🔬 Easy LoRA Extractor

You know the situation: someone trained a checkpoint on top of a base you have, and you want the difference - the character, the style, the concept - as a reusable LoRA. Or you trained one yourself and want to hand out a 100 MB adapter instead of a 7 GB file. That's exactly what the Easy LoRA Extractor does: it takes a base (before) and a fine-tuned (after) checkpoint and extracts the delta as a LoRA.

This is the same trick the "LoRA extraction" scripts and web UIs do, but wrapped as a node with the full Easy LoRAMerger treatment - SVD rank selection, noise-aware thresholding, and a forensic report at the end.

How it works

The pipeline is straightforward under the hood:

1. Both checkpoints get normalized into a shared key space.
2. Keys present in both with compatible shapes get matched.
3. The delta is computed: Δ = tuned − base.
4. Each delta matrix is decomposed with SVD into the low-rank lora_A / lora_B pair that is a LoRA.

The clever parts are the knobs around step 4. svd_mode defaults to auto_energy, which keeps the singular values that retain your chosen energy_threshold (default 0.95 - higher rank, more fidelity; lower, more compression). noise_thresholding is on by default and uses Marchenko-Pastur theory to tell real signal from noise - but the tooltip has the gotcha: turn it off for cross-model deltas, like a Turbo checkpoint vs its Base, where the "noise model" can wrongly reject genuine signal.

It also handles the bane of extraction: models that were trained with a baked-in LoRA attenuation. strength_multiplier set to auto detects that attenuation (via detection_mode - auto_fast samples 30 layers, auto_precise checks all of them) and compensates before the SVD runs.

Inputs worth setting

• checkpoint_base - the unedited "before" state.
• checkpoint_tuned - the "after" state whose concept you're extracting.
• rank / rank_mode - leave rank_mode on auto and let the spectrum decide; it's the upper bound when auto.
• lora_format - native (CivitAI-compatible diffusion_model.* keys, merged QKV) or diffusers (HuggingFace-style transformer.* keys, split QKV). Native unless you specifically need the Diffusers one.
• save_trigger - off by default, which previews the extraction applied to an optional connected model / clip instead of writing a file.

Outputs: model and clip (the extraction applied at strength_model / strength_clip, so you can preview before saving), forensics (the extraction report), and lora_path (where the .safetensors landed once save_trigger is on).

Installing it

Same as the rest of the pack - Manager (search "Easy LoRA Merger") or:

cd ComfyUI/custom_nodes
git clone https://github.com/Terpentinas/EasyLoRAMerger

Restart and you're done. No extra model files.

Gotchas

Cross-architecture extraction (Flux base + SDXL tuned) is a non-starter - the normalizer can't reconcile the key spaces, and the report will tell you so. And the #1 "why is my extracted LoRA weak" fix is letting strength_multiplier run in auto; the #2 fix is remembering that if your tuned checkpoint barely differs from base, there's not much signal to extract. Preview with save_trigger off before you commit anything to disk.