Training data has helped advance generative AI for text and images, but 3D AI has faced a more basic challenge: finding enough high-quality objects to train on. Objaverse-XL aims to narrow that gap with a collection of over 10 million 3D objects.
Why 3D AI needs more data
AI systems that generate text and images have benefited from massive datasets gathered from the web. Building comparable resources for 3D computer vision and generative AI is harder because high-quality 3D data is difficult to acquire.
Objaverse-XL is a response to that shortage. Researchers assembled its objects from several online sources, including Sketchfab, Thingiverse and Polycam. The collection is a tenfold expansion of the Objaverse dataset released in April.
A larger collection gives researchers more examples to use when training and evaluating 3D models. The researchers report promising scaling trends for 3D vision tasks as the amount of data grows from a few thousand assets to 10 million.
Training a model to see new views
The team used Objaverse-XL to train Zero123-XL, a model for novel view synthesis. In plain terms, this task involves generating a view of an object from a different viewpoint.
The researchers say Zero123-XL showed strong zero-shot generalization across several kinds of visual material. These included photorealistic assets, cartoons, drawings and sketches. That range matters because 3D models may need to work with objects that do not all look alike or come from the same visual setting.
The reported results suggest that training on a broader collection can help a model handle examples beyond the specific assets it encountered during training. The source does not establish how the model will perform in every practical setting, but its results point to the value of expanding the data available for 3D vision.
Scale brings questions about collection
The dataset’s size also draws attention to how online 3D work enters AI training collections. At the time of the earlier Objaverse release, Sketchfab said the data had been collected en masse without its or the artists' knowledge. In February, Sketchfab introduced a NoAI tag to prevent this, but the source reports that it came too late for that collection.
This history makes data sourcing part of the story alongside model performance. A large dataset can offer more material for research, while the circumstances of collection affect how creators and platforms view its use.
What researchers hope comes next
The researchers believe that even larger datasets, potentially containing billions of objects, could further improve the capabilities of 3D AI models. That is a projection based on the scaling trends they observed, rather than a result already demonstrated by Objaverse-XL.
They also expect the dataset to support stronger state-of-the-art models and applications such as augmented and virtual reality. Objaverse-XL and Zero123 are the result of cooperation between the Allen AI Institute, Columbia University, UWCSE, Stability AI, LAION and Caltech.
For now, Objaverse-XL offers researchers a much larger foundation for experimenting with 3D computer vision and generative AI. Its contribution is both the scale of its collection and the evidence that more 3D data may help models generalize across varied kinds of visual content.