LyCORIS Loader

If you've ever downloaded a LoHa or LoKr file, dropped it into the stock LoraLoader, and watched it do nothing - or worse, quietly distort your output - this node is aimed straight at you. LyCORIS Loader is the flagship of AbstractEyes/comfyui-lycoris, a small reimplementation of KohakuBlueleaf's LyCORIS loader built around the author's claim that a lot of those modules "don't correctly load into the ComfyUI structure." Whether that's true depends on your file, but this node is the community's workaround for exactly the failure mode people report: training a LoKr in ai-toolkit or kohya, loading it in ComfyUI, and getting a bypassed-looking LoRA that doesn't change the image.

The pack is a single-commit affair - one author (AbstractPhil), Apache 2.0, unchanged since mid-2025, and "tested primarily with SDXL." Set your expectations there: it's a promising tool, not a battle-tested one, and the README says so more honestly than most.

How it works

The README's whole pitch is precomputation. ComfyUI's sampler is additive, so this node does the math up front: it reads the .safetensors from your normal models/loras folder, casts it to your chosen dtype, pulls the metadata, and parses ss_network_args to detect which algorithm it's dealing with. Then it converts the file through ComfyUI's own comfy.lora_convert, hands it to per-algorithm adapters (LoHA, LoKr, LoCon, GLoRA, OFT, BOFT, BLoRA, plain LoRA), and each one pre-calculates the weight diffs and patches a clone of your model and CLIP. You get the same slot in your workflow as any LoRA loader - checkpoint in, sampler out - with the reconstruction work done at load time instead of mid-sampling.

The inputs that matter

Only a handful, and three of them are the standard LoRA drill:

• lora_name - a dropdown of everything in your loras folder. LyCORIS files are just .safetensors; no separate folder to set up.
• strength_model / strength_clip - both default to 1.0, range -100 to 100. Same semantics as the stock node, including the quirk that setting both to 0 returns the inputs untouched.
• lycoris_type (optional) - any by default, which tries to detect the algorithm from the file's metadata. If your file was re-saved and lost that metadata, force LoHA, LoKr, LoCon, or LyCORIS here instead of trusting auto-detect.
• dtype (optional) - float16 default; bfloat16 and float32 exist but float16 is the right call on most GPUs.

Outputs are model and clip, wired exactly like a LoRA loader feeds your KSampler.

Installing it

No package yet, so it's a manual clone:

cd ComfyUI/custom_nodes
git clone https://github.com/AbstractEyes/comfyui-lycoris

Restart ComfyUI and you're done. requirements.txt is literally just torch - everything else is ComfyUI built-ins, and there are no model downloads. One caution: ComfyUI Manager will surface several packs when you search "lycoris," including the original KohakuBlueleaf one. Check the repo URL before you install; the names are near-identical.

Where people get burned

• Plain LoRAs through this node can malform. The README says it flat out: the default LoRA fallback "may malfunction or malform your lora during precomputation." If you're loading a standard LoRA, use the stock LoraLoader. This node is for LyCORIS-family files.
• Not every variant works. LoHA and LoCon are the tested ones; LoKr is "semi tested." OFT literally can't precompute (its adapter needs runtime weight access and returns nothing), and BOFT and blora are untested. Don't blame yourself if an OFT file silently does nothing.
• any is only as good as the metadata. Detection reads ss_network_args out of the safetensors; files stripped of it need a manual lycoris_type.
• SDXL-first. If you're chasing the reported Flux2 + LoKr loading failures, this pack is thematically on target but hasn't been validated there - keep an eye on the repo before trusting it on a 2025+ transformer base.

One light aside: when auto-detect works, you genuinely can't tell this node from a LoRA loader, which is the whole point - the complexity is supposed to be invisible. Try it on a LoHa file you know works, then on one that's been misbehaving, and keep the stock node for everything else.