Black Forest Labs — the Freiburg team that built Stable Diffusion and FLUX.1 — shipped FLUX 3 on July 23. It generates synchronized video and audio from a text prompt. It edits images. And it currently controls robot arms on Audi production lines. All from one model, one training run, one set of weights. This is not a video generator with image editing bolted on afterward. It is a fundamentally different thesis about what a generative model is.
One Architecture, Not Three Stitched Together #
The key decision BFL made with FLUX 3 is training across all modalities simultaneously, not sequentially. The underlying approach is called Self-Flow. The reasoning: images, video, audio, and physical actions are all “projections” of the same underlying physical reality. Train on all four at once and the model builds a richer representation of how the world actually works — how objects hold shape, how things move, how events sound, how forces cause effects.
More than 95% of training compute went to video prediction. That is where the hardest world-modeling happens. The implication for developers building on top of this: one API, one fine-tuning run, and learned representations that transfer across image editing, video generation, audio, and robot control. Fewer vendors, less stitching.
Robots at Audi, Running Today #
The most concrete demonstration of BFL’s thesis is FLUX-mimic, developed with mimic robotics. FLUX-mimic layers a lightweight action decoder on top of FLUX 3’s video prediction path. The logic: the same world-model that makes video look physically realistic — contact dynamics, object weight, cause and effect — already encodes much of what a robot arm needs to know to act.
FLUX-mimic is running dexterous manipulation tasks on Audi production lines right now. Fine-tuning the model for a new manipulation task takes about 30 minutes of robot demonstration data. Prior approaches required 30-plus hours — a 60x reduction in training data requirements. That is a meaningful validation of the core architectural bet: the world-model learned during video training transfers to physical control.
What Is Actually Available Today #
FLUX 3 launched with gated access, and the distinction matters:
FLUX 3 Video: Early access, by application atbfl.ai. Generates clips up to 20 seconds with native audio in one pass.FLUX 3 Action (FLUX-mimic): Early access to selected robotics and commercial partners only.FLUX 3 Image: Not yet released. Coming in the next few weeks.FLUX 3 Dev: An open-weight version planned for later in 2026.Pricing: Not yet announced.
If you are a current FLUX.1 or FLUX.2 user waiting to upgrade your image pipeline, you are not there yet. The image component — what most people use FLUX for — has not shipped. FLUX 3 Video is real, but gated. The practical window for most developers opens when FLUX 3 Image arrives.
On the Benchmarks #
BFL published preference benchmarks showing FLUX 3 Video preferred over Seedance 2.0 and Gemini Omni Flash in 52% of head-to-head comparisons, and over Kling v3 Pro in 60%. A few things to note: these are BFL’s own preliminary numbers on a pre-release model candidate. A 52% win rate is within statistical noise — effectively tied with Seedance and Gemini Omni Flash. Seedance 2.5 and Gemini Omni Flash are already production-ready and widely accessible. FLUX 3 is not, yet.
The benchmarks are directional, not definitive. The interesting claim is not “it beats Kling” — it is the architecture thesis: one model that learns physics well enough to generate convincing video and guide a robot arm. Latent Space’s technical breakdown and VentureBeat’s coverage both flag the limited release context.
What to Watch #
Two releases matter for developers not in robotics. First, FLUX 3 Image in the coming weeks — that is when the creative tooling and developer ecosystem built on FLUX.1 can actually start evaluating an upgrade. Second, FLUX 3 Dev later in 2026, the open-weight release. BFL has a strong track record: FLUX.1 models accumulated roughly 500 million downloads combined across the FLUX and Stable Diffusion era.
The broader claim — that a single architecture trained on the physical world can serve creative professionals and factory robots equally well — is either BFL’s most ambitious idea or its most important one. The Audi deployment suggests it might be both.