Overview
Stable Virtual Camera is a 1.3B-parameter multi-view diffusion model from Stability AI that turns one or more 2D images into 3D-consistent videos with controllable camera movement. It performs novel view synthesis along user-defined camera paths without explicit scene reconstruction, supporting 1 to 32 input images and outputs up to 1,000 frames. It is released as a research preview under a non-commercial license.
Our take
Stability's multi-view video generation with 3D camera control is released for research under a non-commercial licence, which is the decisive fact — weights are available to experiment with and not to build a business on. It is worth understanding as a capability preview rather than evaluating as a product.
Worth it if
Researchers and technically curious users exploring multi-view generation non-commercially.
Look elsewhere if
Any commercial use, which the licence does not permit.
✨ Features
Novel View Synthesis
Generates 3D-consistent new viewpoints of a scene from 2D image inputs.
3D Camera Control
Supports 360°, spiral, dolly zoom, pan, and roll camera movements.
Multi-Image Input
Accepts 1 to 32 input images for richer scene reconstruction.
Flexible Output
Produces videos up to 1,000 frames in square, portrait, landscape, or custom ratios.
⚖️ Pros & Cons
Pros
Cons
💰 Pricing
Research Preview
Free
- Non-commercial research license
- Weights on Hugging Face, code on GitHub
- No commercial use permitted
Categories
❓ Frequently Asked Questions
Can I use this in a commercial project?
Not under a non-commercial research licence. Open weights and open licensing are different things, and this distinction is frequently missed — plenty of AI models are downloadable but restricted commercially. Read the licence before building anything, because discovering the restriction after launch is expensive.
What does 3D camera control add to video generation?
Deliberate, specified camera movement rather than whatever motion the model produces by default, with consistency between viewpoints. Camera behaviour is a large part of why generated video reads as amateur or cinematic, so controlling it is one of the more consequential capabilities in this area.