Enregistré dans:
Détails bibliographiques
Auteurs principaux: Wen, Yuxin, Wu, Jim, Jain, Ajay, Goldstein, Tom, Panda, Ashwinee
Format: Preprint
Publié: 2025
Sujets:
Accès en ligne:https://arxiv.org/abs/2504.10317
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866912326216581120
author Wen, Yuxin
Wu, Jim
Jain, Ajay
Goldstein, Tom
Panda, Ashwinee
author_facet Wen, Yuxin
Wu, Jim
Jain, Ajay
Goldstein, Tom
Panda, Ashwinee
contents We conduct an in-depth analysis of attention in video diffusion transformers (VDiTs) and report a number of novel findings. We identify three key properties of attention in VDiTs: Structure, Sparsity, and Sinks. Structure: We observe that attention patterns across different VDiTs exhibit similar structure across different prompts, and that we can make use of the similarity of attention patterns to unlock video editing via self-attention map transfer. Sparse: We study attention sparsity in VDiTs, finding that proposed sparsity methods do not work for all VDiTs, because some layers that are seemingly sparse cannot be sparsified. Sinks: We make the first study of attention sinks in VDiTs, comparing and contrasting them to attention sinks in language models. We propose a number of future directions that can make use of our insights to improve the efficiency-quality Pareto frontier for VDiTs.
format Preprint
id arxiv_https___arxiv_org_abs_2504_10317
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Analysis of Attention in Video Diffusion Transformers
Wen, Yuxin
Wu, Jim
Jain, Ajay
Goldstein, Tom
Panda, Ashwinee
Computer Vision and Pattern Recognition
We conduct an in-depth analysis of attention in video diffusion transformers (VDiTs) and report a number of novel findings. We identify three key properties of attention in VDiTs: Structure, Sparsity, and Sinks. Structure: We observe that attention patterns across different VDiTs exhibit similar structure across different prompts, and that we can make use of the similarity of attention patterns to unlock video editing via self-attention map transfer. Sparse: We study attention sparsity in VDiTs, finding that proposed sparsity methods do not work for all VDiTs, because some layers that are seemingly sparse cannot be sparsified. Sinks: We make the first study of attention sinks in VDiTs, comparing and contrasting them to attention sinks in language models. We propose a number of future directions that can make use of our insights to improve the efficiency-quality Pareto frontier for VDiTs.
title Analysis of Attention in Video Diffusion Transformers
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2504.10317