Stability AI drops Stable Video 3D to turn single images into 3D meshes

Stable Video 3D (SV3D) is a generative model that turns a single 2D image of an object into smooth multi-view orbital videos and full 3D meshes. Built by Stability AI on top of its video diffusion stack, it targets the biggest headache in single-image 3D generation: warped perspectives and mismatched angles.

What it does

Feed SV3D a single photo of an object, and it predicts what that object looks like from virtually any angle, outputting a continuous orbital video around it.

The release splits into two specific variants depending on your workflow.

SV3D_u is the plug-and-play version. It takes a single image and generates a clean orbital video without requiring any camera parameters or manual conditioning.

SV3D_p gives you director-level control. It accommodates both single images and orbital views, letting creators define custom camera paths for the video output.

Beyond just generating novel video frames, SV3D leverages those multi-view outputs to optimize 3D Neural Radiance Fields (NeRF) and extract ready-to-use 3D polygon meshes.

Under the hood

Earlier tools like Stable Zero123 relied on image diffusion, essentially guessing each new viewpoint frame-by-frame and frequently hallucinating mismatched shapes.

SV3D instead builds on Stable Video Diffusion with added camera trajectory conditioning. Because video models inherently understand frame-to-frame temporal coherence, they translate exceptionally well into spatial multi-view consistency.

For 3D mesh reconstruction, the team introduced a masked score distillation sampling (SDS) loss to sharpen up regions that are occluded or invisible in the predicted views.

It also runs a disentangled illumination model, optimizing lighting separately from the actual shape and texture. That prevents harsh real-world shadows from getting permanently baked into the final 3D asset. Benchmark tests show it noticeably outperforming older alternatives like Zero123-XL and EscherNet.

Licensing and access

The model weights are available now on Hugging Face alongside the research paper for non-commercial experimentation.

Commercial deployment is locked behind a paid Stability AI Membership, so companies looking to bake it into production workflows will need a subscription.

Why it matters

Borrowing the spatial intuition of video diffusion models to fix 3D asset generation is an exceptionally smart move. For indie game studios, 3D artists, and e-commerce teams, turning concept art into an initial 3D mesh in minutes cuts substantial prep work.

That said, it is currently geared toward isolated single objects rather than complex, cluttered scenes, and the resulting meshes will still need cleanup if you require production-grade topology.