🎬 30 seconds in a single pass. Clips run up to 30 seconds in one generation, double the 15-second ceiling in Seedance 2.0. Setup, camera move, action, and resolve fit inside one take, so a full scene no longer has to be stitched from separate clips.

🎬 Character lock across the full take. The model maps baseline facial geometry and clothing at the start and holds them from the first frame to the last. Reusing the same reference keeps the same person recognizable across separate shots, which is what series and campaign work depends on.

🎬 Motion that holds up over 30 seconds. The diffusion architecture was rebuilt and trained on cinematic footage. Crowds, water, and cloth move more plausibly, artifacts are reduced, and a reference camera move can be applied without the physics falling apart mid-clip.

🎬 Prompt adherence up 20% over 2.0. Briefs that combine several characters, props, and camera directions resolve correctly more often on the first generation, which reduces the number of runs needed to reach the intended shot.

🎬 Editing handled inside the generation. Green screen, camera perspective changes, and reference-based edits are part of the model rather than a separate post step. Reference material can be attached across images, video, and audio to steer the result.