Geometric View of the Diffusion Models for Video and 4D Imaging
김수현
27동 220호
0
1812
2025.08.21 14:51
| 구분 | 응용수학 |
|---|---|
| 일정 | 2025-10-17(금) 10:30~12:00 |
| 세미나실 | 27동 220호 |
| 강연자 | Jong Chul Ye (KAIST) |
| 담당교수 | 홍영준 |
| 기타 |
The recent emergence of
diffusion models has driven substantial progress in image and video processing,
establishing them as powerful generative priors. In this talk, we show that diffusion models
naturally constitute a scale-space representation of the data distribution, in
contrast to the classical scale-space representation of the samples themselves.
This geometric understanding makes conditional sampling for image
reconstruction not only more intuitive but also more firmly grounded in theory.
Then, we present recent advances in dynamic diffusion models addressing two key
challenges: high-resolution reconstruction and temporal consistency. With
geometric understanding of diffusion models, we demonstrate scalable and coherent video
generation and reconstruction with state-of-the-art performance.
