Black Forest Labs' FLUX 3 Beats Runway Gen-4.5 and Now Runs Factory Robots

FLUX 3 is a single foundation model jointly trained across video, audio, and images from scratch, with the same backbone now running robots on an Audi factory floor. Black Forest Labs, the team behind the original FLUX image generation models, is betting that content creation and physical AI are the same problem.

One backbone, every modality

FLUX 3 jointly learns from images, video, and audio within a unified architecture. The core idea is that no single modality provides a complete picture of reality: each is a projection of the same underlying world, captured by different sensors. Images give you spatial structure. Video restores time and reveals physics. Audio exposes causal relationships that vision alone misses. The model reconciles all three at once, which forces it to build a coherent internal model of how the world actually works.

FLUX 3 builds on Self-Flow, BFL's approach for aligning multimodal generation and understanding within the same architecture. Self-Flow unifies representation learning and generation in one pass, without relying on frozen external encoders like CLIP or DINO. BFL then scaled compute and data substantially to train across all three modalities simultaneously. The result is a model that gets better at generating video because it understands the world better, not because it memorized more frames.

What it can do right now

FLUX 3 Video with optional native audio and FLUX 3 Action are available for early access. FLUX 3 Image will roll out in the coming weeks. No pricing has been disclosed as of this writing. You can request early access via the official form.

The video capability is the most developed. FLUX 3 generates videos up to 20 seconds long, with audio, in a single pass. Supported generation modes include:

• Text-to-video and image-to-video generation
• Video-to-video: carry a character or element from one scene into a new context
• Keyframe-to-video for controlled transitions between defined moments
• Generative audio continuation from an input video clip
• Multilingual dialogue generation
• Agentic chaining of clips into longer multi-shot sequences
• A wide range of visual styles, from camcorder footage to animation and cinematics

Audio is generated jointly with the video frames, not added in post-processing.

How it compares to other models

BFL's preliminary evaluation used 10-second text-to-video clips at 720p with audio. In head-to-head preference comparisons, FLUX 3 was preferred over:

• Runway Gen-4.5 in 77% of comparisons
• Grok Imagine Video in 69%
• Kling v3 Pro in 60%
• Happy Horse v1 in 59%
• Seedance 2.0 and Gemini Omni Flash in 52%
• Luma Ray 3.2 in 93%

These numbers are self-reported and preliminary. FLUX 3 Image has not launched yet, and open weights are planned for later in 2026. Early evaluators highlight the model's strength in capturing human facial expressions, synchronizing sound with physical events, and handling multilingual output.

The robot connection

The most unexpected part of this launch is that the same backbone already runs physical robots in production. mimic robotics was among the first partners to gain early access to FLUX 3. Together they built FLUX-mimic: a video-action model combining the FLUX 3 backbone with mimic's expertise in robot learning for dexterous manipulation and production deployment.

The connection between video generation and robotics is more direct than it sounds. Video prediction accounts for over 95% of FLUX 3's total training compute. To generate realistic video, the model must learn contact, motion, weight, and cause-and-effect relationships accurately, because errors in any of them are immediately visible. A model that has internalized those physical laws is already most of the way to predicting robot actions.

Robot actions follow the same structural shape as audio: a low-dimensional signal tightly coupled to visual observations. After learning the physical processes behind video and audio, action prediction becomes one more view of the same reality the model already represents.

How FLUX-mimic works

FLUX-mimic trains a lightweight action decoder on top of intermediate features from FLUX 3's video prediction path. The backbone does not need to be retrained for robotics; its internal representations already encode the relevant physics. The decoder reads those representations and outputs motor commands.

During a large-scale training run, BFL added action prediction to the curriculum and observed that human ratings on text-to-video and image-to-video tasks initially dropped by up to 10% as the model absorbed the new modality. After 3,500 training steps, video quality had fully recovered while action prediction was now active. The video model gave up nothing permanently to become a robot controller.

Latency matters in real deployment. The FLUX-mimic backbone can process input to world representation in under 80ms on a single RTX 5090, which is on the same order as human visual reaction time. End-to-end robot reaction time is 101ms.

Already on the factory floor

mimic deployed FLUX-mimic across several production and logistics tasks at Audi: kitting parts into structured trays, inserting electronic control units into tight-fitting fixtures, assembling components, and handling soft materials like seals and cables that conventional automation cannot reliably manipulate.

Audi confirmed the deployment, noting that FLUX-mimic solved soft-body manipulation tasks that stayed manual for decades because the variant diversity of premium production made reprogramming robot cells too costly for each case.

When physical knowledge is already encoded in the model's representations, teaching a robot a new task means mapping that task onto existing understanding rather than rebuilding world knowledge from scratch. The mimic-video paper reports up to 10x sample efficiency for video-action models over standard vision-language-action models.

What's still coming

The full rollout, each tier gated behind an early access phase:

1. FLUX 3 Video — available now in early access, API and private weights
2. FLUX 3 Action / FLUX-mimic — early access for selected research and commercial robotics partners
3. FLUX 3 Image — rolling out in the coming weeks
4. FLUX 3 Dev — open-weight multimodal backbone, planned for later this year

Open weights will enable secure, low-latency local deployment for applications like robotic control systems and let teams fine-tune FLUX 3 on their own data and workflows.

Why the framing matters

The field has largely treated video generation and robotics as separate tracks. FLUX 3 challenges that assumption directly. A model trained to predict the next frame of a video exercises much of the same world-understanding a robot needs to predict its next physical action. If that holds at scale, compute invested in better video models also builds the substrate for physical AI, using the same scaling laws, the same data, and the same architecture. The FLUX-mimic deployment at Audi is the first real-world evidence for that claim. Whether it generalizes beyond controlled factory tasks is the next question to answer.