news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Author: Li Dan

Source: Hard AI

In August, Stability AI, a star startup in generative artificial intelligence (AI), released its latest breakthrough 3D model, Stable Fast 3D, which greatly increases the speed of generating 3D images from a single picture.

Stability AI introduced that Stable Fast 3D can convert a single input image into a detailed 3D asset in just 0.5 seconds, and believes that this innovative model has set a new benchmark for speed and quality in the field of 3D reconstruction. It can quickly generate complete 3D assets, including UV split networks, material parameters, and albedo colors for reduced lighting baking, and users can choose quadrilateral or triangle remeshing, and the processing time will only increase by 100 to 200 milliseconds.

The introduction of Stable Fast 3D represents a major leap forward for Stability AI in reducing AI model processing time.

In March this year, Stability AI released the 3D model Stable Video 3D (SV3D), which is an improvement on Stable Video Diffusion. It can generate the corresponding 3D model with just one picture, which is suitable for new perspective synthesis tasks and 3D generation tasks. SV3D takes up to 10 minutes to generate 3D assets, while Stable Fast 3D only takes 0.5 seconds to complete the same task, which is 1200 times faster.

When Stability AI released Stale Fast 3D, it specifically mentioned that the model surpassed its competitors in several key areas, one of which was that it only took 0.5 seconds to generate 3D assets each time on a GPU with 7GB of graphics memory VRAM, while it took nearly a second on the Stability AI API. In addition, the model has high-quality UV split networks and material parameters, can reduce entangled lighting in textures, and can generate additional material parameters and normal maps.

Stability AI claims that Stable Fast 3D is based on TripoSR, a model developed in collaboration between the company and Tripo AI, which significantly improves the architecture and enhances the functionality. The model is suitable for game and virtual reality (VR) developers, as well as professionals in retail, architecture, design and other graphics-intensive industries. There are multiple use cases in game and film production:

• Use this model to achieve fast inference during pre-production when experimentation is critical.
• The game's static assets (background objects, clutter, furniture)
• 3D Models for E-Commerce
• Rapid model creation for augmented reality (AR)/VR.

Currently, Stable Fast 3D is available through Stability AI's chatbot Stable Assistant and Stability AI API. Users can use the model on the AI ​​open source community Hugging Face, and Stability AI is released under the Stability AI Community License, which is free for research, non-commercial and commercial use.

After Stability AI officially announced the launch of Stable Fast 3D, netizens on social media quickly responded to the post. Some netizens asked how clear the UV network generated by the model was and how many polygons it had. Some netizens said that they liked the animated image attached by Stability AI when it released the model, which showed the machine spitting out a general AI image model and throwing it directly into the trash can. Some netizens asked if the generated 3D results could be printed out.

There have been comments that, currently, content creation for 3D worlds is a bottleneck. We don’t want to pay people to handcraft the Metaverse. Automatic asset generation should be the default, just like in the real world. Stability AI’s new model is a big step in the right direction.