crj/

tokghibli

Since the current Nunchaku node cannot load the original Insert Anything LoRA, I used the following steps to make it compatible for use with the Nunchaku node in ComfyUI: I merged the original LoRA with the fill model and then re-extracted it to obtain a LoRA that is compatible with the Nunchaku node. The exact principle behind this is unclear, but it might be due to some unrecognized keys being lost in the process. Typically, this results in a slight drop in quality.

Additionally, insert-anything-workflow.json is a workflow for ComfyUI that, although designed for the Nunchaku node, can also be used by replacing the Nunchaku model loader with ComfyUI's native UNet loader node. Considering that the original LoRA has a rank of 256 and is quite large, I have created a version with a rank of 64: insert-anything_extracted_lora_rank_64-bf16.safetensors. Simple tests show there are minor differences in the results, but no significant degradation in image quality. Therefore, please download as needed.