WorldFact-Bench: Beyond Image-Internal Plausibility to Image-World Consistency

Key points

• Advances in image generation have made visual authenticity increasingly difficult to assess.1
• Although image forensics now examines both generation artifacts and higher-level visual inconsistencies, a plausible image can still contradict real-world facts or rules.1
• Each pair introduces a specific, evidence-supported factual conflict while seeking to preserve non-target content and visual plausibility.1
• We further propose PERSIST-Agent, which organizes iterative verification around a persistent state linking candidate facts, visual observations, evidence, and verification statuses.1

Sources (1)

• [1]WorldFact-Bench: Beyond Image-Internal Plausibility to Image-World Consistency
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 8, 03:42 AM

  We introduce WorldFact-Bench to evaluate image-world consistency from a single image, without a predefined claim or verification target.

  Advances in image generation have made visual authenticity increasingly difficult to assess.

Extractive summary: sentences quoted from the sources.