FacePercept-Bench

Do synthetic faces sit closer to the centre of vision-model face space?

A reproducible computational demo comparing real and AI-generated faces across CLIP, DINOv2, SigLIP, and Qwen2.5-VL.

Key finding

Across three independent vision encoders, synthetic faces were consistently closer than real faces to a reference centroid estimated only from real faces in the training split, while remaining strongly separable from real faces.

0.844

CLIP linear-probe balanced accuracy

0.862

DINOv2 balanced accuracy

0.868

SigLIP balanced accuracy

Centrality differences (Synthetic - Real)

CLIP: -0.0381 · DINOv2: -0.0279 · SigLIP: -0.0497. All three 95% confidence intervals remain below zero.

Cross-encoder consistency

Encoder	Balanced accuracy	Synthetic - Real centroid distance	95% CI
CLIP	0.844	-0.0381	[-0.0582, -0.0196]
DINOv2	0.862	-0.0279	[-0.0490, -0.0085]
SigLIP	0.868	-0.0497	[-0.0746, -0.0302]

Question this raises

Does representation-space centrality in modern vision models relate in any meaningful way to the perceptual averageness or hyper-realism effects reported for AI-generated faces in humans?

This demo does not equate embedding geometry with human perception. It provides a computational starting point for testing that relationship.

VLM behaviour

Qwen2.5-VL-3B did not show a supported machine-side hyper-realism effect. It predicted REAL for both classes at nearly the same rate, so that result is treated as a near-constant-response failure mode rather than a positive effect.

What the repo includes

Automatic data acquisition

Balanced real and synthetic face benchmark with reproducible sampling.

Multiple model families

VLM classification plus CLIP, DINOv2, and SigLIP representation analysis.

Reproducible analysis

Bootstrap intervals, effect sizes, held-out centroid analysis, figures, and cached outputs.