Extract Kohya Lokr

LoKR is the less famous sibling of LoRA, and it deserves the spotlight here. Where a plain LoRA decomposes a weight delta into two skinny matrices (up @ down), LoKR - Kronecker-based LoRA - factors the delta into a Kronecker product of two full-rank-ish smaller matrices. Net effect: a LoKr can represent a given change with less rank loss than a LoRA of the same size, which is why the community has been gravitating to it for character work on newer bases. This node extracts a kohya-format LoKr from any merge delta, same workflow as Extract Kohya Lora, different decomposition.

How it works

Extract Kohya Lokr takes a base (delta) recipe and runs a truncated Kronecker decomposition, controlled by one key knob:

• kronecker_ratio - default 0.5. This decides how the matrix is split into the two Kronecker factors. 0.5 splits both dimensions roughly in half. Values toward 0 make the first factor tiny (more compression, more loss); toward 1 the opposite. 0.5 is the sane starting point, and honestly the place most people leave it.

The decomposition is rank-1 per factor pair, so the file stays small while the reconstruction quality is surprisingly good - that's the whole LoKr pitch. Output is a single recipe (MECHA_RECIPE) in LoKr weight space. Use it like any extracted adapter: wire it into Add Difference as the delta input on a compatible base, or serialize it. And same as its LoRA sibling, don't expect it to work standalone - it's a delta.

Unlike Extract Kohya Lora there's no rank / approximate-basis tuning here; the ratio is the only real parameter, which keeps the node mercifully simple.

Why pick LoKr over LoRA?

If you're targeting SDXL-era checkpoints and want max fidelity per byte, LoKr usually wins on the same budget. The catch is ecosystem support: not every tool and every base's LoRA stack treats LoKR files uniformly. On current architectures the convention has genuinely shifted - what people casually call "a LoRA" on post-Flux bases is increasingly a LoKr under the hood, as the LoRA knowledge doc in our KB notes. So extracting LoKr isn't a niche flex anymore; it's often the more future-proof choice. If you need maximum compatibility with older loaders, extract the plain LoRA instead.

Install & gotchas

ComfyUI Manager → search mecha → "Mecha Merge Node Pack", or:

cd ComfyUI/custom_nodes
git clone https://github.com/ljleb/comfy-mecha.git
pip install -r comfy-mecha/requirements.txt

Restart after. Dependency is sd-mecha==1.1.7.

Two things that bite: the delta you feed it must be a weight delta (two models subtracted), not a recipe already in LoRA space - decompose once. And kohya-format LoKr keys assume the kohya naming; if your downstream expects the LyCORIS layout, grab Extract Lycoris Lokr instead. They're the same math with different labels.