Black Forest Labs: FLUX.2 Max

FLUX.2 [max] is Black Forest Labs’ top-quality member of the FLUX.2 image-generation and editing family. The manufacturer positions it for demanding final assets, strong adherence to a complex brief and consistency across edits. It creates an image from a written description or uses supplied images as visual references. Its distinguishing family capability is grounding search: a request can draw on current information instead of relying only on the scene described in the prompt.

Real-world information and the visual brief

The maker documents web search for subjects such as current weather, a recent sporting event or a historical setting. Identify the actual location, event and time relevant to the picture, then describe the artistic treatment separately. A grounded weather-card concept could specify the city, which condition to visualize, the exact title and the layout. Search informs the generation; an attractive illustration still needs its visible facts checked against the relevant source before becoming a factual communication.

This search capability belongs to [max] in the published FLUX.2 comparison. The other named members have their own roles: [pro] for production, [flex] for adjustable sampling and typography, and [klein] for compact interactive generation. Keep the selected model name with the result so a later revision does not silently replace the groundable profile with a different edition.

References for demanding edits

The documented request profile takes up to eight reference images; the maker’s interactive interface separately allows ten. Assign a role to each reference, such as character appearance, product design, clothing, room or visual style. A composition can combine the chair from one picture with the room from another while keeping the chair’s silhouette and material. For an existing scene, identify the object to change and the details to retain, including light direction and surrounding geometry.

The family’s examples include background changes, replacement of objects, furniture compositions, text edits and consistent characters. An original final-asset brief might place a supplied watch on dark folded fabric, retain the dial and strap design, and change only the lighting to a soft side-lit treatment. Name which source supplies each feature and review the resulting dial, markings, material and reflections together. A preservation instruction establishes the goal; the returned picture is the evidence to compare with it.

Text, colors and prompt interpretation

The FLUX.2 guide recommends a subject–action–style–context structure, with important elements first. It suggests 10–30 words for quick concepts, 30–80 for ordinary projects and longer descriptions for a complex scene. These are writing recommendations, not hard context limits. Natural language and JSON-structured scene descriptions are both documented ways to organize subjects, placement, light and style. A structured prompt is an image brief rather than a promise of a structured textual response.

For visible lettering, quote the wording, specify its position and describe the typographic hierarchy. Associate each hexadecimal color with a particular object or text element; a color code without a named target is ambiguous. The maker guide uses positive descriptions instead of a negative-prompt mechanism. For a product scene with empty space for a headline, describe that empty area and its intended position directly.

The exact [max] request specification applies prompt upsampling by default. Its disable_pup control turns that expansion off when a carefully written brief should be used as supplied. This is distinct from a claim that every word will appear unchanged in the final image. Preserve the actual brief and selected setting with the result when assessing whether an unexpected visual decision came from the instruction or its interpretation.

Output and a useful final asset

The family guide documents images up to 4 megapixels, with explicit dimensions at least 64×64 and multiples of 16. Up to 2 megapixels is its ordinary-use recommendation. The exact [max] profile lists JPEG, PNG and WebP, with JPEG as the default, and an optional random seed. Choose the output shape for the intended placement before arranging the composition. The model’s image generation is separate from resizing an already finished asset.

Retain the image, references, prompt, dimensions and model identity when selecting the final version. The producer’s FLUX.2 guide warns that its signed result link is temporary, with a ten-minute retrieval window. Inspect fine text, object identity, layout, light and any facts drawn from search at the intended display size. The useful artifact is the chosen visual with enough context to continue an exact revision, rather than a scene whose important details have never been compared with the brief.