seedance-2-5

Call a supported hosted route: fal.ai with model_version 2.5, Volcengine Ark with model 2.5, Runway with model seedance2_5, or a ComfyUI Partner Node with model_family seedance_2.5. There are no local weights; all routes are paid and hosted.

Use reference images rather than text descriptions, build a three-plate character sheet with the face only on a sharp close-up, and apply named locks like COUNT LOCK and SCREEN DIRECTION. State monotonicity rules prevent the model from resetting details between shots.

Seedance 2.5 accepts up to 30 reference images, 10 reference videos, and 10 reference audio clips, with combined video and audio duration capped by the provider (Runway caps at 30 seconds). Smaller, deliberate reference sets outperform maxed-out cluttered ones.

The API routes expose 480p and 720p only. Native 1080p generation exists on the vendor's own web platform, which is not an API route, so check which surface a resolution claim refers to before planning a pipeline.

Vague transitions are the most common cause. Give each shot an explicit time range, camera distance, and framing, separate them with Hard cut markers, and lock location, lens, and screen direction so nothing drifts between segments.

Singing, sustained physical contact like fights, and multi-character scenes are the hardest categories. Workarounds include pre-recorded labeled audio tracks, phone-shot motion references, and heavier reference material per face.