Once, generating 3D images was an extremely challenging task involving complex wireframes, software, and hardware. But now, the situation has changed dramatically. Stability AI recently announced a new generative AI technology called Stable Fast3D, which can quickly generate 3D images from a single image.
The most impressive part is, according to Stability AI, the new model can generate 3D images in just half a second. This processing speed is a significant leap compared to previous models, which might have taken minutes to produce similar results, while Stable Fast3D completes the same task at a staggering 1200 times the speed of its predecessors.
Back in March, Stability AI released Stable Video3D (SV3D), which took 10 minutes to generate 3D assets, and now Stable Fast3D has made significant progress.
Stability AI anticipates that this new model will have practical applications in multiple industries, including design, architecture, retail, virtual reality, and game development. Users can access the model through the Stable Assistant chatbot, the Stability AI API, and the community license Hugging Face.
Stable Fast3D Principles
Stable Fast3D did not start from scratch but evolved from the previous TripoSR model. In March, Stability AI partnered with 3D modeling supplier Trip AI to focus on creating rapid 3D asset generation technology.
In the research paper, researchers detailed the innovative working method of Stable Fast3D. At its core, Stable Fast3D uses an enhanced transformer network to generate high-resolution tri-planes, i.e., 3D volumetric representations, from input images. This network is designed to efficiently handle larger resolutions without significantly increasing computational complexity, thereby achieving finer detail capture and reducing aliasing artifacts.
Researchers also detailed an innovative method for material and lighting estimation. The material estimation network uses a novel probabilistic approach to predict global metalness and roughness values, resulting in improved image quality and consistency.
It is also noteworthy that the Stable Fast3D model can combine multiple elements required for 3D images (including meshes, textures, and material properties) into a compact, ready-to-use 3D asset.
Stability AI is perhaps best known for its Stable Diffusion text-to-image generation technology, but it has been researching 3D at least since November 2023. The March release of Stable Video3D improved the quality of 3D image generation and viewing experience. Moreover, last week the company announced Stable Video4D, adding a temporal dimension to short 3D video generation.
Technical Report: https://static1.squarespace.com/static/6213c340453c3f502425776e/t/66ab9814a3551056403508b4/1722521625313/SF3D-10.pdf
Key Points:
- 😃 Stability AI introduces Stable Fast3D technology, generating 3D images in half a second, far surpassing previous speeds.
- 👍 The new model has practical value in multiple industries, with various access methods available.
- 👏 Stability AI continues to lead the development of image generation technology, from 2D to 4D.