TE DoRA Extract (Ratio)

This is the ratio-flavored sibling of the TE DoRA extractors in ModelUtils: same DoRA trick, rank decided per layer by a ratio instead of a fixed number or a percentile. Keep every singular value above max(S) / ratio, decompose the direction, stash the magnitude in a dora_scale. If you already understand the LoRA Ratio node, this is that, with the DoRA difference step swapped in.

Quick refresher on what changes with DoRA. Where LoRA extraction just computes A - B and decomposes the whole delta, the DoRA path rescales each row of the fine-tuned encoder to the base's row magnitude before subtracting. That separates the update into direction (what gets low-rank-decomposed) and magnitude (what gets saved as the per-layer dora_scale). So the output file has lora_A, lora_B, and dora_scale keys - the DoRA signature. Whether that's worth the extra moving parts is up to you; the community's DoRA enthusiasm cooled mostly because trained DoRAs don't stack like LoRAs do, but for distilling one specific delta the magnitude/direction split is a legitimate choice, not a gimmick.

The inputs that matter

• model_a / model_b - fine-tuned and base text encoders from models/text_encoders.
• linear_ratio / conv_ratio - the rank threshold for linear/attention vs convolution layers, default 2. Higher values keep more singular values (a higher ratio lowers the cutoff, max(S) / ratio), so raise it for higher fidelity, lower it for a tighter, smaller file. Same kohya-style convention as the LoRA Ratio node.
• linear_max_rank / conv_max_rank - hard caps (default 128) so no single noisy layer blows the rank up.
• probe_offset - extra singular values sampled past the cap so the threshold decision sees a full-enough spectrum.
• output_filename / save_dtype / device / lazy_load / force_clear_cache - the shared knobs: output name in your loras folder, save precision (fp16 default), compute device with per-layer CPU fallback on OOM, streaming memory.

It's an output node; output_path tells you where the file landed.

Install

ComfyUI Manager → "Model Utility Toolkit", or:

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

Restart; it sits under ModelUtils/DoRA Extract (TE). The streaming dependency that actually matters is unifiedefficientloader.

Gotchas

Same family warnings: mismatch_mode: skip is the default, so a wrong base means a silently small, useless file. DoRA extraction runs the full SVD per layer, so budget minutes, not seconds. And if the resulting file loads but does nothing, suspect the loader before the extraction - dora_scale support is not universal across LoraLoaders. When in doubt, compare against the plain LoRA Ratio output: if it behaves better in your stack, that's your answer about whether DoRA was worth it for your use case.