Black Forest Labs: FLUX.3 Video

FLUX 3 Video is Black Forest Labs’ multimodal model for generating video with synchronized audio. It creates scenes from a text brief, builds motion around pinned images or continues an existing clip. Speech, effects and ambience can be generated together with the frames, supporting a short scene whose camera movement, action and sound are designed as one sequence.

For text-to-video, describe the subject, setting, visible action, camera behavior and sound. A useful product scene can begin with a wide view, move toward a material detail and end on the product in use. Multi-shot briefs can name each shot and the transition between camera angles, keeping the subject and visual style consistent across the sequence. Typography can appear within the scene, so supply the exact words and their placement when a sign or title matters. The model covers live action, animation, motion graphics and stylised looks.

Image-to-video uses one to ten keyframe images. A single image establishes the opening frame; a pair pins the start and end, while ordered images with timestamps guide specific moments in the clip. These images become frames of the video rather than serving only as general style references. With three or more images that have no timestamps, the producer requires an explicit duration: the first and last pin the endpoints, and the others are placed evenly between them. Timestamped frames must be in time order; automatic duration runs to the last timestamp rounded up, within the 20-second maximum. Prepare keyframes with compatible subjects, framing and visual detail, then describe the motion between them. For a product reveal, the first frame can show the closed object and the last the opened arrangement, with the intervening action defined in the prompt.

Video continuation picks up from the supplied clip’s final frames. The launch workflow uses up to four seconds of existing video and audio as context for the next action, carrying camera behavior, movement and sound across the seam. Choose a source ending that clearly establishes where the subject and camera are going, and describe what happens next. Continuation is different from rewriting an existing shot: FLUX Video Edit is a separate editing model and does not become part of the generation model merely because both accept video.

Text- and image-conditioned generation support clips from 5 to 20 seconds. The mode-specific overview gives video continuation a 5–15-second output range. Duration and aspect ratio can be chosen automatically, or the brief can specify a supported landscape, square or portrait shape, including 21:9, 2:1, 16:9, 4:3, 1:1, 3:4, 9:16 and 9:21. Output is 24 frames per second. Choose the duration around a small number of visible actions rather than trying to fit an entire story into one short generation.

HD output comes directly from generation; FHD, QHD and UHD finish the same scene with a video upsampler. For 16:9, the delivered dimensions are 1280 × 704, 1920 × 1088, 2560 × 1440 and 3840 × 2176 respectively. Dimensions are aligned to multiples of 32, so a named resolution class can differ from a conventional delivery size. Crop the finished frame when an exact standard canvas is needed, preserving the content that matters at its edges. The higher resolution is a finishing step rather than a different interpretation of the shot.

Native audio is enabled by default and can be disabled for a silent clip. For dialogue, quote the line, identify the speaker and describe its delivery; multilingual speech, accents and lip synchronization are part of the model’s audio-video capabilities. Keep dialogue and sound cues connected to the visible event, such as a lid closing or footsteps moving away. Review the voice, timing and spoken wording together with the image rather than treating the sound as an unrelated layer added after generation.

Draft mode produces an HD preview and a draft bundle. Enhance the selected bundle to render that generation at full quality; sending the same prompt as a fresh full-render request is a new interpretation and can produce a different take. Drafts use HD only, while enhancement can select the four resolution classes and defaults to FHD. This gives a practical sequence of exploring the action and composition first, then finishing the chosen take at its intended delivery size.

The producer workflow returns a task identifier before the MP4 is ready, then provides the completed clip when processing finishes. Signed result locations expire after about two hours, so retain the actual asset once the shot is selected. Inspect the keyframe transitions, cuts, typography, subject details and audio seam against the brief. For a longer sequence, preserve the successful shot’s references and ending context as input to the next clip rather than assuming one short request contains the whole production.