Extract Lycoris Lora

This is the "one more checkbox" version of LoRA extraction. Extract Lycoris Lora does exactly what Extract Kohya Lora does - compress a merge delta down to lora_up/lora_down low-rank matrices via SVD - but it emits a LyCORIS-layout LoRA instead of a kohya one. Same rank, same approximate-SVD knobs, different key naming (lora_unet_* / lora_te_* prefixes, LyCORIS conventions). If you've never heard of LyCORIS, don't worry: it's the training library that grew out of LoRA and added a bunch of network types (LoHa, LoKr, Diag-OFT…) on top. Some loaders and some base-model ecosystems expect its naming, and this node exists so your extracted adapter lands on the right side of that fence.

Inputs

• base (delta) - the difference to compress. From a Subtract between two checkpoints, or any delta recipe.
• rank - default 8. Singular values kept. 8–32 is the practical SDXL range; higher rank = more fidelity, bigger file.
• use_approximate_basis (True), approximate_basis_iters (2), approximate_basis_seed (None) - the randomized-SVD tuning. Approximate is way faster and near-invisible below a few iters; seed it if you want reproducible extractions.

Output: one recipe (MECHA_RECIPE) in LyCORIS LoRA space. Same usage as the kohya variant: into Add Difference as the b (delta) slot on a compatible base, or out through the serializer. It's a delta - don't hand it to Mecha Merger expecting a full model.

The merge_checkpointing toggle caches the branch on CPU in fp16 for repeated runs.

Kohya vs LyCORIS: does it actually matter?

Honestly, for SDXL, often no - both layouts load fine in ComfyUI's standard LoRA loader and most trainers. It matters in three situations: your downstream tooling (an external trainer, a script, another pack) explicitly checks the prefix; the base model's docs recommend one over the other; or you're matching the format the rest of your collection uses. The deeper point is the one the LoRA knowledge doc in our KB hammers: "LoRA" the name has outlived the format, and what people call a LoRA on modern bases is increasingly a different decomposition entirely. If you're extracting for a current architecture, check what your base's adapter ecosystem actually expects before picking a layout - LyCORIS-style tends to be the safer bet there.

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, no model downloads.

Feed it a full weight where it wants a delta and you'll get nonsense compression, not an error. Extract from the full checkpoint difference, not from an already-extracted adapter, or you stack the loss twice.