Beyond Time Shifts (ECCV 2026): a reference-free audio-visual synchronization evaluator built on Qwen2.5-Omni-3B, with SynthSync data, R-GRPO training, and the SyncBench benchmark.
benchmark reinforcement-learning multimodal generative-video video-evaluation audio-visual-synchronization omni-llm eccv2026 qwen2-5-omni reference-free-metric
-
Updated
Jul 13, 2026 - Python