Скачать или смотреть T2AV-Compass: Benchmark for T2AV Models

In this AI Research Roundup episode, Alex discusses the paper: 'T2AV-Compass: Towards Unified Evaluation for Text-to-Audio-Video Generation' T2AV-Compass tackles a major problem in evaluating text-to-audio-video generation: existing benchmarks are fragmented and often miss cross-modal semantic alignment, temporal synchronization, instruction following, and perceptual realism. The paper introduces a unified, diagnostic benchmark with 500 semantically dense prompts designed to stress realistic, compositional generation. Prompts are built through a taxonomy-driven pipeline using community prompt aggregation, embedding-based semantic deduplication, long-tail sampling, and LLM plus human auditing, then extended with 100 video-inversion prompts from YouTube clips with dense captioning and verification. It also proposes a dual-level evaluation framework combining objective signal-level metrics and more diagnostic checks to better pinpoint where T2AV systems fail under complex prompts. Paper URL: https://arxiv.org/abs/2512.21094 #AI #MachineLearning #DeepLearning #TextToAudioVideo #MultimodalGeneration #Benchmark #Evaluation

Resources:
GitHub: https://github.com/NJU-LINK/T2AV-Comp...