Most people meet FLUX through text-to-image and never find out that Black Forest Labs also ships four task-specific endpoints that do things a generation model handles badly.
They are worth knowing about because of one property: three of the four take no prompt at all. You send an image, you get an image back. No prompt engineering, no negative-prompt folklore, no seed hunting.
Sourcing note: every endpoint path, parameter rule, and behaviour below comes from Black Forest Labs' own API documentation for each tool, plus their release announcements. Where a tool has a free public demo, the link is BFL's own. Last verified July 30, 2026.
What this guide solves
The pain point: people try to do erase, expand, and deblur work with prompts, fighting a generation model into doing a restoration job. The results are inconsistent because the wrong tool is being used.
The differentiator: this is the four tools side by side with the thing that actually decides which to reach for — what input each one needs (mask? prompt? nothing?) — plus the mask convention that trips up every first Erase integration.
The four tools at a glance
| Tool | What it does | Needs a prompt? | Needs a mask? | Endpoint |
|---|---|---|---|---|
| Erase | Removes an object, reconstructs what was behind it | No | Yes | /v1/flux-tools/erase-v1 |
| Outpainting | Extends an image past its borders | No | No | /v1/flux-tools/outpainting-v1 |
| Deblur | Sharpens a blurry photo | No | No | /v1/flux-tools/deblur-v1 |
| Virtual Try-On | Puts a garment on a person | Yes | No | VTO endpoint |
All four use the same async pattern as the rest of the FLUX API: POST the job, get a polling_url back, poll until the result is ready.
Erase — the one with a mask convention
FLUX Erase removes a masked object and rebuilds the scene behind it in a single call. BFL states it is powered by FLUX.2 Klein 9B, which is why it is fast.
The rule that breaks first integrations: the mask is a black/white PNG at the same resolution as your input image, where white (255) marks the pixels to remove and black (0) marks pixels to keep. Get that inverted and you will erase the entire background instead of the object.
No prompt is accepted — the model uses a built-in instruction. That is a feature: it removes the "describe what should be there instead" guessing game entirely.
Use it for: product photography clean-up, removing unwanted elements from a shot, scene simplification, and privacy-aware editing (removing identifiable details before publishing).
Outpainting — aspect-ratio surgery
Extends an image naturally in any direction, filling the new region with coherent content. You POST the input image plus the target canvas dimensions; the model extends the scene on its own.
No prompt needed. BFL's own example notes the demo output was produced with no prompt at all.
Use it for: turning a square product shot into a banner, reformatting one asset for multiple social placements, or giving a cramped composition breathing room instead of cropping in.
Free demo, no API key: flux-tools.bfl.ai/outpainting.
Deblur — restoration, not reimagining
Takes a single blurry image and returns a sharper version of the same scene. BFL is explicit that it is designed to stay faithful to the input: scene, character identity, colours and composition are preserved.
No prompt and no mask are accepted — the whole image is sharpened automatically. Input can be base64 or an HTTP(S) URL.
This distinction matters. An upscaler or a generation model asked to "make this sharper" will happily invent detail that was never there. Deblur is positioned as restoration, so the output should still be the same photo.
Virtual Try-On — the one that does need a prompt
Generates a try-on result from a person image plus one or more garment references. BFL says the endpoint is optimised for low latency, aimed at interactive uses like virtual fitting rooms and social filters.
This is the only one of the four with real prompt craft, and BFL publishes the formula:
The person of image 1, maintaining exactly their face and pose, wearing the [GARMENT DESCRIPTION] of image 2.
Their guidance on top of that:
- A static prompt works as a default, but a concise garment description meaningfully improves quality — especially for complex garments.
- Be specific about fit and category: "the oversized tee", "the 7/8 length pants", "the cap with 'Keep FLUXing' text".
- When your garment reference is a photo of another model wearing it, describe only what should be tried on. Do not describe what to keep from the person image.
Garment references can be a packshot, a full outfit arranged on one canvas, or an on-model photo. Free demo: flux-tools.bfl.ai/virtual-try-on.
Want the everyday version of this without an API key? Flux 3 AI runs generation, editing, reference workflows and upscaling in the browser. Open the image generator or see the credit plans.
How to choose
- Something has to leave the picture → Erase (build a mask)
- The frame is the wrong shape → Outpainting (give target dimensions)
- The photo is soft but otherwise fine → Deblur (send it as-is)
- A garment has to go on a person → VTO (write the formula prompt)
- You want a different picture → not a Tool at all; that is FLUX.2 text-to-image or editing
The rule of thumb: if you can describe the job without describing the desired image, it is probably a Tool.
Release timeline
Per BFL's announcements, the Tools shipped through 2026: Outpainting on May 14, Erase on May 21, and Virtual Try-On on May 28. Deblur is documented alongside them in the Tools section.
FAQ
Do FLUX Tools need prompts? Erase, Outpainting and Deblur take no prompt. Only VTO does.
What mask format does Erase need? A black/white PNG at the same resolution as the input, white for pixels to remove, black to keep.
Can I try them without an API key? Yes — BFL hosts free demos for Outpainting and Virtual Try-On.
What model powers Erase? BFL documents it as FLUX.2 Klein 9B.
Does Deblur change the image? It is built to preserve scene, identity, colours and composition while restoring detail.
How are Tools billed? Through the same credit system as the rest of the API; see BFL's pricing page and calculator for the per-tool rates.
Bottom line
FLUX Tools are the unglamorous half of the FLUX family and often the half that actually finishes the job. Three of the four need nothing but an image, which makes them far more reliable than trying to prompt a generation model into doing restoration work.
If you are producing images regularly, the useful mental shift is to stop reaching for text-to-image by default. Generate in the Flux 3 AI workspace when you need something new, and reach for a Tool when you need something fixed.
Sources
- FLUX Erase (BFL documentation) — endpoint, mask convention, Klein 9B backing, use cases
- FLUX Outpainting (BFL documentation) — endpoint, canvas dimensions, no-prompt behaviour, free demo
- FLUX Deblur (BFL documentation) — endpoint, fidelity guarantees, input formats
- FLUX Virtual Try-On (BFL documentation) — prompt formula, garment reference guidance, latency positioning
- FLUX Tools model page — product overview
- FLUX Outpainting announcement (BFL, May 14, 2026) — release date and framing
- FLUX Erase announcement (BFL, May 21, 2026) — release date
- FLUX VTO announcement (BFL, May 28, 2026) — catalogue-scale positioning
- BFL API pricing — credit model and calculator
- Image generation quick start (BFL docs) — the shared async request pattern
Scope note: endpoints and behaviour as of July 30, 2026. Flux 3 AI is an independent creator workspace, not affiliated with Black Forest Labs.
