Stable Video Diffusion

Stability AI's first open video model animates a single still image instead of working from a text prompt. The base SVD checkpoint generates 14 frames at 576 by 1024, and the XT variant extends that to 25 frames, giving clips of up to about four seconds. Beyond the source image, control is limited to micro-conditioning values such as frame rate, a motion bucket that raises or lowers the amount of movement, and noise augmentation, so camera moves can be slow or absent. Reference scripts and Streamlit and Gradio demos sit in Stability's generative-models repo, alongside the later SV3D orbital-view and SV4D video-to-4D models. Most people run it through the Diffusers StableVideoDiffusionPipeline or ComfyUI's built-in SVD nodes; with CPU offload, feed-forward chunking, and small VAE decode chunks, Diffusers reports under 8 GB of VRAM, at the cost of speed. The repo code is MIT, but the weights use the Stability AI Community License: research and personal use are free, and commercial use requires registration and ends above 1 million US dollars in annual revenue. Good for anyone wanting a local image-to-video baseline.