A split before any label
Human-created and AI-generated paintings separated along the dominant principal-component directions of a shared CLIP embedding space, a split that appeared before the analysis used a supervised discrimination objective. Principal component analysis, or PCA, is a way of finding the strongest directions of variation in a group of data; here, those directions already aligned with the human-versus-AI distinction. The result is about the geometry of the studied CLIP representation, not proof of a single visual cause.
The work mostly used AI-WikiArt for training and AI-Pastiche and National Gallery of Art collections for testing. Most analyses used CLIP ViT-L/14 at 336 pixels, representing each image from a 336-by-336-pixel input as a 768-dimensional embedding.
The pattern held through basic tests
The analysis followed a progressive framework: it checked robustness, compared basic global image statistics, added increasingly detailed HOG and scattering descriptors, tested whether spatial arrangement mattered, and then used regression-model inversion to inspect movement in CLIP space. HOG provides information about local edge patterns, while scattering descriptors summarize image structure across multiple scales.
Across datasets, image generators, preprocessing procedures and image transformations, the separation remained stable. That pattern argues against simple artifacts or one isolated visual cue as the main explanation. It does not, however, eliminate every generator-specific signature or watermarking mechanism, a possibility the experiments did not fully rule out.
One test showed where the signal was likely to live. With random crops, classification accuracy was 51.6% when the crop covered 10% of the image, but 94.3% when it covered 40%. In other words, tiny fragments carried little of the separation, while a larger view restored high discrimination. The result points to information at an intermediate image scale rather than a cue concentrated in a very small local patch.
The useful clues were spread across the image
Simple color, luminance and variability measures did show some differences between the image groups, but they were modest relative to the gap in CLIP space and did not convincingly account for the dominant directions. Spatial detail mattered more in a test using HOG. A full spatial HOG field predicted the first and second principal components with R2 values of 0.30 and 0.32, respectively; reducing it to a nine-feature spatial average cut those values to 0.03 and 0.02. Taken together, the result is consistent with a weak but distributed spatial signal.
Scattering descriptors offered a stronger, though still incomplete, account. These descriptors track image patterns across scales and their interactions. As the maximum scale increased, the model's R2 for PC1 rose from approximately 0.33 to 0.41, while PC2 rose from approximately 0.44 to 0.49; the gains diminished at larger scales. The numbers describe predictive performance, not a measured causal explanation, and the authors treat scattering as only part of the CLIP-component variation.
Adding more structure did not solve the interpretability problem. When scattering information was kept in image patches so its spatial organization was preserved, the predictive performance fell between the linear and multilayer-perceptron models but did not produce a corresponding improvement in inversion. Inversion here means changing an image to move its representation toward a chosen direction. The comparison therefore did not show that richer spatial bookkeeping made the process more controllable or interpretable.
Small image changes could move CLIP substantially
Direct optimization of CLIP itself produced the sharpest demonstration of the mismatch. Bounded perturbations stayed below 0.05 on the image scale, yet they caused substantial movement along CLIP's principal directions while the altered image remained visually almost indistinguishable from the original. That does not show that people would fail to see every such change: the study reports no formal human psychophysical assessment. It shows that CLIP can be sensitive to changes that are not obvious on visual inspection.
When the predicted scattering representation was inverted instead, the displacement was approximately one quarter as large as in direct CLIP inversion. The authors caution that this comparison is not a direct explained-variance measure, but it reinforces the conclusion that scattering captures only a portion of the variation linked to the dominant CLIP directions. The scattering coefficients themselves also showed a consistent group difference: AI-generated images had larger first-order responses at every analyzed scale and larger, longer-tailed second-order coefficients across combinations of scales. The pattern is consistent with distributed multiscale and cross-scale statistical differences, although the mechanism behind it remains a hypothesis.
A robust pattern, but not a complete explanation
Taken together, these experiments support a narrower conclusion: within the studied CLIP and artistic-image setting, human-AI separation is robust and is consistent with distributed, multiscale image statistics. They do not establish that the same pattern will hold in other vision models or natural photographs. Nor do they show that scattering is the mechanism used by CLIP, or that generator-specific signatures and watermarking are absent. The authors present multiscale statistics as contributors to investigate while leaving the underlying mechanism unresolved.
Paper data and sources
Original title: On the Separation of Human and AI-Generated Images in CLIP Embedding Space
Authors: Andrea Asperti
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text