On the RAID dataset of perceptual responses: analysis and statistical causes

A research team from Universitat de València has published a comprehensive analysis of the RAID dataset, establishing precise human detection thresholds for common AI-generated image distortions. Using Mean Squared Error (MSE) measurements, the study found that Gaussian noise consistently produced the lowest detection thresholds, meaning humans notice it more easily than other distortions like rotation, translation, or scaling. Statistical analysis with ANOVA and Tukey Kramer tests confirmed observers are significantly more sensitive to noise, with Fourier analysis revealing that high-frequency image components act as a visual mask for this type of distortion.

The research went beyond traditional psychophysics by employing the PixelCNN generative model to analyze how image probability affects perception. They discovered that statistical likelihood—how probable an image appears—significantly correlates with detection thresholds for most distortions. This suggests that more "expected" or statistically likely images allow for greater tolerance of imperfections. The study also found that spectral orientation influences rotation perception, adding another layer to understanding how humans evaluate manipulated visual content. These findings provide concrete, quantitative benchmarks that could inform everything from AI image generation quality metrics to compression algorithm development and digital content evaluation standards.