# FLUX 3 Image

> Source: https://aiwiki.ai/wiki/flux_3_image
> Updated: 2026-10-02
> Fact-checked: 2026-10-02
> Categories: AI Models, Generative AI, Image Generation
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "FLUX 3 Image." aiwiki.ai, 2 Oct 2026. https://aiwiki.ai/wiki/flux_3_image
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

FLUX 3 Image is an [image generation](https://aiwiki.ai/wiki/ai_image_generation) and editing model developed by [Black Forest Labs](https://aiwiki.ai/wiki/black_forest_labs) (BFL). It generates images from written descriptions, combines up to ten reference images, and accepts bounding boxes to guide the placement or editing of individual elements.[1] BFL's release notes date the product's release to October 1, 2026.[2]

## Background

BFL introduced the FLUX 3 family on July 23, 2026. The company described a shared architecture trained jointly on images, video, and audio, based on its Self-Flow research. Its rollout separated image synthesis and editing from [FLUX 3 Video](https://aiwiki.ai/wiki/flux_3_video), action prediction, and a planned open-weight backbone called FLUX 3 Dev. The July announcement described image evaluations as preliminary and said image early access would follow in subsequent weeks. That announcement therefore documents the family and its planned rollout, rather than the October image release.[3]

## Generation and editing

A text prompt can describe a new scene without reference images or boxes. For editing, the request supplies an image and an instruction identifying the requested change. Multiple references can contribute different parts of a composition, such as a setting from one image and an object from another. BFL's overview also documents web and image search grounding, enabled by default; disabling it makes generation rely on the supplied prompt instead of external searches.[1]

The product page describes native 2K and 4K generation and illustrates 4K output with an image measuring 5456 x 3072 pixels. Its agent workflow uses a language model to turn a short scene description and aspect ratio into a caption and an element table. The user can adjust that proposed layout before generation.[4]

An edit does not always require a manually drawn box. The editing documentation says FLUX 3 Image expands instructions into a detailed prompt and, for many edits, identifies the named element and supplies a box automatically. The returned `result.prompt` exposes that expanded instruction. References can be identified by their position in the input list or by their content.[6]

## Bounding-box control

The layout prompt contains a scene caption followed by a JSON array. Each generated element has an `id`, a `bbox`, and a `desc`; the caption refers to the identifier in angle brackets. Coordinates are integers on a normalized 0-1000 grid, ordered `[top, left, bottom, right]`, equivalent to `[y_min, x_min, y_max, x_max]`. The boxes are part of `prompt`, not a separate API parameter.[5]

Editing rows use `from`, `src_bbox`, and `tgt_bbox` to connect source elements to their intended output positions. `ref_image_0` identifies the first reference image.[5]

| Intended edit | Row configuration |
| --- | --- |
| Retain an element | Reference an input element and use equal source and target boxes.[5] |
| Move or resize | Reference an input element and change its target box.[5] |
| Add, replace, or recolor | Set `from` and `src_bbox` to `null`; describe the result at `tgt_bbox`.[5] |
| Remove | Reference the source element and set `tgt_bbox` to `null`.[5] |

Boxes guide placement and scale; the documentation explicitly says they are not clipping masks, and elements can extend beyond them.[5]

## API inputs and output

The service accepts `POST https://api.bfl.ai/v1/flux-3-image`, authenticated with an API key in the `x-key` header. Its request schema includes the following fields.[7]

| Field | Meaning |
| --- | --- |
| `prompt` | Required textual request.[7] |
| `images` | Optional references supplied as HTTP(S) URLs or base64 data.[7] |
| `aspect_ratio` | Fifteen explicit ratios, from `21:9` to `9:21`, plus `auto`.[7] |
| `resolution` | `768sq`, `1k`, `1.5k`, `2k`, or `4k`; default `1k`.[7] |
| `safety_tolerance` | Moderation setting from `0` through `4`, with `0` strictest and `2` the default.[7] |
| `grounding` | Enables external web and image research; default `true`.[7] |
| `version` | Defaults to `latest`, the documented allowed value.[7] |

Resolution classes specify equal pixel areas; actual dimensions depend on the aspect ratio. With `auto`, a ratio requested in the prompt takes precedence, followed by the first reference's framing, with square output as the fallback.[7]

Each reference must be at least 256 x 256 pixels and no larger than 16 megapixels.[1] Generation and editing share the same endpoint without a `mode` field. The text-to-image guide warns that unsupported fields, including `seed`, `width`, and `input_image`, produce HTTP `422` validation errors.[8]

Requests run asynchronously. Submission returns a task identifier and `polling_url`, which the client must use because it points to the region holding the task. The client polls until completion and downloads `result.sample` when the status becomes `Ready`; the image URL must be used within one hour. A completed result also reports the expanded prompt and generation duration.[8]

The guide distinguishes blocked inputs from blocked outputs through `Request Moderated` and `Content Moderated` statuses. It also cautions that a failed task can return HTTP `503` with a normal JSON response: the status in that response determines what happened.[8]

## Access and pricing

BFL provides browser access through its playground and a Precise Editing tool.[1][10] The latter accepts an uploaded image through file selection, drag-and-drop, or clipboard paste, and offers sample images for editing.[10] BFL also advertises a commercial weights license allowing companies to fine-tune and deploy the image model on their own infrastructure.[4]

The pricing documentation lists the following per-image charges, checked on October 2, 2026. API and playground prices are the same, with one credit worth US$0.01.[9]

| Resolution class | Documented output size | Price per image |
| --- | --- | --- |
| `768sq` | 768 x 768 | US$0.041.[9] |
| `1k` | About 1 megapixel | US$0.048.[9] |
| `2k` | About 4 megapixels | US$0.100.[9] |
| `4k` | About 16 megapixels | US$0.607.[9] |

The pricing table does not list a separate `1.5k` rate. The submission response includes the cost of the individual request.[9]

## Editing limits

BFL's editing guide qualifies its pixel-preservation claims: pixels outside edited boxes usually remain identical, but shadows, reflections, or nearby lighting can change. It also reports that newly requested elements sometimes fail to appear in very small boxes, citing boxes around 40 x 25 pixels in its tests. These are vendor-documented observations, not an independent guarantee of unchanged pixels. The guide advises giving text its own box and stating additions or removals in the written instruction as well as the element rows.[6]

## References

1. Black Forest Labs. [FLUX 3 Image](https://docs.bfl.ai/flux_3/flux3_image_overview). Documentation, accessed October 2, 2026.
2. Black Forest Labs. [Release Notes](https://docs.bfl.ai/release-notes). FLUX 3 Image entry, October 1, 2026.
3. Black Forest Labs. [FLUX 3: Real World Models](https://bfl.ai/blog/flux-3). July 23, 2026.
4. Black Forest Labs. [FLUX 3 Image](https://bfl.ai/models/flux-3-image). Product page, accessed October 2, 2026.
5. Black Forest Labs. [Bounding boxes with FLUX 3 Image](https://docs.bfl.ai/flux_3/flux3_image_bounding_boxes). Documentation, accessed October 2, 2026.
6. Black Forest Labs. [FLUX 3 Image Editing](https://docs.bfl.ai/flux_3/flux3_image_layout). Documentation, accessed October 2, 2026.
7. Black Forest Labs. [Generate an image with FLUX 3](https://docs.bfl.ai/api-reference/utility/generate-an-image-with-flux-3). API reference, accessed October 2, 2026.
8. Black Forest Labs. [FLUX 3 Text to Image](https://docs.bfl.ai/flux_3/flux3_image_generate). Documentation, accessed October 2, 2026.
9. Black Forest Labs. [API pricing](https://docs.bfl.ai/quick_start/pricing). FLUX 3 Image rates, accessed October 2, 2026.
10. Black Forest Labs. [Precise Editing](https://flux-tools.bfl.ai/precise-editing). Browser tool, accessed October 2, 2026.