Stability AI Ltd., the open-source generative artificial intelligence startup, has been on a roll, releasing new AI models that generate 3D assets from 2D images and capture motion in 3D.
The company announced Thursday the release of a new generative AI model called Stable Fast 3D, which creates 3D models from a single 2D image lightning fast – in about half a second. According to the company, this sets a new standard for speed and quality regarding 3D reconstruction.
“Stable Fast 3D’s unprecedented speed and quality make it an invaluable tool for rapid prototyping in 3D work, catering to both enterprises and indie developers in gaming and virtual reality, as well as retail, architecture and design,” Stability AI’s team said in the announcement.
Starting from a single image, the new generative AI model rapidly generates several valuable 3D assets that can be imported into modern 3D modeling software. That includes a UV unwrapped mesh for texture mapping, material parameters, albedo colors with reduced baked-in illumination and optional quad or triangle remeshing, which adds only 100 to 200 milliseconds to the processing time.
The model can also predict the type of material the 3D asset might be, producing more likely scenarios for rendering. For example, it’s more likely to create genuine shiny surfaces for metal or matte surfaces for leather and walls. That means it can handle many different types of materials, textures and lighting better than other models in the industry, the company said.
The AI model is based on previous work with a model called TripoSR, a fast 3D reconstruction model inspired by Large Reconstruction Models for single images to 3D. However, the new SF3D model features a whole new retrained set of parameters and architecture that allow it explicitly to create fast texturing and meshes at high speeds.
For speed comparisons, it takes 10 minutes for Stability AI’s previous model Sable Video 3D to transform still images into rotating 3D videos for 3D model creation. SF3D spends a half-second per asset generation on a graphical processing unit with 7 gigabytes of VRAM, or close to a second using Stability AI’s developer service.
Stability AI has been rolling out new models at a fast pace. The company released SV3D in March and two weeks ago it introduced Stability Video 4D, an AI model that can take a video of a 3D object from one angle and reproduce it from multiple views. All of these models, including the newly released SF3D, provide high-quality outputs that would be useful in virtual reality, e-commerce, gaming, architecture and other industries where 3D modeling is necessary.
The company said SF3D model code is available on GitHub, with model weights and demo space on Hugging Face. The model is also accessible to developers via the company’s application programming interface and its “friendly chatbot” Stable Assistant.