Objaverse-XL brings 10 million 3D objects to generative AI training


Researchers unveil Objaverse-XL, a dataset of over 10 million 3D objects, to advance AI development in 3D computer vision and generative AI.

Developments in AI have been accelerated by increased access to large amounts of training data. This is true of generative AI systems for text and images, which have been trained on massive datasets crawled from the web.

One of the next frontiers of AI, 3D computer vision and generative AI for 3D, has lagged behind due to the challenges of acquiring high-quality 3D data.

Objaverse-XL contains 10 million 3D objects

To address this problem, a team of researchers has unveiled Objaverse-XL, a massive collection of over 10 million 3D objects.


Culled from several online sources, including Sketchfab, Thingiverse and Polycam, it is a tenfold expansion of the Objaverse dataset released in April.

At the time, 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 – too late, as it turned out.

Zero123 is a generative AI model for 3D

Using Objaverse-XL, the researchers have successfully trained a model for novel view synthesis. The resulting model, Zero123-XL, has demonstrated strong zero-shot generalization capabilities across a range of complex modalities, including photorealistic assets, cartoons, drawings, and sketches.

According to the researchers, experiments have shown promising scaling trends for 3D vision tasks using Objaverse-XL as data scales from a few thousand to 10 million assets. For this reason, they believe that an even larger dataset containing billions of objects would further enhance the potential capabilities of such AI models.

Objaverse-XL and Zero123 are the result of a cooperation between the Allen AI Institute, Columbia University, UWCSE, Stability AI, LAION and Caltech.


Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top