北京雁栖湖应用数学研究院

董昂 , 杨登程 , 赵鑫

Shuoli Liu

2026年08月01日 11:00 至 12:00

Shuangqing-B725

Zoom 928 682 9093 (BIMSA)

When a video diffusion model generates a person frame by frame, small identity errors compound. After a few minutes the face drifts and the character quietly becomes someone else. Existing systems either reset identity at shot boundaries or tolerate this drift. The longest publicly demonstrated continuous character video, Gloria (CVPR 2026), reaches about 10 minutes. We start with a localization question: where exactly does identity live inside a diffusion transformer? By injecting donor features into value streams versus key streams across four backbones, we find that identity is localized in mid layer values. A carrier write there reaches 0.999 cosine similarity to the donor, while an equal energy write to the orthogonal complement stays below 0.33. Prior systems condition on a reference image at the model boundary. We go inside the transformer: localize identity to mid layer value streams, and periodically rewrite only that carrier at a controlled strength along a continuous homotopy path. We call this Yau-Yau Homotopy, after the observation separated filtering structure of Yau and Yau (2008). Theorem 1 guarantees that periodic reanchoring inside a contractive corridor keeps identity error bounded uniformly across the entire horizon. Theorem 2 shows the system's fixed point moves smoothly with write strength, with an explicit bound on how fast you can change the control before identity breaks. One identity, continuously generated for 1,200 seconds, 20 minutes, with ArcFace cosine holding near 0.97 throughout. This is 2× the duration of Gloria, the previous state of the art, across a 41 system audit.

Shuoli Liu is a PhD candidate at Princeton working on decision theory: the mathematical foundations of risk and uncertainty. He applies this to learning theory, preference alignment, and theory AI risks.