Current audio-visual AI is stuck in a short-term memory problem. While models can generate impressive clips, they consistently fail to maintain quality over longer durations. This limitation is the critical bottleneck preventing AI from creating truly cinematic, minute-long content that feels real and coherent.

Researchers have introduced LongAV-Compass, a new benchmark designed to test this exact capability. Unlike previous tools that only evaluate five-to-ten-second clips, this framework assesses minute-long generation across three distinct input types: text, images, and video. It moves beyond simple quality checks to measure how well models preserve identity, narrative flow, and audio-visual sync as time progresses.

The study reveals a stark conflict between current model capabilities and user expectations. By testing eleven major models, the researchers found that while short clips look good, coherence degrades rapidly over time. The benchmark exposes specific failures in semantic alignment and temporal consistency, showing that today’s systems struggle to sustain a unified story or visual identity for more than a few seconds.

The takeaway is clear: duration is the new frontier for generative AI. LongAV-Compass provides the first unified diagnostic tool to measure this gap, pushing the industry to prioritize long-form stability over short-term visual tricks.