The authors first analyzed cross-frame attention in video diffusion models and found that attention responses remained dispersed across multiple regions even after recognizable scene structure had formed, meaning video attention does not reliably associate the same scene element across frames.
Distilling geometry attention into video diffusion models improves long-range consistency
Junfeng Ni and colleagues show that transferring cross-frame correspondence from a frozen geometry model into video attention reduces geometric drift on all tested backbones.
Chinese Tech
Junfeng Ni · Ziyu Zhu · Fan Jiang · Yu Liu · Chiyu Wang · Jiawei Zhang · +4 more
Tsinghua University · AMAP, Alibaba Group · State Key Laboratory of General Artificial Intelligence, BIGAI
Research Digest··3 min read
The authors propose Geometry Attention Distillation (GAD), a training-time method that distills cross-frame correspondence targets from a frozen geometry model into video diffusion attention.
Why this paper
From AMAP, Alibaba Group and 2 others
In one line
Distilling geometric correspondences from a frozen geometry model into video attention improves long-range scene and dynamic consistency without adding inference modules.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§