ComfyUI-LoRA-Merge-EXPERIMENTAL

ComfyUI-LoRA-Merge-EXPERIMENTAL

日本語README

Custom nodes for ComfyUI that support NP-LoRA and SSR-Merge based LoRA merging.

This extension currently provides:

• NP-LoRA Loader (Subject + Style)
  • Fuses a subject/character LoRA with a style LoRA using NP-LoRA's asymmetric null-space projection.
• SSR-Merge Calibration
  • Runs a lightweight calibration pass for SSR-Merge and builds a merged LoRA from internal activation statistics.
• SSR-Merge Loader
  • Applies the result emitted by SSR-Merge Calibration to a MODEL.

Internally, the extension uses ComfyUI's current custom-node API.

Installation

Place this directory under ComfyUI's custom_nodes directory. From the current development location:

Copy-Item -Recurse D:\tools\dev\ComfyUI-LoRA-Merge-EXPERIMENTAL D:\tools\ComfyUI\ComfyUI\custom_nodes\

Restart ComfyUI, then add the nodes from loaders/LoRA.

NP-LoRA Loader

The NP-LoRA loader follows Eq. 12 of the NP-LoRA paper. It attenuates the content LoRA component that overlaps with the right-singular-vector subspace of the style LoRA.

D_merged = D_style + D_content (I - mu/(1+mu) V V^T)

Basic usage:

1. Set content_lora to the subject, person, or character LoRA you want to preserve.
2. Set style_lora to the style LoRA.
3. Start with mu = 0.5.
4. Increase mu if the style is too weak. Decrease it if subject details are lost. mu = 0 is equivalent to a direct additive merge.

SSR-Merge

SSR-Merge performs a short calibration inference before merging. It collects internal activation statistics, solves an analytic router, and absorbs that router into the up-projection so the result can be applied as a normal LoRA.

This extension exposes SSR-Merge as two nodes:

1. SSR-Merge Calibration

   • model / clip
   • lora_1 / lora_2
   • prompt_1 / prompt_2
   • negative_prompt
   • seed, resolution, lambda_reg
   • sampler / scheduler

2. SSR-Merge Loader

   • Takes the ssr merge output from SSR-Merge Calibration and applies it to MODEL.

Calibration behavior

SSR-Merge Calibration runs a fresh calibration pass when the node is executed.

The calibration inference uses fixed lightweight values for steps and CFG:

• steps: 1
• cfg: 1.0

The sampler and scheduler are user-selectable in the node UI.

SSR-Merge logs normal progress at INFO level. The logs include:

• extracted LoRA layer counts
• routed and passthrough layer counts
• calibration start and end for each LoRA
• registered hook count
• activation statistics coverage
• router solve summary
• Loader timing

Current scope

• Model-side U-Net/DiT LoRA
• Linear and 1x1 convolution LoRA
• SSR merge for two LoRAs

CLIP-side LoRA, spatial LoCon kernels, DoRA, LoHa, and LoKr are intentionally excluded from SSR-Merge for safety.

References

• NP-LoRA: Null Space Projection for Subject-Style LoRA Fusion
• SSR-Merge: Subspace Signal Routing for Training-Free LoRA Merging in Diffusion Models
• SSR-Merge official implementation