seedance-2-0

Call the video_selector tool with preferred_provider set to seedance, an operation like text_to_video or reference_to_video, and parameters for aspect ratio, duration, and resolution. The scoring engine routes to whichever configured gateway (fal.ai, Replicate, Runway, Higgsfield) is available.

Open every prompt with a shot-structure declaration such as a multi-shot montage or single continuous POV statement, then add environment detail, character blocks, beat-by-beat choreography with temporal markers like 0-3s, and explicit camera negations like no cuts or no zoom.

Seedance 2.0 is accessible via fal.ai, Replicate, Runway (Enterprise, non-US), Higgsfield, HeyGen Video Agent, BytePlus ModelArk, Freepik, and aggregators like Pollo and PiAPI. OpenMontage wraps fal.ai, Replicate, Runway, and Higgsfield directly.

Use the reference-to-video endpoint with a clear front-facing portrait, reuse the same reference across all shots, and stack identity-lock phrases like the same character, no deformation, no drift, and do not alter clothing category or primary color.

Use the fast variant only for b-roll, previews, or latency-constrained drafts at lower cost. Avoid it for slow-motion, multi-shot, or dolly-heavy prompts, since it routinely misses complex camera and motion fidelity on the first attempt.

Face drift occurs when identity anchoring is weak or clips are too long. Drop duration to 5-6 seconds, tighten identity-lock language, and cull reference images to the three most consistent ones rather than flooding with nine.