ComfyUI Node
Mpi LoRA Extract
Extracts a LoRA from the difference between a finetuned and an original model (or CLIP) and saves it under output/.
Mpi LoRA Extract
- finetuned
- original
◄filename_prefixloras/ComfyUI_extracted_lora►
◄rank64►
◄lora_type▾►
◄algorithmsvd_lowrank►
◄lowrank_iters7►
◄output_dtypefp16►
◄bias_difftrue►
◄adaptive_param0.15►
◄clamp_quantilefalse►
CategoryMpiNodes/KJ
Inputs (11)
| Name | Type | Default | Description |
|---|---|---|---|
| finetuned | MODEL,CLIP | The finetuned model or clip to extract LoRA from. | |
| original | MODEL,CLIP | The original base model or clip to diff against. | |
| filename_prefix | STRING | loras/ComfyUI_extracted_lora | — |
| rank | INT | 641–4096 | The rank for a standard LoRA, or the maximum rank for adaptive methods. |
| lora_type | COMBO | 6 options: standard, full, adaptive_ratio, adaptive_quantile, adaptive_energy, adaptive_fro | |
| algorithm | COMBO | svd_lowrank | SVD algorithm; svd_lowrank is faster but less accurate. |
| lowrank_iters | INT | 71–100 | Subspace iterations for the lowrank SVD algorithm. |
| output_dtype | COMBO | fp16 | 3 options: fp16, bf16, fp32 |
| bias_diff | BOOLEAN | true | — |
| adaptive_param | FLOAT | 0.150–1 | ratio mode: fraction of the largest singular value. quantile mode: the quantile of the singular values. fro mode: Frobenius norm retention. |
| clamp_quantile | BOOLEAN | false | — |
Outputs (0)
No outputs