Research

Paper

TESTING February 24, 2026

Echoes Over Time: Unlocking Length Generalization in Video-to-Audio Generation Models

Authors

Christian Simon, MAsato Ishii, Wei-Yao Wang, Koichi Saito, Akio Hayakawa, Dongseok Shim, Zhi Zhong, Shuyang Cui, Shusuke Takahashi, Takashi Shibuya, Yuki Mitsufuji

Abstract

Scaling multimodal alignment between video and audio is challenging, particularly due to limited data and the mismatch between text descriptions and frame-level video information. In this work, we tackle the scaling challenge in multimodal-to-audio generation, examining whether models trained on short instances can generalize to longer ones during testing. To tackle this challenge, we present multimodal hierarchical networks so-called MMHNet, an enhanced extension of state-of-the-art video-to-audio models. Our approach integrates a hierarchical method and non-causal Mamba to support long-form audio generation. Our proposed method significantly improves long audio generation up to more than 5 minutes. We also prove that training short and testing long is possible in the video-to-audio generation tasks without training on the longer durations. We show in our experiments that our proposed method could achieve remarkable results on long-video to audio benchmarks, beating prior works in video-to-audio tasks. Moreover, we showcase our model capability in generating more than 5 minutes, while prior video-to-audio methods fall short in generating with long durations.

Metadata

arXiv ID: 2602.20981
Provider: ARXIV
Primary Category: cs.CV
Published: 2026-02-24
Fetched: 2026-02-25 06:05

Related papers

Raw Data (Debug)
{
  "raw_xml": "<entry>\n    <id>http://arxiv.org/abs/2602.20981v1</id>\n    <title>Echoes Over Time: Unlocking Length Generalization in Video-to-Audio Generation Models</title>\n    <updated>2026-02-24T15:01:39Z</updated>\n    <link href='https://arxiv.org/abs/2602.20981v1' rel='alternate' type='text/html'/>\n    <link href='https://arxiv.org/pdf/2602.20981v1' rel='related' title='pdf' type='application/pdf'/>\n    <summary>Scaling multimodal alignment between video and audio is challenging, particularly due to limited data and the mismatch between text descriptions and frame-level video information. In this work, we tackle the scaling challenge in multimodal-to-audio generation, examining whether models trained on short instances can generalize to longer ones during testing. To tackle this challenge, we present multimodal hierarchical networks so-called MMHNet, an enhanced extension of state-of-the-art video-to-audio models. Our approach integrates a hierarchical method and non-causal Mamba to support long-form audio generation. Our proposed method significantly improves long audio generation up to more than 5 minutes. We also prove that training short and testing long is possible in the video-to-audio generation tasks without training on the longer durations. We show in our experiments that our proposed method could achieve remarkable results on long-video to audio benchmarks, beating prior works in video-to-audio tasks. Moreover, we showcase our model capability in generating more than 5 minutes, while prior video-to-audio methods fall short in generating with long durations.</summary>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.CV'/>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.AI'/>\n    <published>2026-02-24T15:01:39Z</published>\n    <arxiv:comment>Accepted to CVPR 2026</arxiv:comment>\n    <arxiv:primary_category term='cs.CV'/>\n    <author>\n      <name>Christian Simon</name>\n    </author>\n    <author>\n      <name>MAsato Ishii</name>\n    </author>\n    <author>\n      <name>Wei-Yao Wang</name>\n    </author>\n    <author>\n      <name>Koichi Saito</name>\n    </author>\n    <author>\n      <name>Akio Hayakawa</name>\n    </author>\n    <author>\n      <name>Dongseok Shim</name>\n    </author>\n    <author>\n      <name>Zhi Zhong</name>\n    </author>\n    <author>\n      <name>Shuyang Cui</name>\n    </author>\n    <author>\n      <name>Shusuke Takahashi</name>\n    </author>\n    <author>\n      <name>Takashi Shibuya</name>\n    </author>\n    <author>\n      <name>Yuki Mitsufuji</name>\n    </author>\n  </entry>"
}