【Stable-Fast-3D】超高速開源圖片轉 3D 模型　僅需 0.5 秒即可生成高品質模型！｜僅需一張圖即可生成 3D 模型

Stability AI has introduced a revolutionary AI model known as Stable-Fast-3D, which can convert 2D images into 3D models within 0.5 seconds. This model is an advanced, open-source solution that requires just a single image to generate high-quality 3D models. It has been improved upon the TripoSR project, offering more accurate UV mapping and material parameter prediction for a more realistic appearance. The model can be easily integrated into games and is capable of reducing the impact of lighting and shadows, ensuring more accurate texture maps even when the input photo has environmental lighting. Demos and comparisons with other models are available on Stability AI's GitHub repository and official website, showcasing its superior performance despite not being highly refined. The model has been tested on GSO and OmniObject3D datasets, achieving the highest scores in Chamfer Distance (CD) and F-score metrics, indicating closeness to real models. It is suitable for less critical game decorations, distant objects, and has potential applications in e-commerce 3D modeling, AR, and VR. The model is available for free under the Stability AI community license for commercial and non-commercial use for entities with an annual income of less than one million USD. It requires at least 6GB of VRAM for execution but is also accessible through a free Hugging Face Space for online use.

While Stable-Fast-3D excels in speed, it has certain limitations. The model struggles with objects that have self-occluding parts or depressions, tending to fill in unseen areas as solid. This can result in models with slightly odd internal shapes when only a single perspective is provided. The UV mapping approach used is straightforward but can be limiting for modifications, as it uses six views (top, bottom, left, right, front, back) without separating parts that could be independently mapped, such as chair legs. For objects like cups, a more common approach would be to map the outer wall as a single face and separate the inner wall, bottom, and handle. Despite these limitations, the model is suitable for simple WebAR applications with only a few models. However, for more complex uses, the model's high polygon count may not be ideal for games without adjustments. The video also notes that while the model can generate with either quadrilateral or triangular meshing, this option is not available in the free Hugging Face Space, and the generated models have a high polygon count, which could impact performance if used extensively in games.