Run TryOn or TryOff Inference [ComfyUI-Flux-TryOff]

The pack's other two run nodes - Run TryOn Inference and Run TryOff Inference - each do one job. This one does both, decided by a single input: whether you feed it a garment. Give it a garment and it puts clothes on; leave that input empty and it takes them off. It's the closest thing this pack has to a single "do the thing" button.

How the switching works

The behavior hinges on garment_in, the one input the dedicated run nodes don't have. From the source: if garment_in is provided (not None), it calls the try-on path - concat the garment image next to the masked person, fill the double-width canvas, crop the pair apart. If it's left empty, it calls the try-off path - no garment, so the model fills that half with a synthesized garment and regenerates the masked clothing region as the person without it. Same side-by-side CatVTON mechanism, one switch.

Everything else is the standard run-node surface you've seen in the other two:

• image_in - the person. mask_in - mask over the current clothing.
• pipe - MODEL from the FluxFill pipeline loaders.
• width / height (default 576×768), num_steps (default 50), guidance_scale (default 30), seed, prompt (the two-image "[IMAGE1]/[IMAGE2]" structure).

The naming trap

Here's where it bites. The outputs are hardcoded to garment_image and tryon_image - always. Run it in try-off mode and the second output is still labeled "tryon_image," but it's actually your try-off result (the person without clothes). The code returns it in that slot regardless of mode, and the brief/UI just keep the same label. It's cosmetic - the tensor is correct, only the name lies - but if you're building a workflow that routes outputs by name, know that "tryon_image" from this node is only trustworthy when you fed it a garment. The dedicated TryOff Run Node exists precisely to avoid this confusion, with its honestly-named tryoff_image output.

When to use it vs. the dedicated nodes

If you're experimenting - "what does this model do to this photo both ways?" - this node is convenient: one runner, just add or drop a garment. But for a workflow you'll share or reuse, the dedicated nodes are clearer and fail less confusingly. This is also the node that makes it trivial to toggle try-on/try-off with a simple conditional wiring your choice of garment image, if you're into that kind of graph.

Install & troubleshooting

Pack install is the usual: ComfyUI Manager → search "ComfyUI-Flux-TryOff", or git clone https://github.com/asutermo/ComfyUI-Flux-TryOff into custom_nodes, restart. Models auto-download on first run. The mask comes from LayerStyle's SegFormer in the sample workflows - install LayerStyle via Manager and clone the SegFormer model or you'll hit the segformer_b2_clothes error:

cd ComfyUI/models
git clone https://huggingface.co/mattmdjaga/segformer_b2_clothes

And the standing performance disclaimer: the author tested on a 4080 and calls it "quite slow" - a 4090 or better is the real recommendation, and quantizing via the TryOff Quantizer (8-bit is the documented default) is how you keep it running on smaller cards. Try-on and try-off both inherit all of that.