TE LoRA Extract (Quantile)

Here's the friendliest way to think about a LoRA rank: "how much of the signal do I keep?" This node turns that sentence into a number. You extract a LoRA from a fine-tuned text encoder minus its base, and instead of picking a rank you tell it to keep enough singular values to reach a target percentage of the total - 90% by default. It's the quantile variant of the ModelUtils TE extractors, and it's the one that reads most naturally if you're new to this.

The mechanism is the pack's standard extraction pipeline: for each .weight tensor, compute A - B, run SVD, and keep singular values until their cumulative sum crosses the target fraction. Layers that are mostly noise stop early and stay cheap; layers with real structure get the rank they earn. The result is written as PEFT-style lora_A/lora_B factors into your loras folder, ready for a normal LoraLoader. If the target fraction demands a rank as big as the layer itself, the node just saves the full difference instead of pretending a decomposition helps.

The inputs that matter

• model_a / model_b - the fine-tuned and base text encoders from models/text_encoders.
• linear_quantile / conv_quantile - the cumulative singular-value target for linear/attention vs convolution layers (default 0.9). 0.9 keeps enough rank to cover 90% of the spectrum's mass; bump toward 1.0 for higher fidelity at a bigger file, drop it for compression.
• linear_max_rank / conv_max_rank - safety caps (default 128) so one outlier layer can't balloon the rank.
• probe_offset - extra singular values sampled beyond the cap so the "have we hit 90% yet?" decision is made against a spectrum that isn't truncated mid-answer.
• output_filename / save_dtype / device / lazy_load / force_clear_cache - the usual shared knobs: output name, save precision (fp16 default), compute device (with per-layer CPU fallback on OOM), and streaming vs. eager memory behavior.

Like the other extractors it's an output node: it writes the file and reports output_path. Nothing hangs off it.

Install

Search "Model Utility Toolkit" in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/silveroxides/ComfyUI-ModelUtils
cd ComfyUI-ModelUtils
pip install -r requirements.txt

Restart, and it lives under ModelUtils/LoRA Extract (TE). The dependency doing the heavy lifting is unifiedefficientloader, which lets the node stream both encoders instead of loading them fully.

Watch out

The silent-skip trap is real here too: default mismatch_mode: skip means mismatched or missing tensors disappear without a peep, so comparing an encoder to the wrong base quietly produces a nearly-empty file. And this is a full SVD per layer, so expect minutes of progress bar, not instant. If 90% of the spectrum sounds arbitrary, it is - the Knee variant's whole job is removing that guess, and Frobenius lets you set the target in terms of energy rather than singular-value mass. Quantile is the middle ground: explicit, interpretable, and usually right at 0.9.