The world of artificial intelligence (AI) never ceases to surprise us. Just when we thought we had seen it all, a new development emerges that could change the game entirely. In this instance, the clash of AI and copyright has taken an intriguing turn, as it enters the realm of three-dimensional content.
The Impact of Generative AI Models
AI models like DALL-E 2, Midjourney, Stable Diffusion, ChatGPT, and most recently GPT-4 have the potential to revolutionize various industries. They can generate images, text, and even code, shaking up the job market and raising concerns among those whose hard work has been aggregated as training data.
The question now arises: Should data released under Creative Commons licenses be freely available for AI training? Lawsuits are underway to decipher the legalities surrounding this issue. Companies like Stability AI offer artists the choice to either opt out of training data or be compensated for their contribution, similar to the approach taken by Adobe and Getty Images.
Public databases have also taken steps to address this concern by introducing new terms that allow creators to prohibit the use of their data for AI training.
The Emergence of Generative AI Models for 3D
After conquering text, code, and images, it was only a matter of time before generative AI models turned their attention to 3D content. Although the initial models are still a far cry from their text and 2D counterparts in terms of quality, progress is being made. The main hurdle in developing high-quality 3D models lies in the limited availability of large datasets.
Fortunately, researchers from the Allen Institute for AI and the University of Washington have released “Objaverse,” an extensive 3D dataset. Objaverse boasts over 800,000 3D models, including more than 44,000 animated objects, making it exponentially larger than previous datasets like Shapenet. It also incorporates nearly 400 times more categories, including photorealistic models.
Image: Deitke et al.
The Controversy Surrounding Objaverse’s Data Source
Interestingly, the data for Objaverse was scraped from the 3D platform Sketchfab, which operates under a Creative Commons License. However, the creators of the 3D models included in the dataset were not informed that their data would be collected for AI training. This revelation has ignited controversy, particularly because the dataset incorporates 3D models whose creators had specifically set the “NoAI” tag, a feature introduced by Sketchfab in February to prevent such occurrences.
In fact, Sketchfab has a direct agreement with EpicGames, the company that acquired Sketchfab in 2021, stipulating that the hosted 3D models cannot be utilized for training generative AI models.
Sketchfab’s NoAI: A Case of Good Intentions, Poor Timing
In response to these developments, Sketchfab expressed their surprise, stating, “It appears Objaverse mass-downloaded these models from Sketchfab and redistributed them without our knowledge. So far, all of the models that we’ve seen were made downloadable for free on Sketchfab under Creative Commons licenses.” Alban Denoyel, the co-founder and CEO of Sketchfab, acknowledged that they had implemented the NoAI tag too late, adding, “We understand artists’ concerns and are looking into it.”
The legal implications for the creators involved will likely become clear in the coming months as the trial regarding generative AI models for text-to-image progresses.
For more information about Objaverse, visit the Objaverse project page.
So, the AI copyright dispute marches on, now extending its reach to the three-dimensional realm. As the world grapples with the challenges posed by innovative technologies, it is crucial to strike the right balance between AI advancements and the rights of content creators.