BAAI releases StreamAV-Bench for evaluating streaming audio-video generation systems

AIBAAI releases StreamAV-Bench, a standardized benchmark for bridging video generation and interactive world models, which it says has already been adopted in the Vidu S2 technical report. The paper, code, dataset, project page, and leaderboard are publicly available.

AIBAAI evaluated 13 representative streaming systems under a unified protocol, comparing native joint models with cascaded pipelines. Native joint models, including PixVerse, HappyOyster, and OmniForcing, led in audio fidelity, audio-video sync, and interactive responsiveness, winning 9 of 11 interactive metrics. Cascaded pipelines combining a video model with Foley generation performed better on visual quality, semantic alignment, and long-horizon stability.

AIBAAI, Peking University, Kling AI, Tsinghua University and USTC jointly present StreamAV-Bench, which they call the first comprehensive benchmark for streaming audio-video generation. The authors say existing benchmarks evaluate only completed sequences, which makes it hard to capture how streaming systems evolve over time.

AIBAAI introduces StreamAV-Bench, which evaluates audio-video generation over 180 seconds on two tracks. The Progressive Track tests instruction adherence and temporal drift from one global prompt, while the Interactive Track adds five runtime updates at 30-second intervals to measure switching, latency, and state retention.

AIBAAI describes a benchmark of 320 expert-verified scenarios spanning 8 scene domains, 5 audio domains, 5 subject categories, and 4 visual styles. The scenarios are balanced across content complexity and update complexity. Evaluation covers 32 fine-grained dimensions grouped into 6 capability groups, including Quality and Alignment and Long-Horizon Stability.