Training settings for the Krea 2 LoRAs

#34
by Gaxili9741 - opened

Hi Malcolm, thanks a lot for the Krea 2 person LoRAs, and for sharing them the way you do.

I tried to follow the karengillan set you shared: tight square crops, one face, and no .txt captions at all, only the trigger word. The one difference I made on purpose is the resolution: I trained at 1024 instead of 512.

The trainer is ai-toolkit, on Krea-2-Raw (bf16 raw.safetensors), with the local Qwen3-VL text encoder and Qwen-Image VAE, flowmatch, linear timesteps, AdamW 8-bit, lr 1e-4. First run was rank 16 / alpha 16, base in float8, about 50 steps per image. Second was rank 32 / alpha 32, base in convrot8, 2000 steps. Both come out usable, but in Comfy they need about 1.6 where yours hold at 1.

Looking at one of your files (krea2_terihatcher_v1_onetrainer), it is rank 16, OneTrainer master revision 192c6aca, and every layer has an alpha tensor of 1. Comfy then applies alpha / rank, so 1/16. Even after that, the attention updates in your file are several times larger than mine. That is the part I cannot reproduce.

If you ever have a minute, the settings that would explain it are: rank and alpha as set in OneTrainer, learning rate and optimizer, steps per image, the resolution you actually train at, timestep weighting, whether the base is quantized while training, and whether those alpha tensors of 1 are written on purpose at export.
Also, does that dataset actually train well with no captions, just the trigger word?

No rush at all, and thanks again for keeping these coming.

Sign up or log in to comment