Load Catvton Flux

This is the entry point for the whole pack: the node that gets you an actual working virtual try-on pipeline running inside ComfyUI. Drop it in, and it hands you a model you can feed into CatvtonFluxSampler to put a garment on a person. The catch - and it's a real one - is what it does to get there.

What it is, and why it exists

The wrapper is built on top of catvton-flux, a virtual try-on method that repurposes FLUX.1 Fill dev - Black Forest Labs' own model, trained specifically for masked inpainting and released November 2024 - for garment transfer instead of generic object removal. Fill dev is still the only widely-used open checkpoint actually trained for mask-conditioned work rather than a regular model with a mask bolted on, which is exactly why someone building a try-on tool reached for it. The pack's own author put it plainly when he shipped this: "the best open-sourced virtual try-on model I've ever seen." That's a real compliment from someone who clearly tried the alternatives.

How it works

LoadCatvtonFlux has zero inputs. No file picker, no checkpoint dropdown - you place the node and it does its thing. Under the hood it wraps a diffusers pipeline that pulls black-forest-labs/FLUX.1-Fill-dev straight from Hugging Face the first time you run it, at full bf16 precision, no fp8 or GGUF quantization. That's the whole reason the README spends more words on Hugging Face authentication than on ComfyUI itself.

The single output is CatvtonFluxModel - a custom type that only wires into this pack's own CatvtonFluxSampler. It won't plug into a generic KSampler or anything expecting a standard MODEL.

Installing it

Search ComfyUI-CatvtonFluxWrapper in ComfyUI Manager (update the manager first if it doesn't show up), or by hand:

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

Restart ComfyUI. Then, before you ever queue a run with this node, deal with the gate: Fill dev is a gated Hugging Face repo under BFL's non-commercial license. Go to the model page, accept the license, generate an access token under Settings → Access Tokens, and set HF_TOKEN and HUGGING_FACE_HUB_TOKEN as environment variables before starting ComfyUI. Skip this and you'll hit exactly the error real users have hit: "Cannot access gated repo... Access to model black-forest-labs/FLUX.1-Fill-dev is restricted." Alternatively, activate your ComfyUI venv and run huggingface-cli login once, paste the token, and it downloads fine from then on. And worth knowing up front: the underlying model is non-commercial use only, same as Fill dev itself.

Where people get burned

24GB is not enough, and that surprised a lot of people who assumed Flux fits comfortably on a 3090 or 4090. It usually does - with fp8 or GGUF quantization, which is how most of the Flux ecosystem actually runs it. This node doesn't do any of that; it loads the full bf16 diffusers pipeline, and multiple people in the pack's own launch thread confirmed OOM on 24GB cards. If that's your setup, this loader is the wrong one for you - reach for LoadCatvtonFluxLoRA instead, which works with a local, already-quantized Fill checkpoint plus a LoRA instead of downloading the whole thing fresh.

Where the model actually lands is another common confusion. Diffusers caches to its own default Hugging Face cache directory, not any folder inside ComfyUI's models/ tree - so if you symlink your ComfyUI model directories elsewhere, don't go hunting there for it.

An old bundled accelerate version can throw load_checkpoint_and_dispatch() got an unexpected keyword argument 'strict' on load. The fix that's worked for people: open a terminal in ComfyUI's embedded Python folder and run python -m pip install accelerate -U. Treat it as best-effort - it's not an official fix, but it's gotten people unstuck.