Stability AI has released a new AI model called Stable Zero123, which can generate high-quality 3D object views from a single image. Stable Zero123 significantly surpasses its predecessor, Zero123-XL, thanks to three key innovations.
Stable Zero123 uses a carefully curated training dataset sourced from Objaverse, specifically retaining high-quality 3D objects. This improvement ensures that the generated 3D objects are more realistic.
During the training and inference processes, Stable Zero123 employs estimated camera angles for "elevation conditioning," a technique that enables the model to make more accurate predictions, significantly enhancing the quality of the generated images. Stable Zero123 also introduces a pre-computed dataset and an improved data loader, boosting training efficiency by 40 times.
Stable Zero123 is now available on Hugging Face for researchers and non-commercial users to download and experiment with. It is important to note that the use of this model is subject to certain licensing restrictions, divided into two versions: Stable Zero123 and Stable Zero123C. The former includes some 3D objects licensed under CC-BY-NC, which can only be used for research purposes; the latter uses only objects licensed under CC-BY and CC0, allowing commercial use for users with Stability AI membership.
Additionally, Stable Zero123 is combined with the open-source code threestudio, supporting open-source research in 3D object generation. Currently, a simplified version of the Stable3D process is in a private preview stage. Through this method, users can utilize Score Distillation Sampling (SDS) to optimize neural radiance fields (NeRF) and construct texture-rich 3D models from images generated by the Stable Zero123 model.
The release of Stable Zero123 not only brings significant technological advancements to the field of 3D object generation but also opens up new possibilities for research and commercial applications.
Official Blog: https://stability.ai/news/stable-zero123-3d-generation
Key Points:
🌟 Stable Zero123 can generate high-quality 3D object views from a single image, significantly improving generation effects.
📈 The model achieves more accurate image generation through improved datasets and elevation conditioning techniques.
🆕 Stable Zero123 is available in research and commercial versions, with the latter requiring Stability AI membership.
