Black Forest Labs released FLUX 3 Image on October 1 and 2, filling the last open slot in its FLUX 3 family after FLUX 3 Video went on general availability. The video model got native audio and draft mode. The image model went after something narrower and more consequential for anyone building a pipeline: a machine-readable contract between what you ask for and the pixels that come back.
That contract is layout prompting, and it is the part of this release worth studying. Instead of negotiating composition through prose, you hand the model a plan made of two pieces: one global caption, then an element table, a JSON array in which every row carries an id, a bounding box, and a description of what belongs inside it.
Inside the Layout Prompt Contract
The geometry is deliberately plain. Whatever aspect ratio you choose, the canvas is a 0 to 1000 grid on both axes, and each box is written as [y_min, x_min, y_max, x_max]. Black Forest Labs' own demo builds a twilight beach scene from five rows, including a title element, a domed structure, swimmers and a seated crowd, then cites those ids from a single caption line. An LLM can write that table from one sentence and an aspect ratio, which is the workflow the company is clearly optimising for.
Two details matter more than they look. First, the prompt upsampler will expand your short request into the dense caption FLUX 3 was trained on, and it may suggest extra elements, but every box you drew reaches the model verbatim with the same id and the same coordinates. Anything the upsampler invents survives only if the caption refers to it. Second, editing reuses the same vocabulary: each box can be re-described, replaced or moved, and everything you leave alone comes back untouched, labelled as locked and unchanged in the playground. In one demo, a surfer's wetsuit and board are recoloured in a single request while the wave and the cloud layer are explicitly pinned where they were.
Multi-reference composition follows the same grammar. You can supply up to ten reference images, each assigned a token in the order you add it, starting at ref_image_0, and the model decides where each one sits and how large it renders. The prompt is one line that names the tokens.
Practical notes from BFL's own documentation and an independent week-long hands-on report published October 9:
- There is no negative_prompt, guidance, seed or prompt_upsampling field in the API. Unsupported fields return an error rather than being quietly ignored, so control has to live inside the prompt text.
- No seed means two runs of the same prompt differ. Generate a batch before concluding a prompt failed.
- 15 fixed aspect ratios run from 21:9 to 9:21, plus auto, which requires the images field immediately after the first image.
- Output is native 2K and 4K, not upscaled: the published 4K sample lands at 5456 x 3072 pixels, 16.8 MP from the model itself.
The hands-on report is where the rough edges show. Negation backfires. Asking for a parking lot without cars produced a lot packed with SUVs; describing what should be there instead, a deserted rooftop lot with a row of shopping carts and cold fluorescent light, produced the intended image.
Photographer vocabulary beats quality adjectives. Aperture ranges (f/1.4 to f/2.8 for shallow depth of field, f/8 to f/16 for deep focus), focal lengths (35mm documentary, 85mm portrait, 135mm compression), lighting terms such as golden hour, blue hour and Rembrandt lighting, and even named gear like a Hasselblad X2D or Kodak Portra 400 film all outperformed stacking words like ultrarealistic.
Text rendering is the clearest strength. Strings wrapped in quotes, with font weight, case, colour and alignment specified, came out legible and correctly placed, and a partial headline swap left the rest of a poster stable. In a same-prompt comparison against Seedream 5.0 Pro and GPT Image 2.5 Flare, FLUX 3 Image was judged the most balanced across detail, scene fidelity and composition, and the most accurate at lettering; Seedream added unintended words, while the GPT image had the strongest single headline.
Cost and workflow are worth planning around. The standard loop is a fast 768 x 768 draft, then 2K or 4K once composition is settled, but changing resolution regenerates the picture and can nudge the layout, and a 4K render takes minutes. BFL's pricing calculator lists 0.048 dollars per image at the 768 x 768 tier, and the launch discount of 50 percent on API access ran through October 8.
For teams that need the model on their own metal, BFL sells a commercial weights licence for self-hosted deployment and fine-tuning, and an open-weight FLUX 3 Dev build is promised in the coming weeks. That timeline keeps pressure on an editing tier that is filling up fast: Ideogram shipped an editing-focused 4.5 model days earlier with open weights also on the way, and cheaper hosted alternatives keep cutting per-image prices.
The larger point is architectural rather than aesthetic. FLUX 3 Image behaves like a renderer for a document an agent wrote, with ids and coordinates surviving the round trip and untouched regions provably preserved. That is a much better fit for automated production than a model that treats every request as a fresh negotiation, and it explains why the most overlooked feature in this release is the least visual one.
Comments