Direct videos up to 30 seconds with text, image, video, and audio references across text-to-video, image-to-video, and multi-reference workflows.
AI modelSeedance 2.52.5Create a video from a detailed prompt with timing, camera, action, and audio direction.
Choose a mode, add the required media, write the direction, and generate a Seedance 2.5 video.
BYTEDANCE SEED · NEXT-GENERATION VIDEO MODELMove from isolated demo shots to a production brief. Seedance 2.5 is designed around longer single clips, dense multimodal reference packages, R2V motion control, synchronized audiovisual direction, and precise local revisions for ads, product films, social content, and narrative work.
Longer scenes, richer references, and more precise creative control
Seedance 2.5 is a next-generation AI video model for production-scale creation. It combines single clips up to 30 seconds, as many as 50 multimodal creative inputs, reference-to-video motion control, stronger long-scene continuity, coordinated audiovisual direction, and precise localized video editing in one directing workflow.
In practical terms, this is a shift from asking for one attractive shot to packaging a small production: script beats, character and product references, motion footage, camera intent, sound cues, and targeted revision instructions can all contribute to the same brief.
The important upgrade is not a vague promise of prettier frames. Longer timing, larger reference sets, and region-level edits are intended to reduce the repeated generate-and-stitch cycle that slows professional AI video work.
Combine scripts, characters, products, environments, performance, camera motion, sound, and local revision instructions in one brief so every input serves a specific shot objective.
Each capability answers a production bottleneck: short narrative windows, fragmented references, motion ambiguity, expensive reruns, disconnected audio, or inconsistent localization.
A half-minute single clip gives a commercial, product demo, short-drama beat, or training sequence room for setup, action, transition, and payoff without treating every moment as a separate generation.
Combine scripts, images, video clips, audio, storyboards, character sheets, style frames, camera notes, and brand assets. The value is not the raw count; it is assigning each reference a clear creative role.
Reference-to-video guidance can communicate how a person moves, where subjects sit in the frame, how they interact, and how a camera travels. Green-screen or neutral performance references can isolate movement from final art direction.
Target the object, region, or moment that needs correction while aiming to preserve composition, lighting, motion, sound, and timeline continuity elsewhere. This turns revision into a scoped instruction instead of a full reroll.
Write dialogue, ambience, effects, music intent, timing, and visible performance into the same brief so sound and image follow one shared narrative arc.
Direct dialogue, lip movement, caption constraints, and sound across major languages while reducing accidental text and unrelated music in the final video.
The strongest use cases are not random spectacle. They are jobs where subject identity, product geometry, motion, camera language, lighting, and sound must survive across a deliberate sequence.
Lock packaging shape, material response, logo-safe framing, reflection behavior, and the sequence of reveal shots. Use reference roles to separate product truth from lighting and motion inspiration.
Describe what changes and what must persist across the timeline: the same character, wardrobe, weather, screen direction, emotional state, and motivated camera movement from one beat to the next.
Use performance footage for body movement, still references for identity and garment construction, and audio cues for rhythm. Keep fabric physics and camera speed aligned with the intended editorial mood.
Treat the model like a production collaborator: establish the deliverable, organize references, direct time explicitly, then review against objective acceptance criteria.
Start with the surface and purpose: a 9:16 social hook, 16:9 product film, 1:1 ecommerce loop, training segment, or story beat. State duration, audience, final action, and what would make the result unusable before adding style language.
Label inputs by function: reference 1 for product geometry, 2 for character identity, 3 for wardrobe, 4 for camera motion, 5 for sound timing, and 6 for color grade. Remove redundant or contradictory assets rather than filling the full allowance.
Break a longer scene into ranges such as 0–8s, 8–20s, and 20–30s. For each range specify subject action, camera position and movement, spatial relationships, transition logic, lighting continuity, dialogue, ambience, and music intensity.
First check identity, product shape, screen direction, timing, contact physics, camera motivation, and audio sync. Only then judge texture and polish. If localized editing is available, request one scoped correction and repeat the invariants that must remain unchanged.
Seedance 2.5 moves the family toward longer, denser, and more editable production workflows for projects that need continuity and precise control.
| Workflow | Seedance 2.5 | Seedance 2.0 | Production impact |
|---|---|---|---|
| Single-clip duration | Up to 30 seconds in one clip | 4–15 second short-form workflow | More room for a complete narrative or commercial beat before editing and stitching. |
| Reference package | Up to 50 combined multimodal inputs | Combined image, video, and audio references | Larger casts, brand systems, storyboards, motion guides, and sound references can fit one brief. |
| Revision model | Localized video editing | Regenerate, refine the brief, or continue from returned frames | A targeted fix may preserve more of an otherwise successful take. |
| Motion control | Expanded R2V, performance, and spatial guidance | Text, keyframe, image, video, and audio reference guidance | Movement can be communicated through demonstrated action instead of text alone. |
| Creative focus | Long scenes, dense references, R2V, and local revision | Fast clips, keyframes, and multi-reference generation | Choose the workflow around scene length, reference density, and the way revisions need to be handled. |
Seedance 2.5 fits longer scenes, complex reference packages, and targeted revisions; Seedance 2.0 fits fast iteration and keyframe-driven short-form work.
A strong prompt explains what happens over time, what each reference controls, what the camera is doing, what viewers should hear, and what must never drift.
Name duration, aspect ratio, platform, audience, and objective first. A “30-second 16:9 product launch film with a clear end reveal” gives every later choice a measurable job.
List invariants explicitly: face, age, hair, wardrobe, product proportions, packaging details, room layout, vehicle shape, or any other element that cannot be redesigned between beats.
Use second ranges to describe setup, action, transition, and payoff. Include screen direction and the final frame so the model knows how each beat connects rather than treating the clip as one vague event.
State camera height, framing, lens feel, movement, speed, and why the move happens. “Track left to reveal the product when the subject turns” is clearer than simply asking for a cinematic camera.
Separate dialogue, ambience, effects, and music. Give timing and intensity, note whether lips are visible, and request silence where needed. Avoid asking one reference track to control every audio decision.
Request no accidental captions, no unplanned music, no extra products, no identity substitution, no logo deformation, no camera teleportation, and no change outside a localized edit region.
This structure is intentionally observable. Every instruction can be checked in the result and refined without rewriting the entire idea.
DELIVERABLE 30-second 16:9 premium product film; restrained, realistic, suitable for a launch page. REFERENCE ROLES @Image 1 controls bottle geometry and cap proportions. @Image 2 controls the graphite set and amber-cyan palette. @Video 1 controls the slow turntable motion only. @Audio 1 controls music tempo only. TIMELINE 0–7s — 85mm macro, slow lateral slide across condensation and cobalt glass; quiet room tone. 7–18s — pull back to a medium product frame as one hand enters, rotates the bottle 30 degrees, and exits; preserve label-safe front geometry. 18–26s — match-cut from the blue reflection into a rain-lit street texture, then return through the same reflection. 26–30s — locked hero frame, amber rim light, music resolves, two seconds of visual hold. INVARIANTS Keep bottle shape, cap height, glass color, camera axis, and light direction unchanged. No extra text, no subtitles, no logo invention, no unrelated background music.
Longer scenes and richer reference control matter most when a video must carry brand truth, sequence logic, consistent people, or specific timing—not when novelty alone is enough.
Build a complete 15–30 second hook, demonstration, objection, and payoff while keeping the product, actor, and offer framing consistent across the sequence.
Preserve packaging geometry and material behavior across macro details, handling shots, turntable motion, contextual use, and a clean final hero frame.
Give a scene time for intention, reaction, camera movement, and emotional change while using identity, wardrobe, location, and storyboard references as continuity anchors.
Coordinate performance, camera rhythm, transitions, ambience, and musical timing for vertical edits, creator series, visualizers, and narrative social posts.
Use scripts, product-operation references, approved environments, and voice direction to create a clearer process demonstration or onboarding segment.
Keep the same brand system and visual narrative while adapting spoken language, timing, talent, environment, or market-specific creative direction.
Learn how to use 30-second scenes, multimodal references, R2V, local editing, subject consistency, sound control, and stronger Seedance 2.5 prompts.
Organize timing, references, camera movement, sound, and invariants so each generation has a clear objective and every revision is easier to control.
