DoRA Extract (Fixed Rank)

Full fine-tuning genuinely produces better results than LoRA training, and the community has never really disputed that - the reason people train LoRAs anyway is that fine-tuning a whole checkpoint is expensive and a LoRA is cheap to store and share. This node is the middle path taken after the fact instead of during training: hand it a base checkpoint and a fine-tune of it, and it computes the difference and compresses that difference into a low-rank adapter, at a rank you specify directly. It's the classic "extract a LoRA from a diff" operation this pack's README credits to kohya-ss/sd-scripts' lineage, applied here to produce a DoRA-format adapter specifically rather than a plain LoRA.

How it works

model_a and model_b are your two inputs, and their tooltips spell out the direction unambiguously: model_a is "Finetuned model (A - B = LoRA)", model_b is "Base model (A - B = LoRA)." Get them backwards and you're extracting the negative of what you meant, which the math handles but which will confuse you when you go to apply it.

linear_dim (default 64) and conv_dim (default 32) are the fixed ranks - set separately because attention/linear layers and convolutional layers behave differently under SVD compression, and the pack's own README lists "Fixed" rank as one of five rank-selection strategies it ships (alongside Ratio, Quantile, Knee-detection, and Frobenius-norm - see DoRAExtractFrobenius in this same pack for the adaptive alternative). Higher rank keeps more of the original difference's fidelity at the cost of a bigger output file; lower rank is more compact and lossier. svd_niter (default 2) controls SVD power iterations - the tooltip is direct: higher is "more accurate but slower."

A handful of settings tune the extraction quality and memory footprint: clamp_quantile (default 0.99) clips outlier singular values so a few extreme weights don't dominate the compressed result; min_diff (default 0) skips layers where the base and fine-tune barely differ, which is a legitimate speed and size win - those layers weren't meaningfully touched by the fine-tune anyway. chunk_large_layers (off by default) splits big fused layers (QKV, MLP blocks) into chunks, useful on tighter VRAM budgets. lazy_load and force_clear_cache (both default on) manage memory the same way they do throughout this pack. mismatch_mode (default skip) handles tensors that exist in one model but not the other. skip_patterns (regex, or glob with glob_skip_patterns) lets you exclude specific layers from extraction entirely. save_dtype (default fp16) and device (default cuda) round it out. Output is a single output_path string (output_filename default extracted_lora), and it's a save node - this writes the file when it runs.

Installing it

ComfyUI Manager: search Model Utility Toolkit, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/silveroxides/ComfyUI-ModelUtils

Restart ComfyUI. No downloads needed beyond the two model files you're extracting from.

Where people get burned, and the honest take on DoRA

Get model_a/model_b swapped and the extraction still runs - SVD doesn't care which direction you fed it - but the resulting adapter applies the inverse of the change you wanted. If applying the extracted DoRA makes your output look less like your fine-tune, not more, that's the first thing to check.

The bigger thing to know before you commit time to this: DoRA had a real moment of enthusiasm - "LoRAs and fine tunes are dead to me, all hail DoRA" was a genuine sentiment in 2024 - and it never converted into mainstream adoption. The community's honest, repeated experience is that DoRA adapters don't stack well with each other, and that likeness reproduction is dataset-dependent rather than reliably better than a plain LoRA extracted the same way. If you're extracting this to combine with other adapters later, know going in that stacking is the documented weak point, not a hypothetical one.