RealBench V1 Methodology

RealBench V1 is a visual Turing test for image models: can a model's output pass as a real photograph? It is scored entirely by human votes, not by an automated metric.

How the score is computed

Players are shown one image at a time — a real photo or an AI generation — and guess which. A model's realism score is the share of votes that judged its AI images as real. The more people are fooled, the higher the score.

Reliability

• Models need at least 10 votes to appear on the board.
• Each snapshot is dated; scores drift as more votes arrive.
• Realism is not capability — a model can look photoreal yet fail prompt adherence, or vice versa.