Stability AI drops Stable Video 4D 2.0

Stable Video 4D 2.0 takes a regular, single-angle video of an object and turns it into a dynamic 4D asset you can inspect from any perspective.

Released on May 20, 2025, by Stability AI, the upgraded diffusion model eliminates the bulky multi-camera rigs and reference image setups that usually bottleneck novel-view synthesis. Feed it one object-centric clip, and it spits out coherent multi-angle video frames in one go.

What it does

Generating 3D models in motion from a single view has always been an uphill battle against visual ambiguity. SV4D 2.0 is built to slot directly into digital production pipelines where fast dynamic assets are needed.

Game developers can use it to spin up multi-view sprite sheets for moving characters without tedious manual keyframing. Film and virtual world creators can use it to flesh out background props and dynamic stage dressing straight from live-action reference footage.

Stability AI's internal benchmarks put SV4D 2.0 at the top of the leaderboard across all primary metrics. It bests DreamGaussian4D, L4GM, and the original SV4D on LPIPS for image fidelity, FVD-V for multi-view consistency, FVD-F for temporal coherence, and the overall FV4D composite score. In pure novel-view synthesis, it also outperforms Diffusion^2 and SV3D, delivering noticeably sharper edges and fewer warping artifacts.

How it works

To keep objects from turning into mush as they move across camera angles, the model uses a staged training approach. It first learns spatial geometry on static 3D assets, then layers motion dynamics on top once the shapes are solid.

Under the hood, Stability AI rebuilt the architecture around a 3D attention mechanism. This fuses 3D spatial and temporal tokens together, enforcing consistency across viewpoints and timestamps simultaneously without needing side-angle reference images.

While trained primarily on synthetic sets, it borrows world knowledge from broader pre-trained video foundations, so it handles real-world footage reasonably well. The main catch is fast movement: Stability AI cautions that rapid or violent motion can still trigger visible artifacts.

Why it matters

Dynamic 4D capture used to be the exclusive domain of motion-capture stages and multi-camera photogrammetry studios. Compressing that pipeline down to a single ordinary video lowers the barrier to entry by orders of magnitude for indie game teams and solo visual artists.

Stability AI also avoided the locked-down API trap. The weights are up on Hugging Face, the inference code is on GitHub, and the reconstruction pipeline is fully documented on arXiv under the permissive Stability AI Community License, covering both commercial and non-commercial work. If you build dynamic 3D assets for interactive environments, this is definitely one to pull down and benchmark against your own production footage.