Stable Video 4D 2.0: High-Precision New View and 4D Generation Upgrade Based on a Single Video

Stable Video 4D 2.0 (SV4D 2.0) is a multi-view video diffusion model that generates dynamic 4D assets using a single object-centric video. It is an upgrade over the existing SV4D, delivering higher-quality results even from real-world videos, and is optimized for professional workflows ranging from creating sprite sheets for game characters to supporting film and virtual world assets.

The key upgrades are as follows:

• High-precision 4D output: Through staged training that progresses from static 3D assets to a motion stage, it produces sharper and more consistent 4D results.
• No reference views required: It works directly from a single video without needing separate multi-view reference images.
• Redesigned network architecture: It introduces a 3D attention mechanism that combines 3D spatial and temporal features to enhance spatiotemporal consistency.
• Improved generalization to real-world settings: Despite being trained on synthetic data, it maintains consistent performance on real-world videos by leveraging knowledge from pretrained video models.

According to benchmark analysis, SV4D 2.0 achieved SOTA performance, ranking first on key metrics including LPIPS (image fidelity), FVD-V (multi-view consistency), FVD-F (temporal consistency), and FV4D (4D consistency). It also demonstrated superior novel-view synthesis performance compared to models such as Diffusion^2, SV3D, and SV4D.

SV4D 2.0 is now available for both commercial and non-commercial use under the Stability AI Community License. The model can be downloaded from Hugging Face, and the related code is available on GitHub.