FLUX 3 Image: Bounding Boxes, 4K Output, and Layout as a Configuration File

Black Forest Labs released FLUX 3 Image on October 1, completing the image component of the FLUX 3 family. It is available through the company's API and Playground, with a 50 percent introductory discount running through October 8 and an open-weights release promised in the coming weeks.

The three headline claims are native 4K output, element-level layout control, and region-preserving edits. Each of them is worth taking apart, because the third one is doing more work than the marketing suggests.

The grid replaces the prose

FLUX 3 Image takes a JSON array of bounding boxes appended to the prompt, expressed on a normalized 0-1000 coordinate grid. Each element gets an ID, a text description, and a box in the format [y_min, x_min, y_max, x_max]. The model then places that element inside that box, independent of the output resolution.

This is the part the release is built around, and it is a genuine shift in how layout gets specified. In most text-to-image workflows, spatial arrangement is described in natural language ("a bottle on the left third, a hand entering from the right") and the model interprets that as a soft preference. The result is a distribution of outputs that usually lands somewhere near the description, sometimes not.

A coordinate grid removes the interpretation step. You are no longer describing a layout and hoping; you are declaring one. For a designer working in a layout tool, that maps onto an existing mental model. For a pipeline generating hundreds of product shots that all need the model in the same position, it turns an aesthetic problem into a configuration problem.

There is a caveat that comes with any coordinate system: the boxes do not tell the model what goes in them, and a description that specifies the wrong object will be rendered confidently in the wrong place. Precision in geometry does not automatically produce precision in semantics.

4K without the upscale step

FLUX 3 Image generates natively up to 5,456 by 3,072 pixels (about 16.8 megapixels) across multiple aspect ratios, without a super-resolution pass.

Most image models have topped out around 1024 to 2048 pixels natively, with anything larger handled by upscaling afterward. That is fine for screen delivery and increasingly awkward for print collateral, e-commerce hero images, and large-format display, where the post-processing step tends to smooth away exactly the texture detail that justifies the resolution.

Generating at 4K also changes what a workflow has to validate. An upscaler is a separate stage with its own failure modes, and removing a stage removes a category of QC. Whether the native output holds up at full size, rather than merely avoiding an obvious artifact, is something only hands-on testing will settle.

Region edits and the pixel-identity claim

The third feature is partial editing: regenerate only a specified bounding-box region while leaving everything outside it alone. Black Forest Labs reports that 67.8 to 89.7 percent of pixels remain bit-identical after an edit.

That range is wide, and the spread carries information. The lower end implies meaningful drift across the untouched area; the upper end is close to a true mask edit. Where a given edit lands likely depends on how much the regenerated region constrains its surroundings, lighting spill, shadows, and contact edges all create pressure to alter neighboring pixels for coherence.

Still, publishing a pixel-identity figure at all is a useful move. The drift problem in iterative image editing has been the central complaint against diffusion editing for two years: each successive round of revision degrades the parts you were happy with, until you give up and start over. A model that quantifies how much it preserves has at least stated the target.

FLUX 3 Image also accepts up to ten reference images in a single request, positioning subjects according to the prompt and the bounding-box layout. That combination (multiple references plus declared placement) is where the model starts to look less like an image generator and more like a compositing tool with a generation backend.

Ten references also raises the question of what a "reference" is for. In most current workflows a reference image means style transfer: give the model a look and it approximates it. Positioning each referenced subject inside a declared box is a narrower and more mechanical promise. It says: this object, here. Whether the model honors that under load (ten subjects, a 4K canvas, overlapping boxes) is the kind of thing that shows up in production use rather than in a launch post.

The training note worth noticing

Buried in the technical details: FLUX 3 Image was trained with Black Forest Labs' Self-Flow architecture across image, video, and audio data jointly. The FLUX 3 base is a multimodal flow model covering all three modalities, and the image model is one face of it.

That matters for how to read the roadmap. If image, video, and audio share a training substrate, the "open weights in a few weeks" promise reaches past this release, and speaks to a family of models that will be built on the same base, and about a company positioning itself as the open-weight alternative to the closed multimodal stacks from Google and OpenAI.

It also explains the sequencing. Black Forest Labs shipped the video and audio components of FLUX 3 first and the still-image model last. For a company whose reputation was built on image generation, releasing the image model after the video model is an odd order, unless the image model is the hardest one to get right on the shared base. The bounding-box controls may be less a feature and more a workaround for what joint training across three modalities does to fine-grained spatial fidelity.

What to watch for

For now, the practical read is narrower. FLUX 3 Image is a paid API product with layout controls that previously required a compositing step, native resolution that removes an upscaler, and region edits that come with a published fidelity number. The question of whether any of it is better than the alternatives is a benchmarking question, and benchmarks are what the next few weeks of independent testing are for.

Three things are worth tracking once the weights drop. First, whether the bounding-box API survives the open release intact or becomes an API-only feature, which would be a meaningful split between the paid and free tiers. Second, whether community fine-tunes can hit the same layout precision on smaller hardware, or whether the coordinate behavior depends on scale. Third, whether the pixel-identity figure holds up in independent reproduction or proves to be a best-case measurement on favorable edits.

The broader signal is the one the release notes state almost in passing: the competition in image generation has moved from how striking a single output can be to whether the hundredth output matches the first. Layout grids, region preservation, and pixel-identity metrics are all answers to that question. None of them make a beautiful image. They make a repeatable one.