Researchers have introduced AV-FABLE, a new benchmark designed to detect AI-driven manipulations in long-form videos. Unlike existing benchmarks that focus on short clips, AV-FABLE addresses the challenge of localized manipulations within longer content by separating audio and visual timelines. Initial tests show that current audio-visual large language models (AV-LLMs) struggle with zero-shot prompting for detection and localization, but their frozen perception features retain strong forensic signals. To leverage these signals, a new method called NFSR (Native Forensics via Separated Readouts) was proposed, which uses separate audio and visual decoders to achieve superior performance in identifying manipulated segments. AI
IMPACT Introduces a new benchmark and method to improve detection of AI-generated manipulations in long-form videos, addressing limitations of current AV-LLMs.
RANK_REASON New academic paper introducing a novel benchmark and method for AI manipulation forensics. [lever_c_demoted from research: ic=1 ai=1.0]
- alphaXiv
- arXiv
- AV-FABLE
- AV-LLMs
- CatalyzeX
- DagsHub
- Gotit.pub
- Hugging Face
- National Fund for Scientific Research
- ScienceCast
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
