Run TryOff Inference [ComfyUI-Flux-TryOff]

Try-on is everywhere. Try-off - removing the clothes from a person in a photo - is the niche this pack exists for, and it's a genuinely unusual capability to have running locally. The author's own words: after "heavy experimenting with Try-on," it's nice to have a try-off model to work with. This node runs that model.

How it works

Same side-by-side trick as the try-on runner, with one crucial difference: there's no garment_in. The node takes your photo of a clothed person plus a mask over the clothing, and zeroes out the garment half of the canvas. The pipeline is asked to fill that empty half and handle the masked clothing region on the person.

Concretely, in the code, try-off mode masks the person's clothing region, then generates both halves of the pair in one pass: the left half becomes a synthesized garment image (the model's best guess at what was worn), and the right half is the person with that clothing removed. The two outputs are those two crops:

• garment_image - the model's reconstruction of the garment.
• tryoff_image - the person without the clothes. The output that matters.

That means the model isn't just inpainting clothes away from the body - it's simultaneously producing a plausible "product shot" of what it just removed, which is a nice piece of consistency engineering and also handy if you want to know what it thinks it saw.

Inputs

• image_in - the clothed person.
• mask_in - mask over the clothing you want removed. The sample workflows auto-generate it with LayerStyle's SegformerB2ClothesUltra; a hand-drawn mask works too, just make it tight to the clothing.
• pipe - the MODEL from the FluxFill pipeline loaders (use the auto-downloading FluxFill Pipeline Loader).
• width / height (default 576×768) - per-half resolution; the real canvas is 2× wide.
• num_steps (default 50) and guidance_scale (default 30) - the quantized sample workflow drops steps to 20. Keep guidance up around 30; this fine-tune doesn't behave like plain Flux dev's 3.5.
• seed and prompt - seed to make runs reproducible; the default two-image prompt structure ("[IMAGE1] product shot, [IMAGE2] worn by a model") is fine to keep as-is, since the model mostly needs the framing, not a fashion lecture.

Expectation management

This is not fast, and it's the first thing people bounce off. The author tested on a 4080 and calls it "quite slow" - a 4090 is the stated floor for "performant." One real-world report described "minutes" per run even on a 4090. If it OOMs, add the TryOff Quantizer at 8-bit; that's the documented way to make it fit on smaller cards. Quality caveats: the tiny TAESD1 VAE this pack uses for speed means previews decode soft, and results on complex clothing (layers, translucent fabric) are shakier than on simple garments - that's the model, not your setup.

Install & the usual gotcha

ComfyUI Manager, search "ComfyUI-Flux-TryOff", or git clone https://github.com/asutermo/ComfyUI-Flux-TryOff into custom_nodes, restart. Everything downloads on first run except the mask generator. The classic failure: the example workflows' SegFormer node errors on segformer_b2_clothes because the model file isn't there:

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

Install LayerStyle via Manager for the node itself. A final note on content: there's no NSFW filter in this pack, so try-off on anything beyond benign fashion shots is a decisions-you-own situation, not a technical one.