help with training

#1
by TahirC - opened

I am training a ZIT style lora (drawing style) with 33 images with captions (generated using qwen3.5-4b) , at 1.2k steps i am not seeing any change is output sample

any advice ?

Sorry for a late response. If you are still having trouble with it, there's a few things I'd suggest you check. You should definitely be seeing differences at 1.2k steps.

  1. Trigger word. If you had a trigger word, check if the sample image prompts include it.
  2. Learning rate: You might have too low of a learning rate
  3. Captions: If your dataset captions are too detailed, it might be an issue. Also, check what your caption dropout rate is as well.

I was not getting expected output in training sample generated via ai-toolkit but when I ran it on comfy it was working fine

TahirC changed discussion status to closed

Sign up or log in to comment