A new dataset could help researchers build better generative AI for 3D content. But Objaverse’s scale has also brought a dispute over how its models were collected, whether creators knew their work would be included, and what a platform’s NoAI tag can do when applied after collection.
A larger pool of 3D material
Generative models for text, code, and images have drawn attention to the data used to train them. The same issue is emerging in 3D: early models exist, but their quality is still far from that of text and 2D image systems. The source article identifies a lack of large 3D datasets as a main reason.
Researchers at the Allen Institute for AI and the University of Washington released Objaverse to address that gap. It contains more than 800,000 3D models with descriptions, including more than 44,000 animated objects. The article says the dataset is more than an order of magnitude larger than Shapenet, the largest dataset to date, and includes nearly 400 times as many categories, including photorealistic models.
Those details help explain the dataset’s appeal for 3D AI research: it offers a broad collection of models and descriptions. They also make the way the material was gathered consequential. A large dataset can include work by many creators, so questions about permission and licensing affect more than a single model.
Why Sketchfab creators objected
According to the article, Objaverse’s data came from Sketchfab, where models were available under Creative Commons licenses. The creators whose work was included were not informed that it had been collected for AI training. The dispute intensified because some included models had a NoAI tag, which Sketchfab introduced in February to let authors signal that their work should not be used for AI.
One creator publicly asked how their models had appeared in the dataset after using the tag. The timing is central to Sketchfab’s response: the platform said the models it had seen were downloadable for free under Creative Commons licenses, and that Objaverse had gathered them before the NoAI tag was introduced. Sketchfab said it was looking into the issue and the remedies available to it.
That leaves a practical gap between a creator’s later preference and a dataset that had already been assembled. The article describes the tag as a way to prevent AI use, but also reports Sketchfab’s explanation that the collection happened before the tag existed. The source does not say whether the tag can remove models already gathered or settle what earlier licensing allowed.
Licenses, platform terms, and AI use
Creative Commons licensing is part of the dispute because the models were made downloadable under those licenses. The article notes that lawsuits are underway over whether data released under Creative Commons can also be used for AI training. It does not report a ruling on Objaverse, so the presence of a license alone does not resolve the question raised by this case.
The article also describes a separate agreement between Sketchfab and EpicGames, which acquired Sketchfab in the summer of 2021. Under that agreement, Sketchfab models cannot be used by EpicGames to train generative AI models. That arrangement concerns EpicGames; the source does not say that it governed Objaverse’s collection.
These distinctions matter because platform rules, creator preferences, and the terms attached to downloadable files may not answer the same question. The source reports that Sketchfab acted against AI use by introducing NoAI, while Objaverse’s collection had already taken place. It leaves open how those facts interact legally.
What remains unresolved
For the affected creators, the immediate issue is whether they have legal recourse against the dataset. The article says that may only become clear in the next few months, when the first results of the trial concerning generative AI models for text-to-image are expected to be available. It does not predict what those results will be or how they would apply to 3D models.
The Objaverse dispute therefore brings the broader AI training debate into a new kind of creative work. The dataset may give researchers a much larger resource for 3D generation, while its collection has raised concerns about notice, licensing, and creators’ ability to object. Until the legal questions are resolved, the case illustrates how difficult it can be to apply a newly introduced opt-out to material that was already gathered.