On November 21st, artificial intelligence startup Stability AI announced the release of its new generative video model called Stable Video Diffusion. This marks the company’s first AI capable of synthesizing video based on their popular Stable Diffusion image model.
The new model is currently available in research preview mode. According to the announcement, Stable Video Diffusion can generate video clips up to 25 frames in length from text prompts. Users can customize frame rates and video duration.
The underlying code for Stable Video Diffusion is now accessible on the company’s GitHub page. The announcement also states that researchers can run the model locally by utilizing weights from Stability AI’s Hugging Face repository.
One highlighted capability is adapting the model for downstream multi-view synthesis tasks using single image inputs. Stability AI plans to build specialized video models using Stable Video Diffusion as a foundation, similar to its approach with Stable Diffusion ecosystem.
The announcement emphasizes that Stable Video Diffusion matches or exceeds the quality of comparable proprietary models according to preference tests. However, it is still in research preview mode pending further refinement based on external feedback.
Stability AI CEO Emad Mostaque stated that all new model launches progress through research preview to ensure quality and safety standards before full release. The company hopes researcher feedback during this period will maximize Stable Video Diffusion’s capabilities.
Stable Video Diffusion represents Stability AI’s latest offering in its range of open-source AI models across modalities like image, language, audio and more. The company positions the new model as a leap forward for generative AI video capabilities.
