Stability AI speeds up 3D image generation with Stable Fast 3D

Not all that long ago, it used to be hard, really hard to generate 3D images. It was a process that involved complex wireframes, complicated software and beefy hardware. Today that’s no longer the case.

Stability AI today announced a new generative AI technology called Stable Fast 3D to rapidly generate 3D images from a single image. Oh, and it’s fast too. According to Stability AI, the new model can generate a 3D image in half a second.

This represents a significant leap in processing time, dramatically outpacing previous models that required minutes to achieve similar results. For context back in March, Stability AI released Stable Video 3D (SV3D) which took up to 10 minutes to generate a 3D asset – Stable Fast 3D accomplishes the same task 1200 times faster.

Stability AI expects that the new model will have lots of practical utility across several industries including design, architecture, retail, virtual reality and game development. The model is available for use via Stability AI’s Stable Assistant chatbot and the Stability AI API. The model is also available under a community license Hugging Face.

How Stable Fast 3D works to generate images faster than ever before

Stable Fast 3D isn’t built from scratch but rather evolves from Stability AI’s previous work with the TripoSR model. Stability AI first announced a partnership with 3D modeling vendor Trip AI in March in an effort to build a fast 3D asset generation technology.

In a research paper, Stability AI researchers detail the innovative methods the new model uses for rapidly reconstructing high-quality 3D meshes from single images. The system works by combining several novel techniques to address common issues in fast 3D reconstruction while maintaining speed and improving output quality.

At its core, Stable Fast 3D uses an enhanced transformer network to generate high-resolution triplanes, which are 3D volumetric representations, from the input image. This network is designed to efficiently handle larger resolutions without drastically increasing computational complexity, allowing for finer detail capture and reduced aliasing artifacts.

The researchers also detail an innovative approach to material and illumination estimation. The material estimation network predicts global metallic and roughness values using a novel probabilistic method that has the result of improving image quality and consistency.

Also of particular note is the way the Stable Fast 3D model can combine multiple elements required for a 3D image, including mesh, textures and material properties – into a compact, ready-to-use 3D asset.

From 2D to 4D, Stability AI pushes the gen AI envelope

Stability AI is still perhaps best known for its Stable Diffusion text-to-image generation technology. 

While Stable Diffusion is a 2D image generation technology Stability AI has been working on 3D since at least November 2023, with the release of Stable 3D. Stable Video 3D which debuted in March of this year, brought with it the ability to do basic camera panning for image viewing along with a boost in 3D image generation quality.

Stability AI isn’t stopping at 3D either. Just last week the company announced Stable Video 4D which adds the dimension of time to short 3D video generation.