FLUX.2 Max's feature set is built on the most capable architecture Black Forest Labs has ever shipped. The Mistral-3 24B vision-language backbone gives it world knowledge and spatial reasoning. The Rectified Flow Transformer gives it photorealistic rendering. Combined, they produce an image model that understands what things look like, where they should go, and how they interact with light and physics.

Web-Grounded Generation

FLUX.2 Max integrates real-time web context into the generation process. Ask for a product that launched last week and the model looks it up. Request an image tied to a current event and the model pulls in relevant visual information. This eliminates the "stale training data" problem where models render outdated versions of products, logos, and public figures. No manual reference uploading needed for current subjects.

Character Consistency Across Scenes

Create a character and maintain their facial features, proportions, expressions, clothing, and visual identity across completely different images. Change the scene, change the style, change the lighting. The character stays recognizable. This works across complex edits, multiple references, and changing environments. For campaign work where a brand character or product ambassador appears across 50+ images, this is the feature that matters most.

Multi-Reference System (Up to 10 Images)

Upload up to 10 reference images simultaneously. The model synthesizes identity from multiple angles, contexts, and conditions. This is not single-image style transfer. It is multi-view identity extraction. A product photographed from 5 angles, a character in 3 outfits, and 2 environmental references can all feed into a single generation. The model understands what each reference contributes and composites them coherently.

Retexturing

Redesign surfaces and materials with precision. Swap leather to denim, wood to marble, matte to chrome. FLUX.2 Max preserves the shape, geometry, lighting, and spatial context of the original object while replacing only the surface material. This is a production feature: product teams can explore material variants without re-rendering or re-photographing.

32K Token Context Window

The largest context window among image generation models. Write detailed, multi-paragraph prompts with scene descriptions, character specifications, lighting directions, composition instructions, and style references. The model processes the full prompt without truncation. This is the Mistral-3 VLM backbone at work: it reads prompts the way a language model reads a document, with full contextual understanding across the entire input.

Photorealistic Physics and Lighting

Fabric textures with visible weave. Architectural materials with correct reflection properties. Skin with subsurface scattering. Lighting that follows physical rules: shadows fall correctly, reflections match the environment, and specular highlights respond to material roughness. The gap between FLUX.2 Max output and real photography is narrow enough that the model is used for marketplace product images that look indistinguishable from studio shoots.