Jingxi Chen
Description: TimeRewind: Rewinding Time with Image-and-Events Video Diffusion
nerf (195) d-nerf (90) nerfies (89)
TimeRewind synthesizes the video backward into the pre-capture time with image-and-events video diffusion.
The following visual results clearly show that the TimeRewind method excels in synthesizing pre-capture videos, achieving motion that realistically aligns with the reference videos.
Unlike the standard SVD [5], which inherently synthesizes forward-time videos due to biases in its training data, our method excels in backward-time video synthesis. This distinction is particularly evident in Sequence 1, where SVD predicts the forward motion contrary to the backward movement of a horse, revealing a fundamental limitation in capturing reverse temporal dynamics. Furthermore, SVD struggles with complex motion scenarios, such as fluid and particle dynamics, because it lacks precise motion guid
Jingxi Chen