Stable Virtual Camera Released: Multi-view Video Generation Based on 3D Camera Control

Stable Virtual Camera is a Multi-view Diffusion Model that transforms 2D images into immersive 3D videos without complex reconstruction or scene optimization processes. By combining traditional virtual camera control with the powerful capabilities of generative AI, it delivers precise, intuitive 3D video output as the user intends.

Key features are as follows:

• Dynamic Camera Control: Supports over 14 dynamic camera paths, including user-defined trajectories as well as 360° rotation, spiral, dolly zoom, pan, and roll.
• Flexible Input and Aspect Ratios: Accepts input ranging from a single image up to 32 images, and generates various aspect ratios such as 1:1, 9:16, and 16:9 without additional training.
• Long-form Video Generation: Maintains 3D consistency even in long videos of up to 1,000 frames, achieving smooth loops and transitions.

Technically, it uses a Two-pass procedural sampling approach that can handle any number of input and target views. It first generates anchor views, then renders target views in chunks, producing smooth and consistent results. Through this, it achieved state-of-the-art (SOTA) performance surpassing existing models such as ViewCrafter and CAT3D.

However, in the initial version, quality may degrade in scenes containing people, animals, or dynamic textures like water. There is also a limitation where flickering can occur when the camera path overlaps with objects, when the scene is highly ambiguous, or when the difference between the input image and target views is large.