Using AI Hallucinations to Evaluate Image Realism

New research from Russia proposes an unconventional method to detect unrealistic AI-generated images – not by improving the accuracy of large vision-language models (LVLMs), but by intentionally leveraging their tendency to hallucinate.

The novel approach extracts multiple ‘atomic facts’ about an image using LVLMs, then applies natural language inference (NLI) to systematically measure…