Black Forest Labs Introduces FLUX 3 AI: Ditch Stills for Video and Comes with a Robotic Hand

Black Forest Labs released FLUX 3 on Thursday. For the first time, the company’s flagship model produces video rather than just still images. German AI Lab, known for its FLUX series of image generators, trained the new system with images, video and audio simultaneously within one shared system.

This is what is known as multimodality. That is, a single model learns multiple types of information together instead of separate tools fixed side by side.

The video side has a heading function. FLUX 3 produces clips up to 20 seconds long. Sound is generated alongside the image and synchronized with what’s happening on screen (dialogue, sound effects, ambient noise). In early evaluations, human reviewers preferred the FLUX 3’s output over Runway Gen-4.5 in 77% of head-to-head comparisons and the FLUX 3’s output over Luma Ray 3.2 in 93% of head-to-head comparisons. It appears to be slightly better than Gemini Omni and Seedance, beating these models with 52% of ratings.

Of course, this is a preference test, not a fixed-score rubric. The evaluator simply looks at the two clips, chooses the one that looks and sounds more convincing, and BFL counts how often FLUX 3 wins.

Other than that, this model, in keeping with its tradition, seems to be very capable in stills as well. BFL shared a few images and it looks like FLUX 3 is very versatile and can produce a wide range of styles beyond photorealism.

BFL positions it as more than a content tool. “Models that only learn images can only generate images,” said co-founder and CEO Robin Rombach. The company’s bet is that learning to predict video also means learning its underlying physics (weight, contact, timing), which is exactly what machines need to navigate the physical world.

The bet has the name “FLUX-mimic.” Built using Zurich-based Mimic Robotics, it takes FLUX 3’s video prediction engine and adds a lightweight “decoder.” This is a small add-on component that translates the model’s internal sense of how things move into actual robot movement. Carmaker Audi is already testing the technology in tasks such as installing flexible door seals, which are difficult to handle with traditional automation.

“Audi represents the kind of manufacturing partner we built FLUX-mimic with,” said Stephan-Daniel Gravert, co-founder of mimic. Audi’s Christoph Schneider said the robot “can solve complex soft-body manipulation tasks” that older machines could not handle. BFL says the entire system responds in about 101 milliseconds, which is close to the human visual reflex.

FLUX’s rise didn’t happen in a vacuum. Founded in August 2024 by veteran researchers who helped build the original Stable Diffusion model with Stability AI, Black Forest Labs has announced a Flux model that beats MidJourney and outperforms Stability’s own dominant Stable Diffusion 3.

The open-source Flux Dev and Schnell models earned the title of “best open-source image generator” that AI artists hoped Stable Diffusion 3.5, a redo of Stability, would eventually regain.

That was never the case. FLUX 1.1 Pro became the leader in artificial analysis imaging in October of the same year. However, it wasn’t open source.

BFL released FLUX.2 in November 2025, but it wasn’t very popular. The open source throne held by the original Flux lasted until it was usurped by Alibaba’s Z-Image Turbo in late 2025, rivaling the quality of low-end consumer graphics cards. At the time, one CivitAI user wrote, “This is what SD3 is meant to be.”

FLUX 3 is a revival of BFL, but it hasn’t fully opened yet. BFL says the videos and actions are currently in early access through APIs and select partners, mimicking robotics, and image generation will occur “in the coming weeks.” The Dev version of Openweight, the only tier BFL plans to release for local use, is expected to be released until late 2026.