Black Forest Labs launches FLUX 3 capable of generating images and 20-second video with audio — but in limited release to start

Black Forest Labs (BFL) is increasing its FLUX household past picture technology with today's launch of FLUX 3, a multimodal frontier mannequin educated to know and generate pictures, or mixed audio/video clips as much as 20 seconds from a single immediate — and to increase the identical underlying structure to robotic imaginative and prescient and actions.

The Freiburg, Germany-based AI lab says FLUX 3 is collectively educated throughout these modalities moderately than assembling separate picture, video and audio fashions behind a standard interface.

That distinction is central to the corporate's pitch: BFL needs enterprises to consider inventive technology, simulation, laptop use and robotics as related purposes of a single functionality it calls visible intelligence — fashions, within the firm's phrases, "that may understand, predict, and act throughout bodily and digital environments." This launch marks BFL's first public video technology mannequin.

FLUX 3 will likely be provided by means of 4 product strains: FLUX 3 Video, FLUX 3 Picture, FLUX 3 Motion and the upcoming, open supply FLUX 3 Dev. FLUX 3 Video, with optionally available native audio technology, and FLUX 3 Motion are getting into a gated "Early Access" program now, to which anybody can apply, however which BFL should approve.

There’s presently no public entry by means of BFL's utility programming interface (API) or these of companions but, however the firm says FLUX 3 Picture will roll out within the coming weeks, adopted by basic availability. The restricted preliminary availability rollout echoes the discharge methods of latest fashions from different frontier labs within the U.S. currently, together with Anthropic and OpenAI, although these had been ostensibly for safety issues and resulting from authorities request.

What the corporate has not introduced is pricing, manufacturing service-level commitments, analysis methodology, pattern sizes, rater counts or any image-model benchmarks in any respect. Enterprise patrons due to this fact can’t but calculate whole value of possession or independently reproduce the video comparisons.

One other massive notable omission: FLUX 3 is not launching with downloadable weights presently, nor an open supply license. BFL says quicker and open-weight variations will arrive later this yr, and its technical weblog names FLUX 3 Dev as "open-weight entry to a multimodal spine, for content material creation (video, audio and picture) and motion prediction" — a significantly broader dedication than any earlier FLUX Dev launch, all of which coated pictures solely.

Nevertheless it arrives final within the sequence. Builders accustomed to receiving a regionally deployable FLUX variant alongside — or quickly after — a serious mannequin announcement must wait. That delay doesn’t negate the corporate's dedication, however it’s disappointing given the function open weights have performed in FLUX's adoption up to now.

Flux 3 is rated increased than the competitors, however lacking pricing and benchmarking particulars could forestall speedy enterprise adoption

BFL has printed a number of benchmark comparisons, however they're certified as preliminary — with full benchmark outcomes and methodology to be printed later throughout broader basic availability.

In early head-to-head desire testing on 10-second, 720p text-to-video clips with audio, the corporate says FLUX 3 was most well-liked over Luma Ray 3.2 in 93% of comparisons, Runway Gen-4.5 in 77%, Grok Think about Video in 69%, Kling v3 Professional in 60%, Blissful Horse v1 in 59%, Blissful Horse 1.1 in 57%, and each Seedance 2.0 and Google's Gemini Omni Flash in 52%.

One caveat travels with each a kind of figures, and it comes from BFL itself. The chart carrying the outcomes is labeled a "preliminary analysis of an early FLUX 3 candidate" — which means the numbers describe a pre-release checkpoint moderately than the mannequin now getting into early entry. That cuts each methods: the delivery mannequin could carry out higher, however nothing printed at the moment measures what clients will really name.

Luma Ray 3.2 and Runway Gen-4.5, the place FLUX 3 posted 93% and 77%, are the softest comparisons on the record — established merchandise, however not the fashions at present setting the tempo in unbiased video rankings. These are actual wins, and they’re those least more likely to change an enterprise shortlist.

Seedance 2.0, at 52%, is a statistical coin flip in opposition to a mannequin most Western enterprises can’t at present procure. ByteDance indefinitely postponed Seedance 2.0's worldwide rollout after Netflix, Warner Bros., Disney, Paramount and Sony despatched authorized threats over alleged systematic copyright infringement, and that suspension stays in place. Tying a frozen product is neither a powerful declare nor a harmful one.

Gemini Omni Flash, additionally at 52%, issues far more. Omni is the closest large-platform analogue to what FLUX 3 is trying — multimodal enter, video and audio-aware creation, conversational modifying — and by BFL's personal measurement, the 2 are indistinguishable on 10-second text-to-video high quality.

Google's benefit in that matchup is that Omni is usually out there through Google's Gemini API for $0.10 per second of generated 720p video, or a 10-second clip for round.

One regional wrinkle issues for a German firm's house market. Enhancing uploaded video is unavailable to Omni Flash customers within the European Financial Space, Switzerland and the UK, although modifying video the mannequin itself generated is permitted. A European enterprise that desires to run its present footage by means of a generative modifying cross can’t at present accomplish that on Omni Flash.

Right here's a tough information for enterprises contemplating which video fashions to depend upon:

One structure for media technology and bodily motion

FLUX 3 builds on Self-Flow, BFL's technique for aligning multimodal understanding and technology inside one structure, publicized again in March 2026.

The corporate says it considerably scaled up compute and knowledge to coach throughout video, pictures and audio concurrently, and that testing confirmed video technology and motion prediction don’t require separate foundations — the identical structure may very well be prolonged to motion prediction with out sacrificing what it discovered from video.

"We place imaginative and prescient on the heart of our method as a result of it’s the most signal-rich medium of the bodily world. Pictures convey construction, pictures and video educate spatial relationships, video teaches dynamics, and actions reveal causal relationships. However imaginative and prescient alone is just not the entire image," mentioned Robin Rombach, co-founder and CEO of BFL, in a pre-release assertion supplied to VentureBeat. "True intelligence means perceiving the world: predicting the way it will change, taking motion, and studying from the outcomes. Joint coaching inside one unified structure is what is going to get us there, as a result of every coaching modality strengthens the others. Audio conveys timing, prosody, and bodily occasions that elude imaginative and prescient. Language conveys targets, abstractions, and directions that pixels can’t simply specific."

He put the case extra bluntly elsewhere within the announcement: "You possibly can't cheat actuality. A mannequin that solely learns pictures can solely generate pictures. However the world is just not manufactured from nonetheless frames. It strikes, sounds, modifications, and responds."

BFL says FLUX 3 targets inventive tooling, media, design, e-commerce and bodily AI, supporting video technology with synchronized audio, exact picture modifying, product and materials consistency throughout movement, multilingual technology and robotic motion prediction. It’s already being examined by Canva, Burda, Magnific (previously Freepik), Krea and Picsart.

For inventive software program firms, the attraction is consolidation. A single basis may doubtlessly help storyboarding, picture modifying, product rendering, video variation and localization with out repeatedly translating belongings and directions between disconnected fashions.

For robotics groups, the potential worth is knowledge effectivity. Fashions that already encode movement, object habits and bodily change might have much less task-specific robotic coaching than methods ranging from uncooked demonstrations.

What FLUX 3 Video can really do

The video tier is probably the most concretely specified a part of the launch, and it settles a query that had been circulating as rumor: FLUX 3 generates clips of as much as 20 seconds with audio in a single technology.

Each video output comes with native audio. For comparability, HappyHorse 1.0 tops out at 15 seconds of 1080p with synchronized audio — although BFL has not said what decision its 20-second clips run at, and its printed evaluations had been performed at 720p. Nonetheless, a 20-second lengthy clip from a single immediate is among the many longest but achieved, matching OpenAI's discontinued Sora model.

The potential record BFL printed covers:

• Textual content-to-video technology.

• Picture-to-video technology, both animating from a beginning body or utilizing pictures as visible references.

• Video-to-video technology from a reference clip, carrying parts resembling a selected character into a brand new scene or context.

• Generative video-audio continuation from present video and audio enter.

• Keyframe-to-video technology for managed transitions between outlined moments.

• Multilingual dialogue.

• A broad vary of visible types and side ratios, from candid camcorder footage to animation and cinematics.

• Typography technology and animated design.

• Agentic chaining of particular person clips into longer, multi-shot sequences.

That final merchandise is the one enterprise video groups ought to take a look at hardest. BFL claims the capabilities mix to provide sequences lasting a number of minutes, with visible references protecting characters constant throughout scenes. If that holds up beneath manufacturing circumstances, it addresses the constraint that has stored generative video out of most industrial pipelines: not clip high quality, however continuity throughout pictures.

Additionally it is the potential the place competitors is most direct. HappyHorse 1.1's headline improve is R2V, or Reference-to-Video, which accepts a number of character reference pictures to carry identification steady throughout generated footage — the identical downside, approached on the enter layer moderately than by means of agentic clip chaining. Alibaba additionally claims zero-drift lip sync and has particularly focused the artifacts that mark industrial AI video as artificial, together with facial oiliness and over-sharpening. Character consistency is the place this class is being contested, and each firms realize it.

BFL says FLUX 3 Video is already significantly sturdy at human facial expressions, associating sounds with bodily occasions, and multilingual output. On the picture facet, the corporate says preliminary evaluations performed throughout midtraining present important enchancment over earlier FLUX variations in complicated immediate dealing with and textual content technology, together with high-accuracy textual content in a number of languages. It printed no picture benchmarks or win charges.

FLUX-mimic checks whether or not video fashions can turn out to be robotic fashions

BFL is making use of its unified-architecture thesis by means of FLUX-mimic, a video-action mannequin constructed on FLUX 3 and developed with Swiss agency Mimic Robotics, one of many first companions to obtain early entry.

The technical weblog describes two distinct routes to motion prediction: integrating native motion prediction instantly into FLUX 3, scaling up the preliminary Self-Move work; and utilizing the pretrained video spine as a dynamics-aware basis from which specialised motion fashions could be finetuned with restricted task-specific knowledge. FLUX-mimic is the second route — the FLUX 3 spine mixed with mimic's robot-learning and production-deployment experience in dexterous manipulation.

FLUX-mimic is designed for general-purpose robotic manipulation: serving to robots perceive a visible scene, predict the implications of an motion, and adapt to new duties with far much less task-specific knowledge.

BFL and Mimic Robotics say that relying on job issue, the mannequin could be finetuned for a selected manipulation job with as little as half-hour of robotic knowledge, the place prior approaches have required 30 or extra hours.

"The toughest a part of robotics is knowledge," mentioned Elvis Nava, CTO of Mimic Robotics, in an announcement supplied to VentureBeat. "Each new job usually means hours of a robotic repeating itself. As a result of FLUX-mimic is constructed on high of frontier video fashions that already perceive how the bodily world behaves, it picks up a brand new job in minutes, not days. This fashion, we will leapfrog the present cutting-edge in robotic studying."

BFL argues {that a} mannequin educated solely on pictures can’t perceive a world that "strikes, sounds, modifications, and responds," and that bodily understanding is what produces convincing generated footage. Google makes a virtually an identical declare for Gemini Omni.

Its developer documentation cites "world information" that mixes "an understanding of physics" with Gemini's grasp of historical past, science and cultural context. Its advertising is blunter nonetheless: "Most AI fashions simply predict the following pixel to construct a story or a picture. Gemini Omni is totally different," the corporate posted in June, crediting the mannequin with "an intuitive understanding of forces like gravity, kinetic vitality, and fluid dynamics for extra practical actions that observe real-world logic."

The sensible consequence for enterprise patrons is that world-model language is just not a differentiator. Two of the three main video methods now market bodily understanding as their central benefit, and neither has printed a benchmark that measures it.

There isn’t a commonplace take a look at for whether or not generated water behaves like water, whether or not a dropped object falls at a believable price, or whether or not a sound arrives when the affect does. Human desire scores seize a few of it not directly. Nothing else on supply captures it in any respect.

Open weights helped make FLUX an business commonplace

BFL officially launched in summer 2024 and gained a reputation for itself within the AI business within the intervening two years for its dedication to open sourcing high-quality AI picture fashions beloved by builders, creatives, and enterprises.

The corporate's founders, together with Rombach, Andreas Blattmann and Patrick Esser, beforehand helped create VQGAN, latent diffusion and Stable Diffusion, the latter the open supply expertise that kicked off broad AI technology capabilities for the lots and at present utilized by many AI picture mills and firms.

That attain translated into industrial distribution. FLUX fashions now energy generative options inside Adobe Photoshop, Picsart and Nous Analysis's Hermes Agent, amongst different platforms, and the corporate cites movie director Martin Scorsese amongst skilled customers.

Wired journal described Black Forest Labs as a comparatively small firm that nonetheless turned a number one competitor to Silicon Valley's largest AI labs, with FLUX fashions rating close to the highest of picture benchmarks and changing into a number of the most downloaded text-to-image fashions on AI code sharing neighborhood Hugging Face. The corporate says it now runs a 100-person crew throughout Freiburg and San Francisco.

FLUX.1 Dev, FLUX.1 Kontext Dev, FLUX.1 Fill Dev and associated management fashions, released shortly after the firm's launch, gave researchers and creative-tool builders entry to downloadable checkpoints, native inference and integrations with frameworks together with Hugging Face Diffusers and ComfyUI. FLUX.1 Kontext Dev, for instance, was launched as an open-weight mannequin for analysis and noncommercial use, with generated outputs permitted for industrial functions beneath the relevant license.

The corporate continued that sample with FLUX.2 Dev in late 2025, a 32-billion-parameter open-weight mannequin combining technology and multi-reference modifying. Black Forest Labs known as it the strongest open-weight picture technology and modifying mannequin out there at launch and launched weights, reference inference code and optimized implementations for shopper Nvidia GPUs.

FLUX 3 Dev raises the stakes on that analysis. Earlier Dev releases had been picture fashions. This one is described as a multimodal spine spanning video, audio, picture and motion prediction — which means a single license will govern whether or not an organization can regionally deploy a mannequin that touches each content material manufacturing and bodily equipment. BFL hasn't but shared details about its license, the parameter depend, quantizations or {hardware} necessities.

The corporate frames open weights as an enterprise function moderately than a neighborhood gesture, arguing they allow safe, low-latency native deployment for purposes like robotic management methods and let groups adapt FLUX 3 to their very own knowledge, merchandise and workflows.

The monetary backing behind FLUX 3 is price noting alongside the technical claims. Black Forest Labs is valued at $3.25 billion and has raised greater than $450 million from traders together with a16z, AMP, Salesforce Ventures, Nvidia, Basic Catalyst, Adobe Ventures, Figma Ventures, Canva and Deutsche Telekom's T.Capital.

Source link