1.4 KiB
FL Model Difference to LoRA
Connect two Load Diffusion Model nodes: the fine-tune to finetuned_model, and its original base to base_model. Queue FL Model Difference to LoRA to extract fine-tune - base and save a numbered .safetensors file under the first configured LoRA directory. The filename_prefix can include a subfolder, such as Krea2/Dirtyrealism_difference.
Rank defaults to 128. Higher ranks preserve more of the weight difference and produce larger files. auto uses ComfyUI's selected device; choose cpu if the device lacks SVD support or available memory. Extraction processes one parameter at a time, honors weight patches and quantization scales, and leaves the input models unchanged. Biases and other non-matrix parameters are stored as full differences in ComfyUI's LoRA format. No CLIP extraction is performed.
Apply the exported file to the same base with Load LoRA (Model Only) at strength 1. Low-rank compression is approximate, and quantized source models include quantization differences. Logged retained weight energy measures numerical reconstruction, not image quality. Compare generations before replacing a fine-tune in production workflows.
The included Krea2 workflow selects krea2SATDirtyrealism_v5_fp8 and krea2_turbo_fp8_scaled, saving into loras/Krea2. Restart ComfyUI after installing the node, then open the workflow.