Flux Virtual Try-On

Virtual try-on is one of the classic "the API is just better at this" jobs. Locally, dressing a person in a specific garment means garment-conditioned diffusion models, mask compositing, often a dedicated ID model, and usually a pile of fiddling to keep the person's face and pose intact. Flux Virtual Try-On (Flux VTO) reduces all of that to two image inputs: a photo of the person and a photo of the garment. BFL's dedicated VTO model does the rest, and the node pipes it back as a finished image. It shipped in core in June 2026 alongside Flux Erase.

Same shared mechanics as every BFL partner node: both images go out as base64 to Comfy's proxy at api.comfy.org, BFL's servers run the model, and you're billed from the credits on your Comfy account. Nothing downloads into models/, nothing uses your GPU. The node even strips alpha from your inputs before sending - the person and garment should be plain RGB photos, not cutouts with transparent backgrounds, though a clean shot of the garment on a flat background is exactly what the model likes.

Inputs are refreshingly minimal: person (the model to dress), garment (the clothing to apply), prompt (optional natural-language styling, e.g. "fitted," "untucked," "make the sleeves rolled"), and seed. That prompt input is a nice touch - the model is instruction-capable about fit rather than just doing a literal paste - but it can be empty and the node still runs.

Output is a single IMAGE tensor, already decoded pixels. Save it, send it to an upscaler, composite it back into a product listing - it slots straight into your graph with no VAE step. The result preserves the person (face, pose, lighting) and renders the garment on them with proper drape, which is the thing general image-editors still get wrong. That's the category note worth repeating: try-on is a specialist task, and specialist models in the cloud tend to beat generalist local ones at it, even when the local ones are free.

The practical caveats are the pricing ones. The badge shows a range because BFL prices by something like output resolution or the sizes of your inputs - so the number you see on the node before running it is the honest "about this much." And the usual partner-node rules apply: no bring-your-own-key (Comfy account credits or nothing), results are best-effort with a seed rather than deterministic, and the request can bounce off moderation if either input trips the filter.

Where does it fit? If you're doing e-commerce shots, fashion mockups, or character work where the garment has to be recognizable rather than merely adjacent, this is the node you'll reach for over a generic Kontext-style edit. For one-off experiments, a local model is fine. For "I need fifty garment-on-model renders and my GPU isn't a server," you've just found your workhorse.