Social Gaze Consistency: New Semantic Cue for Detecting AI-Generated Images
Researchers have unveiled a new method called Social Gaze Consistency, which helps spot AI-generated images, particularly when they involve people or minor edits that lack low-level artifacts. This approach analyzes how gaze direction aligns with head and eye movements among subjects. It employs three main strategies: a specially designed dataset that includes gaze-consistent changes, Block-Compositional Caption Supervision, and careful grouping of pairs to prevent the memorization of unique generator traits. This framework offers a fresh perspective on detection, differing from existing low-level techniques.
Key facts
- Social Gaze Consistency is a high-level semantic cue for AI image detection.
- It focuses on gaze direction, head-eye alignment, and pupil placement coherence.
- The method targets person-centric and partial-edit settings.
- A controlled diagnostic dataset with gaze-consistent perturbations is used.
- Block-Compositional Caption Supervision is one of three mechanisms.
- Strict pair-level grouping prevents generator-fingerprint memorization.
- The approach is orthogonal to low-level artifact detection.
- The paper is from arXiv:2605.27348.
Entities
—