FLUX 3 drops in early access with unified video and native audio

Most video generation models feel like Frankenstein projects. One model dreams up the visuals, a second one guesses the sound effects, and a third tries to keep the character's face from melting between shots. The audio rarely syncs with physical impacts, and the physics often look unconvincing.

Black Forest Labs is tossing out the patchwork approach. The team behind the widely popular FLUX image models has unveiled FLUX 3, a unified multimodal foundation model now rolling out in Early Access. Instead of stitching together separate systems, FLUX 3 trains across video, images, and audio simultaneously inside a single architecture.

Native audio and 20-second dynamic clips

FLUX 3 runs on Self-Flow, an architecture developed by the lab to align multimodal generation and understanding far more efficiently than traditional Flow Matching. Their reasoning is simple: in the real world, visual motion and acoustic feedback are tied to the exact same physical laws. Separating them during training only teaches the model incomplete physics.

That foundation enables FLUX 3 to spit out up to 20 seconds of video with native, physically synchronized audio in a single generation. It covers text-to-video, image-to-video with starting frames or reference guides, and video-to-video that carries consistent characters into entirely new settings.

It also tackles multi-shot storytelling headaches. The model supports keyframe-to-video generation for controlled transitions, agentic chaining to string multiple shots into longer narratives, multilingual dialogue, and crisp animated typography.

In early blind evaluations on 10-second 720p clips with audio, users overwhelmingly favored FLUX 3. It outperformed Runway Gen-4.5 in 77% of comparisons and crushed Luma Ray 3.2 at 93%. It also beat Grok Imagine Video 69% of the time, Kling v3 Pro at 60%, and edged past Gemini Omni Flash and Seedance 2.0 at 52%.

From film sets to Audi's factory floor

Black Forest Labs isn't stopping at media generation. They believe that a model that truly grasps physical dynamics can also tell a robotic arm how to move.

FLUX 3 includes native action prediction capabilities. The team teamed up with mimic robotics to build FLUX-mimic, a specialized video-action model that adapts the FLUX 3 backbone for complex robotic manipulation. It is already being put through its paces on real-world manufacturing tasks at Audi.

As for availability, FLUX 3 Video is open now in Early Access via APIs and private weights. FLUX 3 Image will follow in the coming weeks, and the open-weight FLUX 3 Dev model is slated for release over the coming months.

Why it matters

For filmmakers, game designers, and indie creators, FLUX 3 takes a massive bite out of production friction. Chaining three different tools for video, Foley, and voice is tedious; getting 20 seconds of coherent, character-consistent footage with native audio in one prompt cuts workflow complexity in half.

For the AI landscape, it shifts the battle lines. Pure visual realism is no longer enough. Competitors are now pressured to deliver native audiovisual coherence and deep physical grounding under a single hood.

We think pivoting a creative generation backbone into industrial robotics is a brilliant commercial hedge. While consumer media tools fight for shrinking margins, robotics and physical AI command serious enterprise budgets.

The real test will be compute. Generating synchronized video and audio alongside deep physical modeling demands staggering hardware power. When the open-weight FLUX 3 Dev finally arrives, whether local creators can actually run it without enterprise-grade GPU clusters will decide if it dominates the developer community.