Seedance

Seedance is a family of video generation models developed by the Seed team at ByteDance. The first generation, Seedance 1.0, was released in June 2025; later ones are Seedance 2.0 (February 2026), Seedance 2.0 mini and Seedance 2.5 (July 2026).

The models generate video both from a text prompt (text-to-video) and from an input image (image-to-video). Seedance 1.0 produces 1080p footage with smooth, stable motion, rich detail and cinematic aesthetics, and natively supports multi-shot storytelling that preserves character identity and style across successive shots.

Technically, the models jointly learn text-to-video and image-to-video tasks, building on carefully curated data and detailed video captioning during pretraining, followed by supervised fine-tuning (SFT) and video-specific RLHF with a multi-dimensional reward mechanism. Through multi-stage distillation and system-level optimizations, Seedance 1.0 achieves roughly a 10× inference speedup — a 5-second 1080p video is generated in about 41.4 s on an NVIDIA L20.

Newer generations extend the capabilities: Seedance 2.5 generates native clips of up to 30 seconds with support for many multimodal references and offers local editing of selected scene regions without regenerating the entire shot (a beta mode extends clips up to 3 minutes).

Seedance is a proprietary model available via API and through ByteDance apps (Doubao, Jimeng/Dreamina) and the Volcano Engine platform. Its quality has been evaluated on the internal SeedVideoBench-1.0 benchmark and on the Artificial Analysis platform, among others.