Stability AI has unveiled SPAR3D, a new 3D reconstruction model that can create complete 3D objects from a single image in under a second.
At CES, Stability AI partnered with Nvidia to showcase SPAR3D (Stable Point Aware 3D), a system designed to work with Nvidia RTX graphics cards. The model's main selling point is its speed - it can transform a single image into a detailed 3D object almost instantly.
Video: Stability AI
According to Stability AI, SPAR3D offers a high level of control over 3D object creation. The generated point cloud can be edited directly by deleting, duplicating, stretching, coloring or adding new features. SPAR3D is also said to provide accurate geometry and detailed 360-degree views, including typically hidden areas.
Video: Stability AI
The numbers are impressive: it takes just 0.3 seconds to turn a point cloud into a finished mesh, and only 0.7 seconds to create a detailed 3D model from a single image. This speed could make life easier for game developers, product designers, and anyone working on environmental design.
SPAR3D uses a two-part system to create its 3D models. First, a specialized point diffusion model creates a detailed point cloud that captures the object's basic structure. Then, a component called the triplane transformer combines this point cloud with features from the original image, adding high-resolution details for geometry, texture, and lighting.
Free use under community license
The model is available now under Stability AI's Community License, making it free for both personal and business use. Larger organizations bringing in more than a million dollars annually need to contact Stability AI about an enterprise license.
Anyone interested can download SPAR3D's weights from Hugging Face or access the source code on GitHub. Developers can also tap into the model through Stability AI's Developer Platform API.
The timing of this release makes sense, as Stability AI has been focusing on developing more efficient 3D models for a while now. Their partnership with Nvidia feels like a natural fit, especially since Nvidia just unveiled their own text-to-3D system, Edify 3D, in late 2024. They're not alone in this space - Meta, Midjourney, and Luma AI are all working on similar AI-powered 3D modeling technology.
