Black Forest Labs releases FLUX 3 Image with bounding-box layout and native 4K output

Black Forest Labs launches FLUX 3 Image

Black Forest Labs has released FLUX 3 Image, a new image generation model whose page surfaced on Hacker News' front page, with the listing dated 2 October 2026. According to Black Forest Labs, the model is built around precise control: users can dictate where every element sits, edit several regions at once, blend reference images, and output at high resolution without extra tooling.

Layout declared as data

The headline feature is image composition driven by bounding boxes. Rather than describing an entire scene in one paragraph, a user supplies a table of elements, where each entry carries an identifier, a rectangular region of the canvas, and a short text description, and the model fills each region accordingly.

Black Forest Labs demonstrates the approach with a poster titled "Le Festival du Soleil", assembled from five declared elements: a cream-coloured serif title, a faint coastal town, a large concrete dome, silhouetted swimmers in dark water, and a crowd seated on a beach. Each element was pinned to its own coordinates, and the model composed the full scene from that layout combined with a scene-level prompt.

The structured mode is optional. According to the company, FLUX 3 Image also works as a conventional text-to-image model with strong prompt adherence and a built-in sense of composition. One showcased series is a set of 2:3 studio portraits generated from a single prompt that specifies printed shirt text down to colour and cuff details.

Editing, references and resolution

Three further capabilities round out the release:

• Multi-region editing: several changes can be applied in a single pass, region by region, while everything outside those regions stays untouched. The company's example recolours a wetsuit and a board this way, and it describes the editing capability as pixel-accurate, meaning any part of an image can be changed with the original preserved elsewhere.
• Reference-driven composition: up to ten reference images can be combined into a single, deliberately arranged output, which Black Forest Labs pitches as a route to a considered composition with less iteration.
• Native high resolution: images render at full 2K or 4K rather than being upscaled afterwards, which the company says preserves fine textures, faces and colours. A sample scene on the page, a soba shop interior, measures 5456 x 3072 pixels and, per the company, comes straight from the model.

Why it matters

Composition has long been the weak point of prompt-only generators: asking for a person on the left and a sign at the top is a negotiation, not a specification. Bounding boxes change that. They are already the lingua franca of design software, annotation tools and detection pipelines, so a model that consumes them natively can slot into existing production workflows instead of forcing designers to steer placement through prose.

Two of the other capabilities push in the same direction. Rendering typography inside a declared box makes the model viable for posters and layout-heavy design work rather than loose illustration, since titles stop being an afterthought the model mangles. And native 4K output removes the separate upscaling step that most current pipelines bolt on after generation.

The material surfaced via Hacker News covers capabilities rather than distribution or pricing, so the practical question for developers building image tooling is how broadly the model will be made available. If it performs as documented, bounding-box prompting is a credible candidate for a standard interface in the next generation of image tools.