SDXL-KOHYA_BUT_DIFFUSERS_KOHYA_LORA Components Params

Here's the setup that shows up constantly on CivitAI: an SDXL model with kohya-style text encoders and VAE but a diffusers-named UNet, plus LoRAs trained the ordinary kohya way. Three key dialects in one recipe. The SDXL-KOHYA_BUT_DIFFUSERS_KOHYA_LORA Components Params node from comfy-mecha is the per-component weight node built for exactly that combination.

The name is a spec sheet: SDXL-KOHYA = kohya-layout SDXL checkpoint, BUT_DIFFUSERS = except the UNet keys use diffusers naming, KOHYA_LORA = the LoRAs are standard kohya-trained. It's a param node, so it produces a MECHA_RECIPE of per-component weights rather than a merged model. Its inputs:

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

Wire the output into a merge method's (delta|param|weight) input - for applying LoRA deltas to a checkpoint that's Add Difference - and zero out anything you don't want touched.

How it works

Same family machinery: each float becomes a literal per-key weight restricted to its component, combined into one recipe. The value is that the config knows all three key dialects, so you don't have to reconcile checkpoint keys and LoRA keys yourself. After building the recipe, run it through Mecha Merger to land a usable MODEL/CLIP/VAE.

Install

Pack-level: 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. Only dependency is sd-mecha; no model downloads.

Gotchas

Every clause has to match your actual files, or the node quietly becomes the wrong node. UNet keys diffusers-style? Yes. LoRAs kohya-trained (not LyCORIS)? Yes. If either answer flips, the sibling variants - SDXL-KOHYA_BUT_DIFFUSERS_LYCORIS_LORA for LyCORIS LoRAs, plain SDXL-KOHYA_KOHYA_LORA for a fully kohya checkpoint - are the right pick. The pack naming is verbose, but it's doing the documentation work for you.