Researchers have introduced AV-SafetyBench, a new benchmark designed to evaluate the safety of text-to-audio-video (T2AV) generation models. This benchmark addresses the limitations of existing safety evaluations by considering risks that may arise from audio tracks or the joint interpretation of audio and visual elements. AV-SafetyBench includes a taxonomy of 13 categories across four axes and uses 5,200 prompts to assess model outputs from full AV, video-only, and audio-only perspectives. Initial evaluations of five open-source T2AV models revealed significant safety concerns, with Full-AV Unsafe Rates ranging from 25.1% to 49.4%. The analysis highlighted that audio-only and joint audio-visual risks are often missed by video-only evaluations, indicating the critical need for comprehensive safety assessments in T2AV generation. AI
IMPACT This benchmark is crucial for identifying and mitigating safety risks in emerging text-to-audio-video AI models, guiding future development towards more responsible AI.
RANK_REASON The item is a research paper introducing a new benchmark for evaluating AI models. [lever_c_demoted from research: ic=1 ai=1.0]
- Audio-only MPEG-4
- AV-Both
- AV-Joint
- AV-SafetyBench
- Cross-Modal Harm Emergence
- Full averaging scheme for differential equation with maximum
- Hugging Face
- T2AV
- Video Only
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →