AI Fake Image Detector Learns What's Real, Not What's Fake

A New Approach to Spotting AI-Generated Images

As AI-generated images grow more convincing, researchers at Washington University in St. Louis have developed a detection model that inverts the traditional approach: instead of training on fake images, it learns what real photographs look like.

The model, called SimLBR (latent blending regularization), was created by doctoral student Aayush Dhakal in the lab of computer science professor Nathan Jacobs, working with collaborators at Oak Ridge National Laboratory. Dhakal presented the research this month at the IEEE/CVF Conference on Computer Vision and Pattern Recognition.

"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," Dhakal said, according to Washington University.

Why it matters

Traditional deepfake detectors face a fundamental problem: they're always one step behind. When new image generators launch, detectors trained on older fake images struggle to identify their output. By learning the boundaries of authentic imagery instead, this approach offers more durable protection as generative AI continues to evolve—critical for content verification, journalism, and digital trust.

Training in Minutes, Not Hours

The computational efficiency represents a major advantage. SimLBR operates in latent space, converting images into 1024-dimensional vectors rather than processing full pixel data. This allows the model to train in under three minutes on a single GPU, compared with two hours on eight GPUs required by current state-of-the-art approaches.

"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," Dhakal noted. "This is a significant computational advantage and it's much, much cheaper than learning on the entirety of the pixels."

Staying Ahead of New Generators

The fundamental challenge with fake-image detection is timing. When a new AI image generator launches, detectors don't have access to its output for training. By the time fake images from that generator appear on social media, traditional detectors haven't seen them and can't classify them accurately.

Dhakal's team addresses this by treating any significant deviation from real image distributions as fake. "We can think of this approach as a deviation from reality, where we say that any time things deviate enough from the real distribution, then we're going to classify it as fake," he explained. "That makes our detector robust because we're not looking at very specific patterns from one generative model."

The researchers developed two evaluation metrics: reliability, which measures accuracy and confidence with new generators, and worst-case performance, which represents expected accuracy when encountering future generators that differ from training data.

The research received support from the National Science Foundation and Taylor Geospatial Institute. Details were first reported by Washington University's McKelvey School of Engineering.