Load Catvton Flux LoRA

If LoadCatvtonFlux OOM'd on you, this is the node that gets you unstuck. Instead of downloading Black Forest Labs' full FLUX.1 Fill dev checkpoint fresh in bf16, it patches a LoRA onto a Flux Fill model you already have loaded - which means you control the quantization and skip the multi-gigabyte diffusers download entirely.

Why this node exists

Straight from the pack's launch thread: someone asked the author whether a 24GB card was enough. His answer was blunt - the full fine-tuned Fill pipeline wasn't going to fit, so "what I used for now is lora instead of the fully finetuned fill model." That LoRA is xiaozaa/catvton-flux-lora-alpha on Hugging Face, and it's the whole reason a second, lighter loading path exists in this pack at all. The README frames it the same way: use the LoRA workflows "if you'd like more flexibility like using fp8 flux fill model."

This is the standard model-plus-LoRA pattern you'll recognize from basically every other corner of the ecosystem - a small delta file layered onto a base checkpoint instead of shipping the whole thing pre-baked - just applied here to a Fill-dev-based try-on model instead of a style or character LoRA.

How it works

One input, MODEL; one output, MODEL. You feed it whatever loaded your local Flux Fill checkpoint - an fp8 or GGUF quantization loaded through ComfyUI's own diffusion-model loaders - and it applies the catvton LoRA on top, then hands the patched model onward.

Worth flagging honestly: there's no file picker on this node. No LoRA-name dropdown, no path field - that's consistent with this pack's whole loader philosophy (LoadCatvtonFlux doesn't expose a checkpoint chooser either). It's opinionated software: it assumes you want the catvton-flux LoRA and handles the specifics internally rather than making you pick a file. If it's not finding the weights on your machine, the node's own source on GitHub is the place to check where it expects them, since the schema itself won't tell you.

Installing it

Same as the rest of the pack - search ComfyUI-CatvtonFluxWrapper in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/lujiazho/ComfyUI-CatvtonFluxWrapper

Restart ComfyUI. Grab the LoRA from the link in the README, and load your own quantized Flux Fill dev checkpoint the normal way before wiring it into this node - the whole point is that you're supplying the base model rather than letting the wrapper fetch one for you.

Where people get burned

Key mismatches. In the pack's early history, the LoRA's weight keys didn't match a locally-loaded standalone Fill checkpoint - only the diffusers-loaded version worked with it. The author added a key transform specifically to fix this, which is the entire reason this node exists as a separate loader rather than everyone just using LoadCatvtonFlux. If you're on an old copy of the pack and this node behaves oddly, update it.

Wrong base model. The LoRA was trained against Fill dev's extra inpainting conditioning, not vanilla Flux Dev. Feed this node a plain Flux Dev checkpoint and it'll load without complaint but won't do what you want - garment transfer needs the Fill-specific behavior underneath.

Licensing carries through. Both the LoRA and the Fill dev checkpoint it's built on are non-commercial only. That doesn't change just because you're loading a smaller file.