Black Forest Labs Unveils FLUX 3 — a Multimodal Model for Images, Video and Sound

Black Forest Labs has announced the release of FLUX 3 in early access. Unlike its earlier versions, which were geared toward image generation, the newcomer is positioned as a multimodal foundation model that, within a single architecture, learns simultaneously from images, video and sound.

The authors' reasoning runs like this: no single modality describes reality in full. An image captures spatial structure at a specific moment, video adds the dimension of time and the physics of motion, while sound reveals cause-and-effect relationships that are inaccessible to sight. By training on everything at once, the model exploits their mutual constraints: sound must match the impact, motion must obey mass, the future must follow from the past. That is how disparate signals turn into evidence about one and the same reality.

What FLUX 3 Can Do

At its core lies the Self-Flow approach, which, according to the developers, is more efficient than classic flow matching and makes it possible to combine generation and understanding in a single architecture. The model is capable of creating video with native sound up to 20 seconds long in a single generation — from text, from a starting frame, from a reference clip or from keyframes. It claims multilingual dialogue, a broad range of styles — from amateur footage to animation and cinema — and the ability to stitch individual clips into multi-scene sequences while preserving characters.

Special mention is made of handling text inside images and video, as well as generating images in various styles, aspect ratios and resolutions.

Cautious Numbers

The company is candid in calling its assessments preliminary: the model and the tooling around it are still being refined. In early comparisons of 10-second, 720p clips, FLUX 3 was preferred over Luma Ray 3.2 in 93% of cases, over Runway Gen-4.5 in 77%, over Grok Imagine Video in 69%, and over Kling v3 Pro and a number of other competitors in the 52–60% range.

From Images to Robots

The authors' ambitions extend beyond content generation. FLUX 3 can predict actions: the video backbone is used as a foundation from which specialized control models are fine-tuned. One of the first partners was mimic robotics — together they built the video-action model FLUX-mimic for dexterous manipulation, which is being tested on production tasks at Audi. The company's thesis is simple: physical AI and content creation run on a common foundation.

Over the coming weeks and months, Black Forest Labs promises to roll out access to its video, audio and graphics capabilities in stages through an API and private weights, as well as to release an open multimodal model, FLUX 3 Dev. The ultimate goal is to unite perception, action and language in a single model. The company is hiring in Germany and the United States.

Source: Hacker News