​Stability AI launches SPAR3D: generating 3D objects from a single image in one second

At the just-concluded CES show, Stability AI announced the launch of an innovative method called SPAR3D (Stable Point Aware3D). This two-stage 3D generation technology can generate accurate 3D images from a single image in less than one second. 3D objects. The introduction of this technology provides game developers, product designers and environment builders with a new way of 3D prototyping.

SPAR3D is unique in its ability to model visible and invisible parts separately. When generating 3D objects, this method provides accurate geometry and a complete 360-degree view, including details on the back of the object. Users can also directly modify the generated 3D objects through the real-time editing function, such as deleting, copying, stretching and recoloring point clouds, which greatly improves the flexibility of operations and creative space.

The workflow of SPAR3D can be divided into two stages. The first stage uses a lightweight point diffusion model to generate a sparse 3D point cloud, and the sample extraction speed is extremely fast; the second stage uses these point clouds and input images to build a high-detail mesh model. Such a design not only improves computational efficiency, but also ensures output quality and overcomes the limitations of traditional single-image 3D reconstruction methods.

In performance tests, SPAR3D performed well on the GSO and Omniobject3D datasets, outperforming many existing state-of-the-art algorithms. The research team verified the effectiveness of this technology through multiple experiments, especially its advantages in handling image occlusions and detail reconstruction. SPAR3D uses explicit point clouds as intermediate representations, allowing users to further edit 3D models and improving user experience.

SPAR3D function summary mainly includes the following points:

real-time generation capability

Rapid generation: Able to generate complete 3D object structures from a single image in less than one second.

Efficient editing: The edited point cloud is converted into the final mesh in just 0.3 seconds.

Advanced editing features

Point cloud editing: Users can directly operate on point clouds, including deleting, copying, stretching and adding features.

Recolor: Provides the ability to adjust the color of 3D models to enhance creative flexibility.

Accurate structure prediction

Full structure prediction: Ability to provide accurate geometry and detailed 360-degree view predictions, including hidden regions.

High-resolution output: Generate high-resolution tri-plane data via a tri-plane transformer, capturing the texture and lighting of the image.

Open usage policy

Broad application license: Supports commercial and non-commercial use, users can obtain and use the model on multiple platforms.

Developer friendly: You can download the weights through Hugging Face, get the code for use on GitHub, or access it through the API.

Official introduction: https://stability.ai/news/stable-point-aware-3d