Flux 3

• 🏢 Black Forest Labs' newest generation, succeeding the FLUX.2 image family
• 🆕 Released October 2026 as a text-to-image endpoint
• 🎯 Positioned for photorealism, prompt fidelity and complex scene composition
• 🌐 Per BFL, Flux 3 is one multimodal model spanning image and video
• 📏 FLUX.2 predecessors generate photorealistic images up to 4 megapixels
• 🔧 Sibling editing and video variants share the same Flux 3 generation
• 👁️ Video side covers keyframe control, clip extension, prompt-based editing
• ⚡ BFL lists video output resolutions ranging from 720p to 4K

Black Forest Labs

Black Forest Labs is a generative AI company based in Freiburg im Breisgau, Germany, founded by former members of Stability AI. The lab is best known for developing the Flux family of text-to-image models, which generate images from natural language prompts…

Flux 3 is the image-generation entry point to Black Forest Labs' third-generation FLUX family, served here as a text-to-image model. BFL describes Flux 3 as a single multimodal model rather than a set of separate tools, covering text-to-image alongside video generation, keyframe control, clip extension and prompt-based video editing, with output resolutions from 720p to 4K on the video side.

Compared with its same-family predecessors Flux 2 Pro and Flux 2 Max, the generational shift is one of scope. The FLUX.2 models were image-only systems, generating photorealistic images at up to 4 megapixels with support for reference-image conditioning. Flux 3 folds image and video generation into one model family instead, per BFL's own description.

Within this catalog, Flux 3 sits alongside Flux 3 image editing for inpainting and instruction-based edits, and the video siblings Flux 3 text-to-video, Flux 3 image-to-video and Flux 3 First Last Frame. An editing counterpart to the previous generation, Flux 2 Max Edit, remains available for comparison.

No independently verified benchmark figures for Flux 3 image quality are available at the time of writing, so capability descriptions here should be read as vendor-described rather than measured. Users evaluating the model are best served by direct side-by-side prompting against the FLUX.2 entries in this catalog.