Flux 3
Black Forest Labs' first video model. Direct contender to Seedance, Kling and Veo families.
Overview
Black Forest Labs' first cinematic video model: 20-second single-pass clips with native audio, frame-accurate lip-sync in 13+ languages, and keyframe control inside a single generation.
The step change is not resolution or motion quality alone. Picture, sound, dialogue, and camera choreography resolve in one pass. No separate audio model, no sync step.
Note: Flux 3 takes keyframes, not references. There is no subject or style reference input — control comes from frames placed on a timeline. If you need reference-driven consistency, use Seedance 2.5 or Seedance 2.0.
Strengths
20-second one-takes: Single generation, no cuts, no stitching. Subjects and geometry hold across large temporal jumps (day to night, season turn). Clips can be chained past 20s.
Native audio in the same render: Sound is not a second model and not a post-sync step. Events arrive carrying their own sound on the exact frame.
Multilingual dialogue with lip-sync: 13+ languages including English (multiple dialects), French, Spanish, Chinese, Japanese, German, Portuguese, Russian, Italian, Indonesian, Turkish, Hindi, and Punjabi. Mixed-language scenes hold in one take.
Keyframes as a control surface: Set a start image, an end frame, or multiple keyframes inside a clip. The model interpolates between them while holding the intended visual language.
Camera behavior: Orbits, dolly moves, focus racks, and tracking shots preserve parallax and scene geometry through continuous movement.
Aesthetic range: Distinctive spike on nostalgic, retro, and vintage registers — 16mm film grain, archival looks, period-accurate texture.
Ideal use cases
Cinematic ads: TV-grade spots with picture, score, and dialogue delivered as one asset.
Stop-scrolling creative: Viral-register content in the Sora 2 family, with more directorial control.
Multilingual campaigns: One master scene, localized dialogue with accurate lip-sync per market.
Frame-locked production: When specific moments must land exactly, set them as keyframes and let the model resolve the space between.
Weaknesses
Keyframes, not references: Flux 3 takes frames placed on a timeline rather than reference images for subjects or style. Consistency across separate generations requires reusing keyframes, not references.
Render time is a few minutes per generation: Rules it out of fast iteration loops. Explore in a quick model (Gemini Omni Flash, Grok Imagine), then commit the finished idea to Flux 3.
Longer clips reward longer prompts: 20 seconds means more decisions, and the model makes the ones you leave open. Underspecified prompts drift.
How to use effectively
Word order is weight. The model reads the prompt front to back; put the subject and core action first, style and atmosphere after.
Declare absences explicitly. There are no negative prompts. Instead of "no music", write what is present: "only ambient street sound and footsteps."
Quote dialogue verbatim, with a visible speaker. Name who talks, put the line in quotes, state the language and accent if not obvious.
Layer audio with verbs, not adjectives. "Rain hits the awning, a bus hisses past" beats "moody urban soundscape."
Keyframes carry the moments that must land. Set the frames that have to be exact; size the space between them to the action and let the model pace itself.
Use camera language. Locked-off, push-in, pull-back, orbit, tracking — the model executes professional cinematography vocabulary literally.
Last updated