Lora Extract K Jv2 (QQ)

LoraExtractKJv2 is the "I trained a full model but I only want the diff" tool. It takes a fine-tuned model and the original it was fine-tuned from, subtracts the delta, and compresses it into a LoRA file you can use anywhere - at a fraction of the storage and without carrying the whole checkpoint around. It's the author's VRAM-optimized v2 of the extraction node in KJNodes (the KJNodes/lora category is a tell), and the generic_docs folder in the repo has a full write-up of what changed.

How it works

The node gets two model inputs: finetuned_model and original_model. For each weight, it computes the difference between them, then factorizes that delta with SVD into low-rank up @ down matrices - that's the "LoRA" form. The core dials:

• rank - target rank of the extracted LoRA (also acts as a cap for the adaptive methods). Higher rank = more fidelity, bigger file.
• lora_type - standard and full are the simple ones; adaptive_ratio, adaptive_quantile, adaptive_energy, and adaptive_fro pick rank per layer based on singular-value statistics instead of a flat cap. For most people, standard at rank 8–32 is the starting point.
• algorithm - svd_linalg (accurate, slower) vs svd_lowrank (faster, approximate).
• adaptive_param - the knob that tunes whichever adaptive method you picked (singular-value ratio, quantile, or Frobenius retention).
• bias_diff - whether to extract bias deltas too (default on).
• output_dtype - fp16 default, or bf16/fp32.

The real headline here is the v2 VRAM machinery, which exists because video checkpoints are enormous and this used to be an out-of-memory generator. vram_usage_mode (conservative ≈50% VRAM, balanced ≈75%, aggressive ≈90%+) is the master dial; below it sit blocks_in_gpu (how many transformer blocks stay resident), batch_size (layers processed at once), prefetch_ahead, memory_threshold (when to auto-offload), and debug_mode for performance telemetry. The author's docs claim 30–120% faster extraction than v1 depending on mode. If you're on a 12GB card, stay on balanced; if you're on a 4090+ and impatient, aggressive is usually safe.

There are no outputs - this is an output node that saves the LoRA using filename_prefix (default loras/ComfyUI_v2_extracted_lora) into ComfyUI/models/loras.

When to use it

The honest use case: you trained or received a full fine-tuned checkpoint and want a shareable LoRA version of just the change, or you're curious what a fine-tune actually altered and want to inspect/export the delta. It's also how people salvage knowledge out of big merged checkpoints. It is not a replacement for proper training - the SVD-compressed delta is a lossy approximation of a fine-tune, and identity/style fidelity degrades as you drop rank.

Installing it

Ships in siraxe/ComfyUI-WanVideoWrapper_QQ (install via Manager by searching WanVideoWrapper_QQ/SA-Nodes-QQ, or git clone https://github.com/siraxe/ComfyUI-WanVideoWrapper_QQ.git into custom_nodes and restart). No special deps beyond standard ComfyUI - the VRAM tricks are pure PyTorch. Load your two checkpoints with the usual model loaders and wire them in.

Gotchas

• VRAM first, speed second. If extraction OOMs, drop vram_usage_mode to conservative before touching anything else - the whole point of this version is that it degrades gracefully.
• Save file missing? Check the console output; debug mode prints exactly where it's writing.
• Standard pack caveat: the repo renamed to ComfyUI-SA-Nodes-QQ in v1.3.4, so delete any stale wanwrapper_qq folder in custom_nodes if old workflows reference it.