embedding with some different models i mixed
These are dopeee
The second one
In third the color is changed for some reason and you're saying that can be avoided too
Yep. Well itβs random. Sometimes it get the color right. Could be just the model, but still canβt get rid of the blurriness. Going to start the training during my lunch break.
done. was busy, so finished it a bit late. 503 images for my dataset. trained on lots of nsfw/sfw. friendly with lots of models. can dress up her too. does not require intense weights to do so.
other model testing for samples
Lord's work!!!!
Which models are these though? They look sick
my own custom mix
i modified it a bit today
the initial ones were just anythingv3. the other ones are the mixes
I love it! To get it working do I need to put it in "embeddings" folder and that's all?
To anyone wondering: you have to use the file name as prompt if you downloaded "myeula.pt" then you gotta write myeula.pt
Yes. Drop it in the embedding folder and use that file name in ur prompt. You can rename it to anything if you like too.
Oh, that's great news. Really good job with that embedding ;D
Thanks. Check out my other ones on here if youβre interested. I also made an Arlecchino. Aka Eula alter lol. that came out great too.
Yeah, I actually downloaded that too
That was complete coincidence
Sabamen (Hypernetwork)
Here's the discord friendly version where the mods should be okay with it. Trained with 349 images at 576x576 resolution at full precision, Learn rate 0.000001 or 1e-06 for all 50k. Model used AnythingV3 and how to get her is to use the tags that she uses or otherwise it'll just be a style `masterpiece, best quality, 1girl, Sabamen, bangs, blush, bow, standing, school uniform, closed mouth, full body, grey hair, hair bow, looking at viewer, multicolored hair, pink hair, red eyes, short hair, streaked hair, two-tone hair, virtual youtuber, outdoors, pointy ears,`
Sabamen (Hypernetwork)
Just a tip for using: If you find that the colors are changing with your specific prompts then use a lower step number hypernetwork.
Thanks for providing all of them
someone made a TI for this
The last one got removed. That may be it.
Trente (trente30m) Style TI
trained on Cafe's TI, 65 images total (6.5k steps)
idk what to say about this one, but overall the result is pretty close to what I ideally want
Sunohara Shun
Trained on 289 images, 10k steps.
`0.0005:5000, 0.0001:10000`
There's a bunch of other Azur Lane/Blue Archive characters in the mega but I can't be arsed to make a thread and preview for all 10 of them.
https://mega.nz/folder/9roxBAgB#ADC1xp6GY8j4K3h7F_c7fA
Okabe Rintarou from Steins;Gate
Trained it for 14 epoch at 21203 steps. There are 45 files in the dataset.
extremely based
Yor TI
Missymuscle style TI
This embedding mostly generates muscular (and busty) women, but you can try it for other things too. Tested on Anything V3 model
Sonozaki Mion
6000 steps with 70 images in the dataset, using Colab. TI is less overfitted so it should perform better than the other ones I uploaded <:momi:930625714464632842>
Title image created with the following prompts:
mion-6000, solo, 1girl, green hair, parted bangs, green eyes, blush, happy, white shirt, yellow vest, (red tie), looking at viewer, close to viewer, white background, empty background, masterpiece, best quality, gyokai, ponytail, forehead, closed mouth
portrait by bad-artist, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, 3d, twintails
Watanabe Akari - More than a married couple
Hi everyone, this was my first embedding, so i don't think it was perfect.
I hope you understand <:SannyoSmile:1032853715956273224>
`Trained it with 60 files in dataset, all SFW! 30000 steps.`
**__Tip:__** add `blue eyes` prompt! The AI made a lil mess with it, maybe because i just used anime scenes of her.
.
I think you forgot to post the file π
OMFG! Just a little detail...IUAHIHAIUHAUHIAU
Also it seems the Ai messed Akaris eye color up since hers is green/aqua blue
<:cirnoWide:930962869401571348>
Zankuro Hypernetwork
For model [925997e9]
*model hash*
hypernetwork gang <:gigachad:962417952290922586>
also what settings did you use?
I'm doing some weird shit by using commit `bd4f0fb9` of webui before multi-layer hypernetworks were added, 0.0000025 learning rate over 30000 steps, post-multi-layer addition requires too much VRAM to load my input data
78 768x input images tagged with deepdanbooru and then tags tweaked manually to remove major things like incorrect characters
did u remove the artist tag from the prompt files before training btw
maybe thats why all my hypernetworks are so shit
<> ping
yes
<>
a
why hash...
updated model
Mafuyu Diffusion [Finetuned Model]
Finetuned model based on Anything V3 using 200 images with tags related to "cat girl, cat ears, animal ears" scraped from danbooru and pixiv public images.
The model seems to have a strong bias towards a specific Japanese Artist: "γΎγ΅γ"'s drawing style, hence the model name. This is despite the fact that after the training, a dataset reverse search showed only 67 of 200 images used is originated from the artist.
Credits: training process heavily inspired by <>'s Yohan Diffusion dataset.
prompt examples on civitai.
https://civitai.com/models/2706/mafuyu-diffusion
img2img:
any tips on how to use?
nvm the civitai has it
Ai Hayasaka - Kaguya Sama
bad-image-v2 negative embedding
Works like a supercharged bad-artist.
Replaces your negative prompt.
Trained off of experimental12b, may not work with other models (tested with ElysiumV2 and seemed fine)
Yes and the hands do be like that.
https://huggingface.co/Xynon/models/tree/main/experimentals/TI
if you want the model used in the sample photos you can find it here (experimental12b):
https://huggingface.co/Xynon/models/tree/main/experimentals
bad-image-v2 negative embedding
(all pictures with experimental12b model, I do not guarantee good results with anything else)
v1 exists too but it's worse, only uploaded for legacy purposes
whats the experimental12b model?
it's in the repo
experimental model
as the name implies
ah
it looks good!
thanks!
is it possible to load these models with the StableDiffusionPipeline?
I wouldn't see why not
though I haven't tried
do i need a vae or will it work with the default sd one?
it works without a vae, but I always run with wd1.3 or fd vae though since it's better
cool
thanks!
is it good when you mix it with the bad-artist emb ?
which training step version have you used in the preview images?
never tried, but probably not as they would conflict
you could try though
mostly 39000 steps, some 27000 steps
you can take it from the exif
I forgot to mention, it's all clip skip 1
outdone yourself
I've been trying out the embedding, and the quality of images it makes is great and I appreciate the work you put into it, but there is one big issue I've run into (and based on your preview images, it seems that it's happening for you too). When using the embedding, I often get compression artifacts generated in the image, sometimes to the point of ruining the entire picture. They aren't very noticeable in yours since you have a busy background that masks them, but you can see them around the characters' hair and on their faces. Here is an example where they are more obvious (using the elysium anime model). You can see them all around her outline and even in the colored areas. I noticed all of my images are fuzzy or have a weird texture (as seen in the attached image) when using it and I think this is part of that. I've tried AnythingV3, elysium anime, and experimental12b, and it occurs on all of them, even with clip set to 1. Without the embedding, those models produce crisp images for me. Sometimes adding the normal NAI negatives removes/lessens the effect, but on others it doesn't, so mostly random. I'm wondering if there is something I can do to stop them from generating like this like maybe I have some settings wrong. edit: turning on the VAE can sometimes help reduce it but not all the time, so might just be a placebo
The embed seems to mess up ahegao. But otherwise works pretty well.
I also get fuzzy details around the eyes or patterned elements - just like in your image
yeah I've seen this occur sometimes but I haven't seen it be that severe
I'm currently not at my computer but could you send a screenshot of your webui parameters?
I have a possible idea of what might be causing this so I'll try recreating that image later
thanks for helping. here are my settings. it's a long story, but just know that i'm not up to date at all with the current webui, but hopefully it isn't an issue here since it doesn't seem to be an isolated incident
Yeah I'm noticing the crunchiness more and more, not entirely sure why
Don't use it as your first negative embed token. I've been testing out different combinations of negative embeds lately and figured out the further it is away from token 1, the better the image is going to be on the side of crust.
This is a sample image using the combination `
oh damn, you're right, more distance from the start of the negatives does help
this should probably just have the same effect as lowering the weight,
such as (bad-image-v2:0.6)
also, sorry for not getting around to finding the issue,
I've been pretty burned out for the past few days
FIRST TIME
No rush, just get around to it when you can
I've tried reducing the emphasis too, but it doesn't seem as effective as giving it distance and other negatives to worry about
Arina Makihara from Waku Waku 7
Character from a retro game with fluffy short hair and long, floppy bunny ears. 6000 steps with 40 images in the dataset, using Colab. NSFW images were included in the dataset, so negative prompts nsfw, nipples, large breasts may be needed.
Title image shows the character and outfit that this TI was trained on.
Prompts used in title image: arina-6000, solo, 1girl, white background, empty background, (blue bodysuit), cleavage, brown hair, short hair, (long droopy dog ears), brown eyes, looking at viewer, medium breasts, chubby, muscular, close to viewer, white vest, red fingerless gloves, goggles on head, evil smile, blush, crystals, sparkling eyes
negatives: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, 3d, nsfw, nipples, large breasts
<> Very nice, wpuld you mind sharing your traning parameters?
it's the default/recommended from colab <:SakuyaSweat:956420860951031849>
this is one of my first TIs so its by chance it turned out well
SadAtmospheric
https://huggingface.co/datasets/IkariDev/SadAtmospheric
Alisa Character from artist Momopoco
Trained until 10k steps (the last one got corrupted).
2k step is the best
prompt `upper_body` to consistently get the face (I might have put too many body references in training data).
also prompt `black skirt, white jacket, school uniform` to semi-consistently to the school uniform.
Generated on Anything-V3
Trained on NAI
checkpoints-20221228T195338Z-001.zip
all embedding checkpoints
9750 step checkpoint got corrupted as well
haha yea 9750 really did not care
I'm not the one who asked for it, but thank you all the same! Love Momoko's work
Imma work on artstyle the next chance I get
I never did style before so does black and white data affect the result?
Some artwork donβt have colors
I don't know, unfortunately. I imagine it could
Iβll just use colored ones then
Thx
Naoto Shirogane (Persona 4)
First attempt at an embedding, I think it came out pretty well.
Trained on 112 images to 10000 steps. I mainly focused on getting the hat and the hair correct, since she has a bunch of different outfits.
bro that first image tho
Bluethebone Style TI
Trained on Anything v3 for 25k steps on ~250 images from the artist Bluethebone (https://twitter.com/bluethebone).
Using the prompt at full strength leads to corrupted results - I've found that using it at 0.75 weighting strikes a good balance.
Do note - this TI was trained on mostly NSFW, so it may spontaneously generate nudity.
LAM style TI
trained from 87 hand-picked images on pixiv of the artist https://www.pixiv.net/en/users/17429 to 20000 steps
i personally use 20000 steps version for more style and 14000 version if the 20000 overcooked with prompt
can't seem to download the files, discord's giving me a 403 error?
not sure ill reupload them
can you check it?
<> yeah downloads are working now, thanks
It nicely replicated LAM's eyes, wow
<>
mind if i do a model of this artist
np
what were the settings for this? and, which model did you use? nai or the anything model? π
for training i use the default model on andite cafe ti colab nai-wd
anything-v3 for previews
scale at 5 DPM++ SDE at 10 steps highres on
Marin Kitagawa TI (Sono Bisque Doll wa Koi wo suru My Dress Up Darling)
7k steps, 202 images.
Responds to Anything 3.0 really nicely, with both outfits available via prompting.
Most images had her doing the v/peace sign, so I would prompt that out as well.
it let me post a bikini example!
β€οΈ
Artist Momopoco Artstyle TI
Artstyle TI trained on 300 images.
Trained up to 20000 steps
Images generated with megumin TI I made for fun
Too lazy to find the best one I'll just attach then all on here
xy grid of all embedding: https://drive.google.com/file/d/1i_f1_ckt8LuVm5YNcROoGJDkpXJ5ys3o/view?usp=sharing
checkpoints-20221229T200352Z-001.zip
I might have forgot to use same seed :p
im not waiting another 30 minute for it to generate sry
It's all good! I think the grid still gives a good style comparison.
I'll play around with some of the checkpoints later and post what I find out
alr, thank you
i think all of them works pretty well tbh
whats the y axis?
the strength of the embedding, its very strong
Two experiments, one on Cocoa and one on Anythingv3, both with the vae. The basic prompt was as follows, where momo\_simp\_# is a checkpoint from the zip
`sitting, 1girl, momo_simp_5250
Negative prompt: bad-image-v2-27000, bad-hands-5, bad-hands-3
Steps: 28, Sampler: DPM++ 2S a Karras, CFG scale: 7, Seed: 3290092727, Size: 608x960, Model hash: 625a2ba2, Denoising strength: 0.7, Clip skip: 2, ENSD: 31337, First pass size: 0x0`
General comments: Works better on Anything as Momo's art clashes with the Yohan base of Cocoa, the TI really likes miqo'te if you don't specify away from it, and probably I should have run more than 28 steps to get better hands. However, the style on Anything is *very* Momoko!
Holo by kawakami rokkaku (hypernetwork) v2
So here's **BEST GIRL** trained on one of my favorites artistes. Trained on anythingv3 at 690x690 1e-06 for all 50k. Tags to use `holo, kawakami rokkaku, 1girl, brown hair, long hair, wolf ears, animal ears, wolf girl, wolf tail, red eyes,` but I think you can get away with just using `holo, 1girl, wolf ears`. This is v1 I will make another one to hopefully get closer to the artstyle and better quality.
Here's the training previews.
The tail wants to become a wolf alot tho.
thanks for best girl hypernetwork!<:babStare:1016190056895103006>
https://cdn.discordapp.com/attachments/1019446913268973689/1059343982880104468/xy_grid-0028-3273442385-masterpiece20best20quality20holo201girl20collarbone20shirt20brown20hair20long20hair20blunt20bangs20red20eyes20wolf20ears20animal20ears.png
https://cdn.discordapp.com/attachments/1019446913268973689/1059343193663078490/xy_grid-0027-411636409-masterpiece20best20quality20holo201girl20collarbone20shirt20brown20hair20long20hair20blunt20bangs20red20eyes20wolf20ears20animal20ears.png
https://cdn.discordapp.com/attachments/1019446913268973689/1059342446393311322/xy_grid-0026-347102059-masterpiece20best20quality20holo201girl20collarbone20shirt20brown20hair20long20hair20blunt20bangs20red20eyes20wolf20ears20animal20ears.png
https://cdn.discordapp.com/attachments/1019446913268973689/1059344858021634058/xy_grid-0029-1868127497-masterpiece20best20quality20holo201girl20collarbone20shirt20brown20hair20long20hair20blunt20bangs20red20eyes20wolf20ears20animal20ears.png
https://cdn.discordapp.com/attachments/1019446913268973689/1059345647473528872/xy_grid-0030-3159338339-masterpiece20best20quality20holo201girl20collarbone20shirt20brown20hair20long20hair20blunt20bangs20red20eyes20wolf20ears20animal20ears.png
So V2 is a little better
less red hair highlights
Holo by kawakami rokkaku (hypernetwork) v2
50k-80k in there
Miorine Rembran
Request: https://discord.com/channels/930499730843250783/1057047354223640647/1057047354223640647
`Trained it with 72 files in dataset! 30000 steps.`
Tip: Add `gray eyes` prompt! Her hair is a lil messy to understand, and I dont know why the AI didn't get her eyes color π€
Are you willing to train one that doesn't feature her uniform? I'd like to be able to prompt her in other outfits.
Put enough negative weights on the outfit and weights on the positive prompts and u can get her out of it
You can do it!
In the examples that I posted, I putted `green uniform` prompt
Thanks, I'll give it a shot
Testing Casual and Maid dress
Maid Miorine is a major mood
Arlecchino (Genshin Impact)
as by request and plus it was an embedding i was planning to do. 10k steps. 400 image for my data set. friendly with every model. can do nsfw. was trained with lots of variety of outfits, nsfw poses/actions, and sfw. enjoy.
getting the x-shaped pupil is impossible. so, oh well. photoshop that in yourself. that's what ill be doing.
Arlecchino (Genshin Impact)
May (Guilty Gear Strive)
Two different embeds, since her anchor kinda polluted the training.
FINAL was trained on the same images with the anchor edited out, but I still think the first embed is slightly higher quality if you don't get any weird chunks of anchor.
In retrospect, FINAL is probably good enough. I'd probably stick to that, but more options never hurt
what model did you use for the more realistic looking one?
Baobhan Sith (Fate Grand Order)
Baobhan Sith TI trained on 50 images and 10k steps
God bless you
My first successful TI I hope yall enjoy it
Itsuki Nakano (Quintessential Quintuplets)
IRyS Embedding
IRyS embedding trained at 6.5k steps
Oh shit she doesn't have a fucked shnozz
theyve updated her model
Tsukumizu (tkmiz)
TI embedding, trained on NAI, 16 vectors per word, .005 learning rate, 35000 steps.
dataset: 179 items, captions and images downloaded from booru
I originally posted this here, with some of my other embeddings: https://gitlab.com/mwlp/sd
August Von Parseval (Azur Lane) TI
Trained on Colab for 14k and 11k steps, her base skin and her maid skin.
This embed is entirely seed dependent, you **will** roll a lot to get her likeliness, but most of the time it gets 90%.
Recommended prompts:
`(hair over one eye:1.2), black thighhighs, skirt` for her base skin
and just `(hair over one eye:1.2)` for her maid skin.
For some reason the TI likes putting her arms up so slap `arms up` at the negative. Idk how to fix the ribbon on her mouth on the maid variant.
Clyde doesnt let me upload sample images its not NSFW damn it <:Angry:689707904277676098>
Trained up to 16 on avp, only 11k on -maid
Rise Kujikawa (Persona 4)
Back at it again with the P4 girls.
10k steps, 175 images. Trained on Clip Skip 2
Trained to emphasize her school uniform over her idol uniform, but both were in the dataset, so you might make that work too, who knows.
Atago Race Queen (Azur Lane) TI
Original post kept getting nuked by ClydeBot even when censored to hell and back, so face only it is. Direct all complaints of censorship to Clydebot, thank you.
https://mega.nz/folder/9roxBAgB#ADC1xp6GY8j4K3h7F_c7fA
Technically not my embed, as training data was kindly provided to me by another person who taught me how to train embeds in the first place. I only ran the training process.
Power (Chainsaw Man)
Trained up to 1600 steps. Sadly does not capture her pupils. Attached are step counts of 1600, 1500, and 1000.
Iino Miko (Kaguya-Sama Love is War)
Trained for 13500 Steps
Found 4000 steps to work the best
heres the other checkpoints.
checkpoints-20221230T055814Z-001.zip
Nice work. I have a chika embed I will share, so then it's just kaguya left
i wanna do it
lemme do it
or did you start already
fire
why not do kei as well π
kei shirogane?
huh
for some reason i always thought its rey or something
:/
ye kei shirogane
go for it!
Although I have just started one (I'm too impatient and am doing some yuri picks atm with chika and ai, need kaguya to complete the scene!)
((and I had already collected the dataset)) Let's just both do one π
gib prompts π
https://discord.com/channels/930499730843250783/1044620915377451058/threads/1059163282851770438
kei is a must
imma do kei then
did the kei one
https://discordapp.com/channels/930499730843250783/1059546518845132862
π awesome!
Nakano Itsuki from Quintessential Quintuplets
Trained it for 1 epoch at 15599 steps. There are 311 files in the dataset, all SFW.
uh , sushi salmon did this embed already
Let's go, embed war! π
oh, i didnt see, it was already training for a while D:. but then again, who cares! the more embeds the better!
WD 1.4 Megumin
A very poor Megumin TI trained using WebUI (Only way I know how to train using 1.4 as a model)
PS: Attachments reuploaded below, since apparently they did not upload correctly the first time.
Bizarre <:ASthink:582656811568594945>
works now <:ReimuSchizo:962460660397113414>
I really like it. It's awesome.
Reisalin Stout (Atelier Ryza) TI
16k steps trained on colab, 190-ish images (no NSFW but includes rating:q, and score above 50 in danbooru)
Cropped to faces since I know Clyde will block this again.
I think I borked my dataset since the outfit is not 100% accurate and leans more towards bikinis
.zip contains steps 3k to 16k
grids, i forgot to lock the seeds but yeah
edit: why is this not blocked wtf
>probably broken because TI leans towards bikinis
I'm not gonna complain about a nice feature
Good TI!
I was aiming for her outfit and that's why I nuked the NSFW from the dataset lmao, but yeah still a nice outcome
What model did you use for training? I tried to make my own TI with Colab, but it's not half as good as this one.
The default on the colab, `nai-wd`
I see, thanks!
And one more question, how many images do you recommend to make a good dataset?
I want to make an embedding of Mio from Xenoblade 3
**Greater than 100 is ideal**, but you can get away with 50 or less. (I've trained with only 30 ish before <:XD:281010861802520576>)
The more complicated the character you'll probably need more than 50
I will look for that amount, thank you very much <:momi:930625714464632842>
Kaguya Shinomiya (Kaguya-Sama Love is War)
9k steps, 144 images.
can do other outfits \o/
what prompts did u use to get consistent school uniform outfit and the shrine maiden outfits o
<> the images should have the meta in them still
let me know if not
ahh
thx
for the school uniform, theres literally no prompts
its just the default as 90% of the training images used it
Sushi Yuusha Toro (Black Souls) Hypernetwork
My first hypernetwork for WD1.4 e1
https://civitai.com/models/3685/black-souls-hypernetwork-for-wd-14
Using "sketch" or/and "jaggy lines" can improve results and I do not recommand to use WD 1.4's "ideal negative prompt" (worst quality,normal quality,lowres,etc...) except maybe "bad anatomy" and "jpeg artifacts".
20k steps with 149 images.
I also used the 1024 module.
||I'M GOING TO ALICE||
I KNEEL
Yo this is amazing
I made a toro hypernetwork before as well but this seems fairly more updated
Zero Two (Darling in the Franxx)
Completed. Data set of 210 images. 10k steps. Can do NSFW reliably. Compatible with all models. Enjoy. will post other images shortly after this post.
Great results with this.
Which model is this? It's awesome!
<> which outfits was this trained on?
Random stuff. Bunny suit, her bodysuit, causal, lewd stuff, etc. I only outfit I pruned was her bodysuit, but all generation are here in her default red outfit
Thank you. It's a model I've been working on fine tuning. I'll send you a message when it's ready for release.
Devmgf style TI
TI trained on Devmgf's art, can generate both sfw and nsfw, "normal" and very buff girls
Kei Shirogane (Kaguya-Sama Love is War)
Trained with 15 images (I have high standards alright?)
Trained on NAI, generated with Anything-V3
Trained for 20000 steps
Anything after step 1000 all look fine
15000 step save is attached
checkpoints-20230102T180711Z-001.zip
only 15 images, what a gamer
basing in your samples `:o` is guranteed, it seems like the images you used have the same expression <:AMxd:745244205512785940>
Elin From Tera
The zip contains the embedding's and the small dataset I used.
Generated using Anything V3
<:SakuyaSmug:1001217158430339103> Nice
(Elin 4 as I kept missing settings when making the embed)
Made using DreamArtist
So this seems to be more of a specific look of animal girl than the orginal
The first embed is at the top and is more like the Elins from Tera than this experiment.
Ranni the Witch
Arise, ye tarnished
https://mega.nz/file/tWNhGLSJ#v_zi8qQ8aHnYt3WvhLKcUesBl3pWpbgdkygLq0J6MPg
Doll joints mixed with multiple arms can lead to janky generations. LORA training is probably the way forward to get an on model Ranni. For now you have to hope the model forgets anatomy.
I like the 2000 step count version, but packaged all my training incase anyone wants to be explorative.
Her fur cloak is difficult. (fur cloak:1.2) at the start of the negative works but requires some finessing because it seems linked to the size of her witch hat.
The results are genuinely fantastic mate, do you know how many images you used? And it was done on LoRA?
Roughly 100 images
It's just an embedding, I'm just waiting for LoRAs to get more support before I start doing them.
ah
||love the samples by the way||
<:MomijiWide:1028509125044359259>
lol thanks
<:the_fool:934663674717675560>
thxxxx
very nice!~
what is the password?
Fantastic. Is the keyword "ranni the witch" with spaces?
it's a TI so it's whatever you name the .pt file
Thanks
Atsuko Kagari From Little Witch Academia
Really like this, any chance for you'll share your embedding?
I just forgot to upload it
I'm gonna delete this post and reupload it with the embedding at the top
<:thanks:770223821172113428>
Atsuko Kagari from Little Witch Academia
Wasn't paying attention and didn't notice that I forgot to upload the TI in the original post.
The hands
https://tenor.com/view/yes-patrick-admiring-wow-amazed-gif-14338920
Fallenshadow (vtuber) TI
Primarily trained on portraits so it won't understand much about her costume except the colors.
Ilya (Princess Connect!) TI
Trained on colab, 4k steps on batch of 3 with 96 images
Batch of 3 so this is technically **12k steps**
Her mouth can by funky at times, so slap `skin fang, tongue` on the negatives.
This embed is NSFW generation friendly
`.zip` file contains steps 1k to 4k
Head only crop for samples again since I do not trust Clyde.
Grid of the `.pt`s
<>
<:prayge:899678417077800971>
