Contents
In brief
Low-level generative artifacts are largely handled; scene semantics remain. The paper proposes Social Gaze Consistency—mutual gaze/head/pupil coherence—as a cue for AI-generated image detection.
What they studied
A controlled paired-edit dataset, Block-Compositional Caption Supervision, and evaluation on FakeVLM and Effort.
Key findings
- +3.7 pp balanced accuracy on COCOAI Interaction; +1.3 pp on Person.
- Both real and fake recalls improve—not an all-fake strategy.
- Training on one inpainter generalizes across generator families.
What this means for developers
Combine pixel and semantic moderation signals; use paired counterexamples in VLM eval; fixed reasoning skeletons stabilize caption supervision.
Limitations
Person-centric scenes; code release pending acceptance.