Black Forest Labs Unveils FLUX 3 AI: Ditches Stills for Video-With Robot Hands

Black Forest Labs has been released FLUX 3 On Thursday, and for the first time, the company’s model produces a video instead of still images. The German AI lab, known for its FLUX line of image generators, trained a new system for images, video, and audio simultaneously, within a single shared system.

That’s what is known as multimodality: one model learning several types of knowledge together instead of separate tools built side by side.

The video part is the title. FLUX 3 creates visuals of up to 20 seconds, audio that is created alongside the image and is connected to what is happening on the screen – dialogue, sounds, surround sound. In early testing, reviewers preferred the FLUX 3 output over the Runway Gen-4.5 in 77% of head-to-head comparisons and the Luma Ray 3.2 in 93%. It seems to be slightly better than Gemini Omni and Seedance, beating those models in 52% of reviews.

Of course, it’s a preference test, not a standard guide: reviewers simply look at two videos and choose the one that looks and sounds better, and BFL calculates how many times the FLUX 3 wins.

After all, the model seems to be very good at the rest of the pictures, following its legacy. BFL shared a number of images, and the FLUX 3 appears to be very versatile and capable of creating a variety of styles beyond just a snapshot.

BFL lists this as an additional tool. “A model that only learns images only creates images,” said co-founder and CEO Robin Rombach. The company’s bet is that learning to predict video also means learning the underlying physics — weight, contact, time — which is what machines need to navigate the world.

That bet has a name: Image of FLUX. Built by Zurich Robotics, it takes the FLUX 3 video prediction engine and adds a lightweight “decoder” – a small chip that interprets the internal logic of how things move in real-time robot motion. Car maker Audi is already testing tasks like installing flexible door seals, a task that conventional machines struggle to handle.

“Audi represents a manufacturing partner for our FLUX-mimic,” said co-founder Stephan-Daniel Gravert. Audi’s Christoph Schneider said robots now “solve complex, physically-disruptive tasks” that older machines couldn’t handle. BFL claims that the entire system performs in about 101 milliseconds, close to human performance.

FLUX’s ascension did not happen in vain. Founded in August 2024 by former researchers who helped develop the original Stable Diffusion models at Stability AI, Black Forest Labs developed a Flux model that beats MidJourney and continues to disrupt Stability. Fixed Mixing 3.

The open source versions of Flux Dev and Schnell took the title of “best open source image editor” which AI experts expect Stable Diffusion 3.5, Stability’s do-over, to take back.

It didn’t. The FLUX 1.1 Pro went above and beyond Forum for Artificial Analysis images that October. That wasn’t open source, though.

BFL released FLUX.2 in November 2025 but it was not popular. An open crown with the original Flux has existed until Alibaba Z-Image Turbo removed it at the end of 2025, comparing its quality to low-cost consumer cards. “This is what SD3 was meant to be,” one CivitAI user wrote at the time.

FLUX 3 is the return of BFL, and it has not been opened yet. Video and Actions are starting to become available now through APIs and choosing partners, modeling robotics among them, and creating images will follow “in the coming weeks,” on BFL. The open Dev version, the only part of BFL that is planned to be released for local use, is not expected to happen until later in 2026.