Lydia (high res + Lora guide) (Patreon)

Took me a while but I finally wrote it! This is the method I've been using for now, I still need to learn a lot so I'm sure I'll change many things in the future. I tried keeping this guide as simple as possible so if I left something out feel free to ask.

----

Webui

1. Download this https://docs.pinokio.computer/download/
2. Once installed, open it, click discover and scroll down to find "kohya_ss"
3. Click download, then install and launch it

Training data

• For my use case and workflow, a single image of the character's face should be enough and is what I did for all of mine.
• I usually make an upper body shot of the character directly facing the viewer with a simple grey background. I make these images with my usual workflow.
• The way the image looks influences what the lora creates quite a bit (things other than the character too), I'm still testing many things at this step so I'm not sure what the best way of doing it is but most of what I've tried ended up being fine or great.
• Making and adding more training images is what I'll be looking into next, so for now I don't have much to say about it other than it should definitely improve the lora if done well.

Training and usage

• Open the webui, go to the LoRA tab, in "model quick pick" I choose "custom" and use my model.
• Somewhere on your pc create a folder and inside create 3 more, one called "img", one "model" and one "log". Back in the training webui, go in the "Folders" tab and copy paste your folder paths. You can change the model output name too if you want.
• In the "img" folder, create another folder called "1_charactername female" where you will place the training image (for this method I don't caption the image).
• Then in the Parameters tab I use these settings:

Train batch size: 1
Epoch: 150
Save every 150 epochs
Caption extension: .txt
Mixed and Save precision: bf16
Uncheck Cache latents
LR Scheduler: constant
Learning rate: 0.0005
Optimizer: AdamW8bit
Resolution: 512,512
Network Rank and Alpha: 16
Noise offset (in "advanced"): 0.05

• You can click "Configuration file" up top and save the parameters you set as a file which you can later load again.
• In the "advanced" tab you can also check "Color augmentation" and "Flip augmentation" (but don't leave them checked when saving the config file because there's a bug that re-checks Cache latents if you did and they don't work together I think).
• Now you can click Start training. It only takes me about a minute on my gpu but even on slower cards it should be very quick.
• Once it's done, you will find your lora file in the model folder you created, you can copy and paste it to your sd webui directory and use it.
• Another tip: I like to move or delete everything out of the "img", "log" and "model" folders once I'm done so next time I load the config file I can just reuse it without new folder paths.
• To use your lora, in your sd webui prompt this "..., <lora:loraname:0.4> charactername female, ..." and I still include keywords for almost every attribute of the character and other parts of the image and use my usual workflow.