Seedance 1.0

What it is

Seedance 1.0 is a video-generation model from ByteDance Seed, the AI research arm of ByteDance. Its technical report was submitted to arXiv on 10 June 2025 with 44 authors, and the model was announced on the Seed blog on 11 June 2025.^[1][2] It handles two tasks that had usually been split between separate systems — text-to-video and image-to-video — inside a single model, and it generates 1080p video made up of multiple shots.^[3]

Many shots in one clip

The distinctive claim of Seedance 1.0 concerns continuity across a cut. The model keeps subject, style and atmosphere consistent from shot to shot, and it can produce a 10-second video containing two or three shot transitions.^[2][3] A cut is where generated video most often breaks down: the same character returns with a different face, a different lens, or different light, and the viewer immediately notices. Holding those properties across a transition is a different problem from making a single attractive shot, and it is the problem this model is built around.

Training

The arXiv abstract frames the design goal as balancing three things: how closely the output follows the prompt, whether the motion looks plausible, and visual quality.^[1] These pull against one another — a model can obey a complicated prompt and produce stiff movement, or produce beautiful movement that ignores part of what was asked. The paper also describes a post-training stage built on fine-tuning plus video-specific RLHF, that is, reinforcement learning from human feedback, in which the model is adjusted using human judgements of sample quality rather than only a fixed target.^[1]

Engineering for speed

Two engineering details are documented. Multi-stage distillation together with system optimizations gives roughly a 10x speedup at inference, to the point where a 5-second 1080p video takes 41.4 seconds on an NVIDIA L20 GPU.^[1] Separately, a thinner VAE decoder — the part that turns the model's internal representation back into pixels — doubled decoding speed without visible loss of quality.^[2] Cost per second of output is a practical constraint on video generation, so how long a single clip takes to produce shapes what the model can be used for.

Evaluation

As of early June 2025 the model ranked first on both Artificial Analysis video leaderboards, the text-to-video board and the image-to-video board.^[2][3] Internally, ByteDance evaluated it on SeedVideoBench-1.0, which uses 300 prompts for text-to-video and 300 for image-to-video, with human raters scoring the results.^[2] ByteDance's own page for the model says it performs well on that benchmark in prompt adherence, motion quality and aesthetics, but no numeric scores are published there.^[3]

Availability and what followed

Seedance 1.0 is offered through an API and through the BytePlus Playground.^[3] Seedance 2.0 followed in February 2026.^[4] Taken together, the release combined 1080p multi-shot generation in one model with a documented effort to cut inference time, and the first places on two public leaderboards were the visible sign of that at the time.^[2][3]