When directly loading the LoRA of the ICEdit project with ComfyUI, it prompts "lora key not loaded," and the image quality is very poor. I analyzed the inference code of the ICEdit project using AI, and it appears that the message "lora key not loaded" originates from the additional four expert layers introduced by the MOE structure. Using some AI-generated code, I applied a simple weighted average to these four layers and obtained a model that can be loaded normally in ComfyUI. However, this modified model cannot be used together with the "nunchaku" project.
To address this, I first merged the LoRA with the "flux fill" model and then extracted it again (this is the only simple method I know), resulting in a LoRA compatible with the nunchaku project. From my testing, the performance is somewhat better than directly loading the original LoRA using a UNET loader. Although dynamic weighting is no longer available, at least it works.
averaged_ICEdit-MoE-LoRA.safetensors is a model obtained by using AI code to perform weighted averaging on the ICEDit LoRAs;
average_icedit_lora_rank_32-bf16.safetensors is a compatible nunchaku model obtained after fusing the averaged LoRA with FLUX.1-Fill-dev and then extracting it again.
