Virtual try-on has a specific failure mode: the garment transfers fine, but the person changes. Different face, different pose, subtly different body — and the result is useless for a product page, because the customer is no longer looking at themselves.
Black Forest Labs' VTO endpoint addresses this directly in the prompt, and their documented formula opens by pinning the person down before the garment is ever mentioned.
Sourcing note: the prompt formula, garment reference guidance, and positioning below are quoted from Black Forest Labs' own FLUX VTO documentation. The demo link is BFL's. No third-party try-on benchmark is cited, because neither BFL nor anyone else publishes a neutral one. Last verified July 30, 2026.
What this guide solves
The pain point: most try-on writeups show you the output, not the prompt. VTO quality is unusually sensitive to how you describe the garment, and the difference between a generic prompt and a specific one is visible immediately.
The differentiator: this is the official formula with BFL's own rules about what not to say — the part that fixes on-model reference images, which is where most catalogue workflows actually break.
The formula
BFL publishes it as the core prompt:
The person of image 1, maintaining exactly their face and pose, wearing the [YOUR GARMENT DESCRIPTIONS] of image 2.
Read the structure, because every clause is load-bearing:
- "The person of image 1" — establishes which input is the subject
- "maintaining exactly their face and pose" — the identity lock, stated before the garment
- "wearing the […] of image 2" — the transfer, with a slot for description
BFL notes the static version works as a default — wearing the garments of image 2 — but that adding a concise garment description meaningfully improves quality, especially for complex garments.
Describe the garment like a product listing
Their guidance is to specify fit and category. Their own examples:
- the oversized tee
- the cap with "Keep FLUXing" text
- the 7/8 length pants
Notice these are the exact attributes a merchandiser would write: cut, length, and any text on the item. "A shirt" gives the model latitude it will use; "the oversized tee" does not.
The rule that fixes on-model references
This is the most valuable line in BFL's documentation and it is easy to skim past.
When your garment reference is a photo of a different model wearing the item, describe only what should be tried on — and do not specify what should be kept from the person image.
So if the reference photo shows a model in both a top and trousers, but you only want the top transferred, the prompt names only the top. Adding "keeping the original trousers" is the instinctive move and it works against you: it puts the person image's clothing into the description, which is exactly the region the model is trying to replace.
Say what moves. Stay silent about what stays.
The three reference types
BFL documents three shapes of garment reference, all supported:
| Reference type | What it is | When to use |
|---|---|---|
| Packshot | A single garment on plain background | You have individual product shots |
| Composed outfit | Multiple garments arranged on one canvas | Full-look try-on in a single call |
| On-model | The garment worn by a different person | You only have lifestyle or campaign imagery |
The composed-outfit option is the one worth planning around: it means a complete look does not require multiple sequential calls, each of which is another chance for identity drift.
Building the product imagery that feeds this? Flux 3 AI is a browser workspace for generating and editing product visuals, references and campaign frames. Open the image generator or see the credit plans.
Why it is built for interactive use
BFL states the endpoint is optimised for low latency, and positions it for virtual fitting rooms and social media filters — interactive contexts where a user is waiting.
That framing has an integration consequence. If your UX has a person standing in front of a camera, you are budgeting for a round trip they will sit through, so the usual advice applies harder than normal: prepare garment references ahead of time, keep them small, and do not stack a second model call on top.
There is a free interactive demo at flux-tools.bfl.ai/virtual-try-on — worth running your own garments through before writing any integration code.
A checklist for catalogue work
- Person image with a clear, unobstructed pose
- Garment reference: packshot if you have it, composed canvas for full looks
- Prompt opens with the identity lock, not the garment
- Garment described by cut, length, and any printed text
- For on-model references: name only the item being transferred
- Nothing in the prompt about what to keep from the person image
FAQ
What is the FLUX VTO prompt formula?
The person of image 1, maintaining exactly their face and pose, wearing the [garment description] of image 2.
Does the garment description matter? Yes. BFL says a concise description significantly improves results, particularly for complex garments.
Can I use a photo of someone else wearing the garment? Yes — describe only the item you want transferred, and say nothing about what to keep.
Can I try on a whole outfit at once? Yes, by arranging multiple garments on a single canvas as one reference.
Is there a free way to test it? BFL hosts a free interactive demo.
Is it fast enough for a live fitting room? BFL positions the endpoint as latency-optimised for exactly that use case.
Does VTO need a mask? No. Unlike Erase, it works from the person image and garment references plus the prompt.
Bottom line
VTO output quality comes down to two habits: lock the identity in the first clause, and describe the garment like a product listing. The counterintuitive rule — say nothing about what should stay — is what makes on-model references work.
Run your own garments through the free demo before you build anything. And if you need the product photography that feeds a try-on pipeline in the first place, start in the Flux 3 AI workspace.
Sources
- FLUX Virtual Try-On (BFL documentation) — prompt formula, garment description guidance, on-model rule, reference types, latency positioning, demo link
- FLUX VTO: Virtual Try-On at scale (BFL, May 28, 2026) — release and catalogue-scale framing
- FLUX Tools model page — where VTO sits in the family
- Fashion use case guide (BFL docs) — related clothing workflows
- Character & Style Consistency (BFL docs) — identity preservation across edits
- Image generation quick start (BFL docs) — the async pattern VTO shares
- BFL API pricing — how Tools calls are billed
- FLUX.2 model page — multi-reference capability underpinning garment transfer
- Official FLUX.2 prompting guide — general prompt rules that still apply
- FLUX Erase (BFL docs) — the mask-based sibling tool, for contrast
Scope note: documented behaviour as of July 30, 2026. Flux 3 AI is an independent creator workspace, not affiliated with Black Forest Labs.
