Seedance 2.5 ranks 1st in Multi-Image to Video on Design Arena with an Elo of 1400.

With a significant score gap from many other top models.

The benchmark gives video models multiple reference images and compares how well they combine them into a coherent video while preserving the subjects, visual style, and scene continuity.

This setup is actually closer to real production, where we already work from reference packs rather than one perfect source image.

When the model preserves those inputs, artists spend fewer iterations repairing identity drift after generation.