Anyway to extract the lora weights into diffuser types?

[tingxueronghua]

Thanks for your efforts and contributions.

My question is that, it is quite hard to use scepter because all the modules are deeply encapsulated and it is really hard to modify it without any documents. I really want to find out how to load the model weights by peft.

Is there any suggestions on how to convert it to convenient formats?

[tingxueronghua]

By the way, it seems that the module names are not aligned with Flux-dev-fill, which makes it hard to convert.

[maochaojie]

Please refer to function ‘merge_comfyui_lora’ in ‘https://github.com/ali-vilab/ACE_plus/blob/main/modules/flux.py’ to complete the conversion of LoRA weights to the original parameter format. The correspondence between the names of diffusers and Flux-dev-fill weights can be referenced in the merge_diffuser_lora function found in https://github.com/ali-vilab/ACE_plus/blob/main/modules/flux.py. In fact, diffusers are already compatible with models that use different parameter names such as comfyui and blackforest.