ByteDance Seed releases the Seedance 1.0 video model

Background

A video generation model creates moving pictures from a text description or a still image. The task is harder than image generation because every frame has to stay consistent with the frames before and after it, so the subject, lighting and motion remain coherent over time.

Release

ByteDance Seed published the Seedance 1.0 technical report on arXiv on 10 June 2025 and announced the model on its blog on 11 June 2025.^[1][2]

Capabilities

Seedance 1.0 generates 1080p video from text or from images. It can produce 10-second clips that include two or three shot transitions, meaning cuts within a single clip from one camera setup to another.^[2]

Leaderboards

The model ranked first on both the text-to-video and the image-to-video leaderboards of Artificial Analysis, a site that compares AI models on published tests.^[2]

Speed

A 5-second 1080p video takes 41.4 seconds to generate on an NVIDIA L20 GPU. The report credits about a 10x speedup to distillation and system optimizations. Distillation trains a smaller or faster model to imitate a larger one, which lets it run with fewer steps.^[1]

Team and access

The technical report lists 44 authors.^[1] The model was offered through an API and through the BytePlus Playground.^[3]

Generating video is expensive because a clip contains many frames, and each one must be produced. For a service shared by many users, speed matters as much as image quality. A shot transition is an editing decision: a director cuts between camera setups to tell a story, so a model that produces such cuts in one generation has to plan the whole sequence rather than fill in each frame on its own.

Leaderboards are one way to compare video models, but a first-place finish depends on the set of prompts used and the rules by which outputs are judged.

Significance

By publishing the report, ByteDance Seed made its method available for others to study.