santa hat
deerdeer nosedeer glow
Sign In

Problem with training character lora with too details

I have collected like 20/30 images of thic character but after i train the model (i did it like 3 times) it turn out good but with an orribile face at every image generated. I used 10 repeats and polynomal for this training should i add more repeats?.. Change model.. If someone could gave me some advice (beside creating a folder with only the face) i would be greatfull

2 Answers

I can't say what you would need to change, just what works for me. I use cosine with (3) restarts and keep the learning rate quite low at 1e-4. You should aim to have somewhere around 400 steps/ images*repeats per epoch. Then it takes somewhere around 15 generations. And don't forget that often even the best trained models create terrible faces, unless you use the hi-res fix.

Check if the resolution of your images is at least 512x512. If not, you should probably upscale them. When training with kohya scripts, make sure to check the "enable bucketing" flag, I had big troubles with quality when it was off.

Try adding more repeats, for characters I usually do at least 6000 steps. With a 30 images dataset, that would be 200 repeats (or epochs, they are interchangable).

For a dataset this small, I think any lora archtecture will do good. But you can check some more presets here in civitai articles, in the training section.

Your answer