AV-FABLE Benchmark for Long-Form Audio-Visual Forensics
October 23, 2026
AV-FABLE introduces a dual-timeline benchmark for detecting sparse manipulations in 998 hours of long-form audio-visual video. It uses 16,650 twin pairs to enable modality-specific temporal localization, exposing errors missed by joint-timeline evaluation methods.
HOW THIS AFFECTS YOU
●
researcherYou can use this to evaluate how well AV-LLMs detect localized deepfakes in long videos.
●
policyThis provides a more rigorous framework for auditing long-form media authenticity.