Extract Anima Checkpoint Diff LoRA

This custom ComfyUI node extracts the UNet weight difference between two similar Anima checkpoints and compresses that delta with SVD, then saves it as a compact LoRA-style safetensors file.

It is designed for cases where you want to capture the behavioral difference between two closely related checkpoints without distributing a full merged model.

What it does

• Loads Checkpoint A and Checkpoint B

• Compares only matching UNet weight tensors

• Computes the delta: B - A

• Applies SVD low-rank factorization

• Saves the result as a LoRA-style safetensors

• Optionally saves a JSON report

• Includes a ComfyUI progress bar

Supported targets

• Linear layers

• 1x1 conv layers

• Optional 3x3 conv layers

Useful for

• Preserving only the “difference” between two related checkpoints

• Studying how one merge differs from another

• Creating lightweight experimental delta modules

• Comparing nearby model variants without saving a full checkpoint

Main options

• rank: controls compression strength / capacity

• alpha: saved as LoRA alpha metadata

• min_diff_norm: skips layers whose difference is too small

• include_conv3x3: whether to include 3x3 convolution weights

• save_report_json: exports detailed layer-by-layer report

• save_to_models_loras: saves directly into the LoRA folder

Notes

• This node targets Anima UNet weights only

• Best results are expected when using closely related checkpoints

• It is intended as a delta extraction / experiment tool, not a perfect full-model reconstruction method

• The saved output uses LoRA-style key naming for compatibility-oriented workflows

Why use this?

If you have two checkpoints from the same model family and want a more compact way to preserve only their UNet difference, this node gives you a practical workflow:
checkpoint diff → SVD compression → lightweight LoRA-style file