SDXL-KOHYA_BUT_DIFFUSERS_LYCORIS_LORA Components Params

Of all the Components Params nodes in comfy-mecha, this one has the longest name and arguably the least common use case - which is why it exists. The SDXL-KOHYA_BUT_DIFFUSERS_LYCORIS_LORA Components Params node handles a merge where the checkpoint is kohya-layout, the UNet keys are diffusers-style, and the LoRAs are LyCORIS networks (LoCon/LoHa). Three dialects, one config.

It's a param node: it emits a MECHA_RECIPE of per-component weights rather than a merged model. Inputs:

• diffuser - UNet (diffusers naming)
• clip_l, clip_g - text encoders (kohya naming)
• vae - autoencoder

Feed the output into a merge method's (delta|param|weight) input - Add Difference is the one for applying LyCORIS LoRA deltas onto a checkpoint - and set anything to 0 to skip it.

How it works

The mechanism is identical across the family: each float becomes a literal per-key weight restricted to its component, then combined into one recipe. The node's whole value is that its config understands all three key dialects, so a LyCORIS LoRA, a diffusers UNet, and kohya text encoders can coexist in one merge without hand-mapping keys.

Install

Standard pack install: 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 ComfyUI. Dependency is just sd-mecha; no model files to fetch.

Gotchas

Let the name be your checklist. If your LoRAs are actually standard kohya LoRAs, step down to SDXL-KOHYA_BUT_DIFFUSERS_KOHYA_LORA; if the UNet is also kohya-named, drop the BUT_DIFFUSERS clause and use SDXL-KOHYA_LYCORIS_LORA. LyCORIS networks are worth getting right because their key structure differs from plain LoRA in ways that a wrong config will silently ignore - one reason people think a merge "did nothing" is that the LoRA clause didn't match the file they loaded.