FLUX 3: Black Forest Labs Unifies Video, Audio, and Robot Action

Update — August 5, 2026: FLUX 3 Video is now generally available via API and partner tools. Native resolution moved from the July preview's 720p up to 1080p, with 2K, 4K, and open weights coming soon. It ships with a Draft mode — a fast, cheap preview render at $0.06/second that, once approved, renders the same shot at full quality while preserving subjects and composition — plus dialogue support in multiple languages and video continuation. See Black Forest Labs' launch clip below.

Every major video generation lab eventually gets asked the same question: if your model actually understands what happens next in a video, why can't it tell a robot what to do next in the real world? On July 23, 2026, Black Forest Labs answered that question directly — not with a research paper, but with a shipped robotics product already running production tests inside Audi's factories.

The Berlin/Freiburg startup — founded in 2024 by former Stability AI leaders and known for the FLUX image models that outran Stability's own follow-ups — announced FLUX 3, a single multimodal architecture jointly trained across images, video, audio, and action prediction. Alongside it: FLUX-mimic, a robotics model built with mimic robotics, running today on real hardware in a real automotive factory.

TL;DR

The actual bet: video generation and robot action are the same problem

Robin Rombach, Black Forest Labs' co-founder and CEO, framed the announcement around a claim that's easy to skim past and important to sit with: "True intelligence means perceiving the world: predicting how it will change, taking action, and learning from results." He followed it with the line that explains the whole product line in one sentence: "A model that only learns images can only generate images. But the world is not made of still frames."

That's the throughline connecting FLUX 3 Video, FLUX-mimic, and the Audi deployment. They aren't three separate announcements bundled for a press day — they're one architectural bet. If a model learns to predict "what does this scene look like one frame later" well enough across enough video, it has implicitly learned something close to "what does this physical action look like one moment later." Video prediction and action prediction, in BFL's framing, are structurally the same task wearing different outputs.

This is why FLUX 3 is trained jointly across modalities rather than as separate specialist models bolted together after the fact. Most labs ship an image model, then a video model, then — if robotics comes up at all — a separate action model trained from scratch on robot demonstration data. BFL's pitch is a single backbone that already understands physical dynamics well enough that adapting it to a new modality (audio, then action) is a fine-tuning problem, not a from-scratch training problem. That's also the argument Yann LeCun has been making about why LLM-only approaches struggle with physical-world tasks — you need models that learn world dynamics, not just token sequences. FLUX 3 is Black Forest Labs' direct answer to that gap, built from the video side rather than the language side.

FLUX 3 Video: what's actually shipping, and how it compares

FLUX 3 Video is in early access as of the announcement. The headline specs: up to 20 seconds of video in a single generation, tested at 720p resolution, with native audio generated as part of the same process rather than added in a separate pass afterward.

That "native" distinction matters more than it sounds. Most video generators that added audio did it as a bolt-on second model — generate the video, then generate matching sound for it. FLUX 3 treats audio as a first-class training signal from the start, which BFL says shows up specifically in how well the model associates sounds with physical events: footsteps landing, objects impacting surfaces, the timing of a door closing. It's also noted for capturing human facial expressions more reliably than prior generations of video models.

Supported capabilities span the range most builders would ask for:

• Text-to-video and image-to-video — animation and reference-image modes
• Video-to-video — with character and element continuity preserved across edits
• Keyframe-to-video — controlled transitions between specified start/end frames
• Multilingual dialogue generation
• Typography and animated text — a persistent weak point for prior video models
• Agentic chaining — stringing generations together into multi-shot sequences lasting several minutes, not just seconds

Black Forest Labs published its own blind-test comparison results. Treat these as a starting data point, not a verdict — they're self-reported, and BFL explicitly flagged them as preliminary and likely to shift as more testers weigh in:

A few things worth reading into that table rather than around it. The margin against Luma Ray 3.2 (93%) is the widest, while the margin against Gemini Omni Flash (52%) is nearly a coin flip — meaningfully different competitive pictures depending which rival a reader actually cares about. No third-party Elo-style arena has ranked FLUX 3 Video yet, which is the honest caveat: these are one lab's own preference tests on its own model, run before the wider public got access.

Early creative-tool testing partners named in the rollout include Canva, Burda, Magnific (formerly Freepik), Krea, and Picsart — signaling BFL's near-term go-to-market is the same creative-tooling ecosystem that already integrates FLUX image models, not a consumer app of its own.

FLUX-mimic and the Audi production floor

The more structurally interesting release is FLUX-mimic — because it's the clearest test of whether BFL's "video and action are the same problem" thesis actually holds up outside a demo.

FLUX-mimic is built jointly with mimic robotics, combining the FLUX 3 backbone with mimic's existing expertise in dexterous manipulation and production deployment — the kind of pairing world-model-to-robotics deals have become common across the industry this year. Black Forest Labs describes two technical approaches sitting side by side: native action-prediction directly within FLUX 3's shared representation, and specialized finetuning of the video backbone for task-specific robotics work when a narrower, faster model is needed.

The efficiency claim is the number worth remembering: fine-tuning FLUX-mimic for a new task takes as little as 30 minutes of robot data, compared to "30 or more hours" with prior approaches. If that holds up under independent testing, it's a roughly 60x reduction in the data-collection bottleneck that has made robot task-switching slow and expensive — closer in spirit to how quickly Xiaomi's UMI-based data pipeline is trying to compress robot learning curves, though via a different route (video-pretrained backbone versus scaled teleoperation data collection).

Elvis Nava, CTO of mimic, put the mechanism plainly: "Because FLUX-mimic is built on frontier video models that understand how the physical world behaves, it picks up new tasks in minutes, not days." That's the practical payoff of the joint-training thesis — a robot doesn't need to relearn physics from scratch for every new task if the backbone already encodes how objects move, deform, and interact from having watched enough video.

FLUX-mimic is already undergoing real production testing at Audi — not a lab demo, an actual factory floor. Per leadership at the production lab involved, quoted in coverage of the deployment: "these robots solve complex soft body manipulation work that would have been simply impossible with conventional robotics." The examples given point at exactly the kind of tasks that have historically defeated rigid industrial robotics arms: cable handling, soft or deformable materials, and other jobs where the object being manipulated doesn't hold a fixed shape. The framing from Audi's side isn't robots replacing people wholesale — it's employee assistance and production-line flexibility on the specific tasks conventional automation couldn't touch.

FLUX-mimic deep dive — Self-Flow, decoders, and factory latency

Black Forest Labs’ July 23 FLUX 3 × mimic technical post sharpens the architecture beyond the launch headline.

Why video is the hard modality. Joint training across images, video, and audio puts over 95% of compute into video prediction — contact, motion, weight, cause and effect. Audio is low-dimensional (under 0.5% of tokens in a 720p AV clip). Actions are the same shape: low-dimensional robot state tightly coupled to vision. Once the backbone models physical processes, action prediction is another view of the same reality — not a separate foundation model.

Adding actions doesn’t permanently tax video quality. In a large run, human ratings on text-to-video / image-to-video fell up to ~10% when action prediction entered the curriculum, then recovered after about 3,500 steps while keeping action capability. One backbone, two product families: content creation and physical AI.

Decoding the world model (mimic-video lineage). FLUX-mimic trains a lightweight action decoder on intermediate features from FLUX’s video prediction path — pioneered in mimic-video. Success needs both a high-quality world model (generation quality) and disentangled, usable representations. BFL’s Self-Flow work unifies generation and representation learning: lower Fréchet distances across modalities and higher robot-control success vs plain flow matching. FLUX 3 is Self-Flow scaled — tens of millions of hours of general video plus hundreds of thousands of hours of human/robot manipulation footage.

Sample efficiency & recovery. Self-Flow cut steps to a target action success rate roughly in half vs a video model without it; mimic-video reported up to ~10× sample efficiency vs classic VLAs. FLUX-mimic combines both. HN viewers flagged the window-trim reseat clip: multiple grasp recoveries without every failure mode in the demo set — behavior that has to come from world knowledge, not rote imitation. (Google’s earlier VLA timing-belt work shows this class of recovery isn’t unique to BFL — but factory soft-body deploy at Audi is the proof point that matters commercially.)

Fast enough to act. Backbone depth dominates latency. Better representations → more capability per parameter → smaller/faster backbones. BFL cites under 80 ms input→world-representation on a single RTX 5090, with mimic’s full stack (decoder, IPC, real-time chunking) reaching about 101 ms reaction — order of human visual reaction time. Factory tasks called out: kitting into trays, ECU insertion, component assembly, seals/cables.

Audi Production Lab (Christoph Schneider): partnership with mimic testing/deploying FLUX-mimic on soft-body work conventional cells couldn’t economically re-engineer for premium variant diversity — assist employees, expand flexible automation.

Frozen-backbone action decoding already beats prior VLAs in BFL’s reported trials; finetuning backbone + decoder reaches their state-of-the-art success rates. Treat numbers as lab-reported until third parties replicate — same caveat as the video preference table.

BFL also co-signed the July 24 Open Weights & American AI Leadership letter; FLUX 3 Dev open weights remain a later-2026 promise, separate from this closed robotics early access.

Rollout timeline: what's out now, what's coming, and to whom

Black Forest Labs is staging the release rather than shipping everything at once:

1. FLUX 3 Video — early access now, via API and private weights to initial partners
2. FLUX 3 Action / FLUX-mimic — early access now, to selected research and commercial robotics partners
3. FLUX 3 Image — rolling out "in the coming weeks," with improved complex-prompt handling and multilingual text rendering across a wide range of styles, aspect ratios, and resolutions
4. FLUX 3 Dev — an open-weight version of the multimodal backbone, planned for later in 2026, enabling local deployment and customization

No pricing has been disclosed for any tier as of this writing. For a company whose FLUX.1 image models became one of the most widely fine-tuned open releases in the diffusion ecosystem, the Dev tier's eventual open weights will likely matter more to the developer and researcher community than the API tiers — but that's a later-2026 story, not a today story.

Honest limitations

A few things are worth being direct about, rather than letting the announcement's momentum carry past them.

The benchmark numbers are self-reported. Every preference rate in that comparison table came from Black Forest Labs' own testing, against its own choice of comparison models and prompts, published in its own announcement. BFL itself calls the numbers preliminary and expects them to shift. There is no independent arena-style leaderboard (the kind that exists for LLMs) ranking FLUX 3 Video against competitors yet. Reasonable to treat as a credible signal from a serious lab — not as settled fact.

The tier most designers actually want isn't out yet. FLUX 3 Image — the direct successor to FLUX.1 and the tier that competes head-on with Midjourney, Qwen-Image, and other still-image generators — is explicitly "rolling out in the coming weeks." Everything shipping today is Video and Action/robotics. If your interest in FLUX 3 is still-image generation, there's nothing new to actually use yet.

Open weights are a later-2026 promise, not a current option. FLUX 3 Dev doesn't have a release date, license terms, or model size disclosed. Anyone hoping to self-host or fine-tune FLUX 3 locally — the way many builders currently do with FLUX.1 — has nothing to run today and no committed date to plan around.

Robotics production testing is still testing. Audi's soft-body manipulation deployment is real and running, but it's described as production testing, not a completed rollout across a full assembly line. The 30-minutes-of-data claim for new-task fine-tuning is a strong number worth watching for independent replication as more partners get access to FLUX-mimic.

The bottom line

FLUX 3 is Black Forest Labs' bet that video generation and robot action prediction are close to the same underlying problem — a model that predicts the next frame of a video is exercising much of the same world-understanding a robot needs to predict the next physical action. FLUX 3 Video (20-second, 720p, native-audio clips) and FLUX-mimic (already testing at Audi, learning new tasks from as little as 30 minutes of data) are the first proof points for that thesis, and they're genuinely shipping today, not vaporware. The parts to watch before drawing firm conclusions: BFL's comparison numbers are self-reported and preliminary, FLUX 3 Image hasn't launched yet, and open weights are a promise for later in 2026, not something available now. Whether the unified-architecture bet actually outperforms specialist video models and specialist robot-learning stacks over the next few months is the real test — this announcement is the starting line, not the result.

Related on explainx.ai

• Open Weights & American AI Leadership letter — BFL is a signatory
• How diffusion image generation actually works
• Moebius: 0.2B parameters matching FLUX.1-Fill-Dev on inpainting
• World Labs acquires SceniX for robotics simulation
• Xiaomi Robotics: 100K hours, UMI, and VLA models
• Xiaomi Robotics U0: world foundation model for embodied synthesis
• 1X Neo's 25-DOF hands and the Physical API
• Yann LeCun on LLMs, physical agents, and the Moravec paradox
• Figure AI: robots outnumber humans milestone

Sources: Black Forest Labs' FLUX 3 announcement and FLUX 3 × mimic technical blog (July 23, 2026), Hacker News discussion, press coverage, and mimic / Audi public statements on factory deployment.

This post reflects Black Forest Labs' announcement and public statements as of July 24, 2026. Rollout timelines for FLUX 3 Image and FLUX 3 Dev are forward-looking and may shift; pricing was not disclosed for any tier at publication time. Self-Flow / latency / success-rate figures are vendor-reported.