Black Forest Labs launches FLUX 3 Video
The Freiburg lab's first video model is generally available through its API, producing clips of up to 20 seconds with native audio and lip-synced dialogue in more than 14 languages. It supports keyframes, continuation from existing footage, multi-shot clips, and a cheaper draft mode.
A draft mode priced well below final is the feature that maps to how production actually spends: explore cheap, commit expensive. It is also the most serious European video model yet, which matters to clients who care where their tools and data sit.
Black Forest Labs built its reputation on images, and FLUX 3 Video is its first move into motion. The release is notable less for any single specification than for the shape of the product: it is built around iteration, with a cheap preview stage and a full render that is meant to keep what the preview got right.
What the lab’s announcement lists: general availability through the BFL API and select partners, clips of up to 20 seconds, HD (720p) output with Full HD (1080p) via upscaling, and native audio covering dialogue, effects and ambience. Lip-synced speech is supported in a long list of languages, among them English dialects, German, French, Spanish, Japanese, Hindi and Punjabi. Control comes through text, a start image, an end frame or multiple keyframes; continuation from up to four seconds of existing video and audio; and multi-shot generation that switches scenes and camera angles within one clip. In the lab’s own human-preference evaluation, FLUX 3 led in text-to-video and tied with Seedance 2.0 in image-to-video. Pre-release risk testing was carried out by a third party, Cinder.
The draft mode is the feature we would test first. The promise is that a draft that works can be rendered at full quality with the same subjects, composition and motion. If that holds, it maps onto how commercial and previs work already runs: many cheap options in front of a director or a client, then a committed render of the chosen one. If the final render reinterprets the draft, with a different face or a changed camera move, the mode becomes a source of false approvals, which is worse than having no draft at all. Anyone building a pipeline around it should check draft-to-final fidelity on their own material before a client signs off on a draft.
Four-second continuation is more useful than it sounds. It means a generated shot can pick up from a live-action plate, sound included, which makes the model a candidate for extending takes, bridging edits or testing alternates rather than only producing standalone clips. Combined with keyframes, that is a toolset aimed at editors and VFX artists as much as at prompt writers. The 720p native resolution, on the other hand, means upscaled output for anything bound for broadcast or cinema, and finishing teams will want to judge that upscale against real delivery specs.
The European angle is real but should not be overstated. A Freiburg lab gives clients a European vendor option, which matters in some data and procurement conversations, but the nationality of a vendor is not the same as data residency or provenance, and the announcement does not address marking or C2PA. With the EU AI Act’s labelling duties in force since August 2, that is the question European buyers should put to BFL directly. Also unstated: pricing figures and licensing terms. The roadmap mentions an open-weight FLUX 3 Dev variant and generation from combined image, video and audio references; an open-weight model in particular would change how studios could run it on their own infrastructure.