Black Forest Labs' FLUX 3 Generates Image, Video, Audio, and Robot Actions From One Model

Black Forest Labs' FLUX 3 Generates Image, Video, Audio, and Robot Actions From One Model

6 min readJuly 25, 2026

Quick verdict

Black Forest Labs put image, video, audio, and action prediction into one model called FLUX 3. That last one, action, is the surprising part: the same architecture that makes a picture can also predict what a robot should do next. A spinoff called FLUX-mimic already runs robot control on a single on-prem GPU and is being trialed with Audi. For most people the near-term win is simpler than robotics, though. One model that handles image and video and audio means one thing to learn and, eventually, one bill instead of three.

What actually shipped

FLUX 3 is a single multimodal model that generates across four modalities: image, video, audio, and action. Earlier FLUX releases were image models, strong ones, but still one job each. Version 3 treats these outputs as the same kind of problem and learns them together, extending the lab's earlier Self-Flow research. The pitch from the team is that a model with a better internal picture of how the world moves makes better video, and that same understanding transfers to controlling things in the physical world.

The robotics side is where that claim gets concrete. FLUX-mimic is a video-action model built on FLUX 3, trained on robot and wearable data for general-purpose dexterity. It runs on a single on-prem GPU rather than a datacenter, and Black Forest Labs says it is already being tested with Audi. Alongside it, a separate system called GEN-1 now adapts when the robot's "hand" changes mid-task, which points at policies that generalize across different hardware instead of being rebuilt for each gripper.

Why one model beats three

If you make things with AI today, you probably juggle tools. One service for images, another for video, a third for voiceover, each with its own login, its own credits, and its own quirks. A unified model does not automatically beat a specialist on every axis, and the honest read is that dedicated image or video models may still edge it on specific quality benchmarks. What it changes is friction. You describe what you want once, in one place, and the pieces come back sharing a visual and temporal style instead of looking stitched together.

The style consistency is the underrated benefit. When your image tool and your video tool are separate models trained separately, a still frame and a clip of the same scene rarely match. A model that learned them jointly keeps them coherent, which matters a lot for anyone producing a series of assets rather than a single hero shot. If you are weighing generators for that kind of work, our roundups of the best AI image generators and best AI video generators are worth reading next to this launch.

Video: FLUX 3 video generation in action

A look at where FLUX 3 lands on quality and what the video side can do right now.

The part worth being skeptical about

"One model for everything" is a strong marketing line, and strong marketing lines deserve a second look. Unified models tend to be generalists, and a generalist rarely tops a well-tuned specialist on that specialist's home turf. The action-prediction claim is the most interesting and the least proven: a trial with one carmaker is a promising start, not a track record. Sample efficiency and world-modeling gains are real research directions, but "the video model transfers to robot control" is the kind of clean story that gets messy once it meets a factory floor.

There is also the open question the whole industry keeps circling. If a single model absorbs image, video, and audio, the tools built around separate models start to look like temporary scaffolding. That is good for creators who want less to manage and awkward for the companies whose whole product is one modality done well.

Why it matters

For two years the pattern was a new specialist model every month, each best at one thing. FLUX 3 bets the other way, that the next advantage comes from collapsing those specialties into one system that shares what it learns across them. If the bet pays off, the practical result for anyone making content is fewer subscriptions, less tool-switching, and assets that actually match each other. If it does not, we get another capable image model with a big roadmap. Either way, "buy one model, get four modalities" is now the pitch every competitor has to answer.

FAQ

What makes FLUX 3 different from earlier FLUX models?

Earlier FLUX releases were image generators. FLUX 3 is a single model that handles image, video, audio, and action prediction together, trained so that understanding gained in one modality helps the others.

What is FLUX-mimic?

It is a video-action model built on FLUX 3 for robotics, trained on robot and wearable data. It runs on a single on-prem GPU and is being trialed with Audi, using the model's world understanding to control physical movement.

Does a unified model beat dedicated image or video tools?

Not necessarily on raw quality. Specialists can still win on their own benchmarks. The advantage of a unified model is consistency across outputs and far less friction from juggling separate tools and subscriptions.

Can I use one subscription for all my AI generation?

That is the direction unified models push toward. Until it fully arrives, the practical move is a workspace that pools multiple models, which we cover in one subscription for all AI models.

Sources

Further reading

Try all the models mentioned in this article

Admix gives you GPT-5, Claude, Gemini, and 350+ AI models in one app. Compare responses side by side. Free to start.

Start free on Admix

Related articles