<h1 id="extract-anima-checkpoint-diff-lora">Extract Anima Checkpoint Diff LoRA</h1>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.<h3 id="what-it-does">What it does</h3><ul><li>Loads Checkpoint A and Checkpoint B</li><li>Compares only matching UNet weight tensors</li><li>Computes the delta: B - A</li><li>Applies SVD low-rank factorization</li><li>Saves the result as a LoRA-style safetensors</li><li>Optionally saves a JSON report</li><li>Includes a ComfyUI progress bar</li></ul><h3 id="supported-targets">Supported targets</h3><ul><li>Linear layers</li><li>1x1 conv layers</li><li>Optional 3x3 conv layers</li></ul><h3 id="useful-for">Useful for</h3><ul><li>Preserving only the “difference” between two related checkpoints</li><li>Studying how one merge differs from another</li><li>Creating lightweight experimental delta modules</li><li>Comparing nearby model variants without saving a full checkpoint</li></ul><h3 id="main-options">Main options</h3><ul><li>rank: controls compression strength / capacity</li><li>alpha: saved as LoRA alpha metadata</li><li>min_diff_norm: skips layers whose difference is too small</li><li>include_conv3x3: whether to include 3x3 convolution weights</li><li>save_report_json: exports detailed layer-by-layer report</li><li>save_to_models_loras: saves directly into the LoRA folder</li></ul><h3 id="notes">Notes</h3><ul><li>This node targets Anima UNet weights only</li><li>Best results are expected when using closely related checkpoints</li><li>It is intended as a delta extraction / experiment tool, not a perfect full-model reconstruction method</li><li>The saved output uses LoRA-style key naming for compatibility-oriented workflows</li></ul><h3 id="why-use-this">Why use this?</h3>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

Files with SHA256: a8a2c198b061a559adc90e5c6d19d5a645ee179283160a5bd4dde61935c89dd7