Model uses real image to train AI to look for fakes
Nathan Jacobs’ lab tackles detecting AI-generated images with the real thing
AI-generated images have become increasingly more sophisticated from early ones that showed humans with more than five fingers on a hand, making it even harder to determine if photos are authentic. A team of computer scientists in the McKelvey School of Engineering at Washington University in St. Louis has developed a model that learns to detect fake images by learning which are real.
Aayush Dhakal, a doctoral student in the lab of Nathan Jacobs, professor of computer science & engineering, and collaborators at Oak Ridge National Laboratory created a simple yet efficient model trained to detect a real image. Dhakal introduced the model, SimLBR (latent blending regularization), at the IEEE/CVF Conference on Computer Vision and Pattern Recognition 2026 in June 2026.
“These generators keep getting better and better, and they don’t show any signs of plateau yet, so I think in the future, it will be impossible for a human to determine if an image is fake,” he said.
Dhakal said one of the big differences of his team’s approach is that it operates in the latent space, which means it projects very high-dimensional pixels into relatively lower dimensions using a foundation model. They use that to convert images into a 1024-dimensional vector, so the model only learns in the 1024-dimensional space.
“Our method requires under three minutes of training on a single GPU, compared with two hours on eight GPUs for the state-of-the-art approach. This is a significant computational advantage and it’s much, much cheaper than learning on the entirety of the pixels.”
Dhakal said focusing on detecting fake images puts a user a step behind because the technology changes as image generators are upgraded.
“When a new AI-generated model launches, you won’t have access to the images it previously created,” he said. “Once you have that access, you can train the model, but when images from that model are surfacing in social media or online, the detector has not seen those and is not going to be able to classify them well.”
Dhakal and collaborators developed two metrics they used to evaluate the accuracy of the detectors: reliability and worst-case performance. A higher reliability score indicates a detector that achieves high accuracy and low uncertainty, which gives them an idea of how confidently they can expect the detector to work with new AI generators. Their worst-case performance represents the expected performance of a detector when it meets a future generator that deviates from its training.
“We can think of this approach as a deviation from reality, where we say that anytime things deviate enough from the real distribution then we’re going to classify it as fake, and that makes our detector robust because we’re not looking at very specific patterns from one generative model.”
Dhakal A, Khanal S, Sastry S, Arndt J, Dias PA, Lunga D, Jacobs N. SimLBR: Learning to Detect Fake Images by Learning to Detect Real Images. IEEE/CVF Conference on Computer Vision and Pattern Recognition, June 2026. https://arxiv.org/abs/2602.20412
Support for this research was provided by the National Science Foundation (OAC-2232860) and Taylor Geospatial Institute.
