What is Stable Video Diffusion?
Stable Video Diffusion is Stability AI's open latent video diffusion model that turns a single still image or a text prompt into a short, coherent video clip. Here is how it works, the model family, and what you can build with it.
A simple definition
Stable Video Diffusion (SVD) is a generative video model released by Stability AI in late 2023. It is a latent video diffusion model: it takes a single conditioning image as input and predicts a short sequence of new frames that animate the scene. Built on top of the same latent diffusion foundation as Stable Diffusion image models, SVD extends that approach into the time dimension — denoising a stack of frames together so that motion stays smooth and the subject stays anchored to your original image. Pair it with a text-to-image step and a written prompt can become the starting frame for the clip.
How Stable Video Diffusion works
- 1
Encode to latent space
Your input image (or text-to-image frame) is compressed into a compact latent representation that the diffusion model can work with efficiently.
- 2
Condition on the frame
Stable Video Diffusion conditions every generated frame on that initial latent, so the output stays anchored to your source image rather than drifting away.
- 3
Denoise across frames
A temporal U-Net denoises a whole stack of frames at once, learning how pixels should move between them to produce coherent, flicker-free motion.
- 4
Decode to video
The denoised latents are decoded back into RGB frames and assembled into a short clip — typically 2 to 4 seconds at 14 or 25 frames.
Key concepts
Latent video diffusion
SVD extends the image-era latent diffusion approach into the time dimension, denoising sequences of frames instead of single images.
Image & text conditioning
Drive the clip from a single still image, or pair it with a text-to-image step so a prompt becomes the starting frame.
Motion bucket control
A motion_bucket_id parameter sets how much movement the model adds — low values give subtle drift, high values give energetic motion.
FPS & frame conditioning
Conditioning on target fps and frame count lets you tune the pacing and length of the generated sequence.
The model family
Stable Video Diffusion ships as a small family of checkpoints, each tuned for a different frame count and level of consistency.
SVD (14 frames)
The original image-to-video checkpoint
The first release generates 14 frames at 576x1024 from a single conditioning image — a fast, lightweight baseline for short motion clips.
Learn moreSVD-XT (25 frames)
Extended, smoother sequences
SVD-XT is fine-tuned to produce 25 frames, giving longer, smoother motion that holds together better than the 14-frame base model.
Learn moreSVD 1.1
Refined consistency at 25 frames
SVD 1.1 fine-tunes SVD-XT at fixed conditioning settings for more consistent, higher-quality output with reduced flicker and artifacts.
Learn moreCommon use cases
Animating stills
Turn a photo, render, or illustration into a short looping motion clip without filming anything.
Concept previews
Quickly visualize how a scene could move before committing to a full production or shoot.
Social content
Generate eye-catching short clips from a single frame for Reels, Shorts, and feeds.
Research & fine-tuning
As an open-weights model, SVD is a base for researchers experimenting with video diffusion and custom fine-tunes.
Product & art motion
Add subtle, cinematic movement to product shots, portraits, or artwork.
Prototyping pipelines
A reproducible, locally runnable model for building image-to-video features and demos.
Strengths
- +Open weights you can run and fine-tune locally
- +Strong frame consistency anchored to the source image
- +Tunable motion via motion bucket and fps conditioning
- +Efficient latent diffusion keeps compute reasonable
Limits to know
- -Short clips only — 14 to 25 frames, roughly 2 to 4 seconds
- -No native audio generation
- -Limited control over precise object or camera trajectories
- -Best at ambient motion, not complex multi-shot storytelling
Frequently asked questions
Stable Video Diffusion (SVD) is an open generative video model from Stability AI. It is a latent video diffusion model that turns a single still image — or a text-to-image frame — into a short video clip of smooth, coherent motion.
SVD encodes your image into a latent space, conditions on that frame, then uses a temporal U-Net to denoise a whole stack of frames at once. The result is decoded back into a short video where motion stays anchored to your source image.
The motion_bucket_id parameter controls how much movement the model introduces. Low values produce subtle, ambient motion; high values produce more energetic, dramatic motion. It is one of the main dials for tuning a generation.
There are three main checkpoints: the original SVD generating 14 frames, SVD-XT generating 25 frames for longer sequences, and SVD 1.1, a refined fine-tune of SVD-XT with more consistent, higher-quality output.
SVD produces short clips — 14 frames for the base model and 25 frames for SVD-XT and SVD 1.1, which works out to roughly two to four seconds depending on the target frame rate.
Yes. SVD is released with open weights, so it can be downloaded and run locally on a capable GPU or used through hosted services. Its open nature also makes it a popular base for research and fine-tuning.
Try Stable Video Diffusion for yourself
Start from an image or a prompt and watch SVD bring it to life.