BIGUS'S GUIDE ON HOW TO TRAIN LORAS LOCALLY (ON YOUR COMPUTER)


# THINGS THAT YOU WILL NEED #

Python 3.12.9
https://www.python.org/downloads/release/python-3129/

Python 3.10.0
https://www.python.org/downloads/release/python-3100/

OneTrainer
https://github.com/Nerogar/OneTrainer


# INSTRUCTIONS #

1) Install OneTrainer by pressing the bright green <> Code button and Download Zip
2) Unpack the file, put it wherever you want, then run install.bat
3) You will get a command window open up. If you installed everything correctly, it should start installing torch and other dependancies.
4) Once its finished and it tells you to close the window, run start-ui.bat

5) The UI of OneTrainer will open. From here, it's as simple as following this guide on OneTrainer's wiki to get your first generation set up.
https://github.com/Nerogar/OneTrainer/wiki/Onboarding-Guide-for-Newcomers


6) Once you get the basic jist of the instructions, there's four things I recommended to set up.

6.1) Set up image sampling by going into the sampling tab, pressing add sample, and writing in a basic prompt for the AI to run and generate an image with. I personally recommend setting "sample after" to 5, and the blue thing next to it to EPOCH.

6.2) Go into the training tab, set your optimizer to PRODIGY, Learning rate to 1.0 - then scroll down a bit and toggle off both Train Text Encoder 1 and Train Text Encoder 1 Embedding.

6.3) Go to the backup tab, set Save Every to 0 and Never, and set Backup After to 5 and to EPOCH. This will save your LORA every 5 epochs.

6.4) Head to the model tab, and under Base Model, go look for the best model you normally use for Stable Diffusion and just select it. I personally use Noobai-XL Epsilon Pred 1, and it works just fine.

7) That's pretty much it. Head to LoRA, set LoRa Rank to 64 if it isn't set already, and now you're ready to train. Just press the bottom right button "start training".


NOTE: For me, I have an RTX 2060 and it took me more than 18 hours to generate roughly 70 epochs and 2241 steps. There is no set standard on what point the LoRA can be considered "ready", however, if you want to get a good idea if your LoRa is working, then check the tensorboard while training is ongoing. It's the button next to "start training". From there, you'll get a tab in your browser and you can check the image tab to check all the image samples you had set up to generate over time.

8) Once you've generated enough steps/epochs and are satisfied with your image samples, then simply stop training, and head to the workspace directory you had saved (if you were following the wiki guide as instructed). Then it's merely a matter of going into the backup folder, picking the level of steps at the end of the filename you're happy with, and then just take out the .safetensor from the folder and rename it to whatever you want your lora to be.