Stability AI introduces Stable Zero123
December 13 2023
Stable Zero123 is the latest view-conditioned image generation model, outperforming its predecessor Zero123-XL by delivering higher quality 3D object views. It leverages an enhanced training dataset culled from Objaverse, incorporates elevation conditioning to improve prediction accuracy, and utilizes an efficient dataloader that accelerates training by 40 times. Despite its advancements, Stable Zero123 requires the same VRAM as Stable Diffusion 1.5 for generating a single view but needs more resources, specifically 24GB VRAM, for 3D object generation processes. The model, intended for non-commercial and research purposes, is now available for download on Hugging Face. To facilitate research in 3D object generation, the open-source code of threestudio has been updated to support Stable Zero123, allowing for text-to-3D generation through a process that combines Score Distillation Sampling and NeRF optimization.
Back to Breaking AI News
What does it mean?
- view-conditioned image generation model: A type of machine learning model designed to create images based on specific viewing conditions or angles.
- 3D object views: Visual representations of objects in three dimensions, indicating that they can be perceived with depth and not just as flat images.
- elevation conditioning: A technique used in model training that takes into account the angle of elevation when predicting or generating images, likely to improve accuracy for perspectives from different heights.
- dataloader: A component of machine learning systems that efficiently loads data, such as training datasets, into the model for processing.
- VRAM: Video Random Access Memory, a type of memory used to store image data for processing by the graphics processing unit (GPU), especially important in high-performance visualization tasks like image and video rendering.
- Stable Diffusion 1.5: A specific version of a machine learning model known as "Stable Diffusion," designed for generating images and possibly other media based on certain input criteria.
- Hugging Face: An AI company that provides a platform for sharing and collaborating on machine learning models, likely mentioned here as a hub where the model can be accessed and downloaded.
- open-source code: The source code of a software that is made publicly available and can be used, modified, and distributed by anyone.
- threestudio: Probably a software library or framework for 3D modeling and generation, which has been updated to integrate with the specified model.
- text-to-3D generation: The process of creating three-dimensional objects based on textual descriptions using machine learning models.
- Score Distillation Sampling: A machine learning technique, possibly used to generate or refine predictions by distilling knowledge across different model outputs or scores.
- NeRF optimization: Refers to optimizing a Neural Radiance Field, a method used to render 3D scenes with complex light interactions from a neural network, tailored to create more accurate and realistic 3D images.