Researchers at The University of Texas at Austin have devised a new framework for training Artificial Intelligence (AI) models using severely distorted images. This novel technique, named Ambient Diffusion, allows AI to generate ideas from images without directly reproducing them—an issue which poses the risk of copyright violation in prevalent AI models. Traditional AI models, such as DALL-E, Midjourney, and Stable Diffusion, often use copyrighted images in their large datasets and could inadvertently reproduce these copyrighted works, leading to legal issues.
The Ambient Diffusion approach, however, reverses this problem by deliberately using distorted data for model training. Alex Dimakis of the Electrical and Computer Engineering department at UT Austin and Constantinos Daskalakis of MIT were among the study’s team members. They accomplished this by training a Stable Diffusion XL model using a dataset encompassing 3,000 celebrity images. Initially, the team observed that the AI models trained using clean data tended to directly replicate training examples.
Interestingly, when researchers corrupted the training data by masking up to 90% of pixels randomly, the AI models still managed to yield high-quality, unique images. This corruption process ensures the AI does not access the unaltered versions of original images, thus preventing direct replication.
The Ambient Diffusion framework supports the production of innovative, high-quality images that bear no resemblance to the training data, regardless of its corruption. Giannis Daras, who led the work, noted that their framework offered control over the balance between performance and data memorization. He further explained that with an increase in data corruption, there is a corresponding decrease in the AI’s memorization of the training set.
According to the researchers, Ambient Diffusion’s applications extend beyond addressing copyright issues. For instance, it could be beneficial for scientific and medical applications, especially in research areas where uncorrupted data is hard to come across or too expensive to procure, such as black hole imaging or specific MRI scans. This is particularly important in fields like astronomy and particle physics, which often deal with a limited amount of uncorrupted data.
In conclusion, if the Ambient Diffusion model is further optimized, it has the potential to help AI companies produce creative, high-quality AI content without infringing on the rights of original content creators, hence, preventing unwanted legal troubles. However, whether it will alleviate the concerns that AI could potentially reduce job opportunities for genuine artists remains a matter of discussion. Currently, this framework does ensure that copyrighted works are not accidentally reproduced.