Stability AI drops Stable Video Diffusion

Stable Video Diffusion is Stability AI's first foundation model for generative video, built to turn single still images into dynamic video clips.

Built on top of the original Stable Diffusion architecture, the model is available as a research preview with both code and weights openly accessible.

What it can do

Stability AI launched two distinct image-to-video models under this release, differing mainly in output length.

One generates 14 frames and the other outputs 25 frames. Both allow customizable frame rates ranging anywhere from 3 to 30 frames per second.

According to Stability AI, external human preference evaluations found that the base models outperformed leading closed-source video models.

The architecture is also designed to be adaptable for downstream tasks. With fine-tuning on multi-view datasets, it can handle multi-view synthesis from a single 2D image, opening doors for 3D asset generation.

Access and restrictions

The code is live on GitHub, and the model weights required to run it locally are available on Hugging Face. Technical deep dives are detailed in an accompanying research paper.

For those looking for an interactive setup, Stability AI opened a waitlist for an upcoming web tool featuring a text-to-video interface targeted at advertising, education, and entertainment.

There is an important catch: this is strictly a research release.

Stability AI emphasized that Stable Video Diffusion is not intended for commercial or real-world production at this stage. It is meant to collect feedback on safety and quality before any wider rollout.

Why it matters

Until now, AI video has been dominated by proprietary, closed-door platforms where users pay subscriptions for black-box generations.

Dropping the weights for an open foundation video model fundamentally shifts the playground. Researchers and developers can finally look under the hood, experiment with local hardware, and begin building fine-tunes.

Generating 14 to 25 frames only amounts to a few seconds of footage, and commercial restrictions keep it out of client work for now. But just like Stable Diffusion transformed image generation, this release provides the open base layer that the open-source community will likely spend the coming months building upon.