Black Forest Labs unveils FLUX 3, a multimodal video model learning from images, audio, and video

Black Forest Labs, the independent research lab behind the widely adopted FLUX.1 image model, has released FLUX 3, a multimodal foundation model designed to generate video by learning jointly from images, audio, and video during training. The consolidated architecture is the key differentiator: rather than treating each modality as a separate input stream, FLUX 3 learns the relationships between them. As Black Forest Labs explains in their release, images capture spatial structures at a point in time, video restores the temporal dimension and reveals physical dynamics, audio reveals causal relationships between mechanical phenomena and acoustics that vision alone cannot detect, and language links these perceptions to goals and instructions. Learning from one modality produces a good model of that projection; learning from all simultaneously reveals deeper truth—the sound must match the impact, the motion must obey mass, the future must follow from the past, and the modalities become evidence of a single underlying reality rather than separate channels.

The model excels at text-to-video generation, accepting both simple prompts and dense, cinematic descriptions. Users can describe visuals alone or pair them with audio details (e.g., "the roar of the engine and the crunch of sand under tires"), and FLUX 3 integrates both. The model also defaults to multiple scene cuts unless told otherwise, allowing cinematic pacing without manual stitching. For image-to-video, users provide a start frame and an end frame along with a description of the physical transformation—the model then generates a continuous shot showing the natural change between them, preserving camera angle and object identity while morphing conditions (e.g., a rusted junkyard car becoming a restored, roadworthy vehicle). A new parameter, start_video, allows video continuation: users supply an existing clip, and FLUX 3 picks up from the final frame while preserving momentum, framing logic, and audio. Timestamped prompts enable multiple scenes and camera angles in a single generation; users write [0-4s]: wide shot, [4-8s]: close-up, [8-12s]: tracking shot directly into the prompt, and the model cuts between them in one pass. FLUX 3 is not locked into photorealistic hyper-realism; it can generate stop-motion claymation (with visible fingerprints and seams), comic-inked art with chromatic aberration, and vintage VHS home-video footage complete with autofocus hunting, blown highlights, and interlaced noise—all prompted descriptively rather than via style tokens. The model is accessible via Cloudflare AI Gateway, which handles provider credentials and billing through Unified Billing, eliminating the need for a separate Black Forest Labs API key.