comfyanonymous
Support "transformer." LoRA prefix for Z-Image
This PR adds it to Z-Image the "transformer." LoRA prefix that is supported for other models like Qwen. For some reason the code for Lumina2 is used for Z-Image.
Please try to be consistent in your LoRA loading code. Thank you.
Can you give me an example lora of this format and put which tool produces it as a comment in the code?
Can you give me an example lora of this format and put which tool produces it as a comment in the code?
upcoming OneTrainer support, and yes I could do that, but my thinking is that there is a prefix that works for everybody and especially is the same for all upcoming models - so commenting that this is for OneTrainer might not be good.
Why I think this is a good prefix for Qwen, Z-Image and more:
Those models use diffusers keys, because they originally started out as diffusers format models. It is likely that future model releases also do this, because they want day-0 diffusers support when they release the model.
Diffusers format follows a directory structure for the components. Using that as prefix in the lora to denote for which component that key is seems logical - and isn't new, it's already implemented in Comfy for Qwen:
If you don't like this solution, no prefix is also an option - so same keys as in the transformer files: https://huggingface.co/Tongyi-MAI/Z-Image-Turbo/blob/main/transformer/diffusion_pytorch_model-00001-of-00003.safetensors
but currently a `diffusion_model." or lycoris prefix is expected