SpaceTimePilot is a method for generative rendering of dynamic scenes across space and time — letting a model synthesize any view at any moment of a 4D scene. To train and evaluate it, we introduce a temporal warping augmentation and built CamXTime: the first dataset to provide full grid rendering of dynamic 3D scenes — rendering the complete motion at every single camera pose along every trajectory. For each of our 744 dynamic scenes, we define 4 distinct camera trajectories, and for each trajectory we render the full scene animation from all 120 individual camera positions.
This produces a 120 × 120 rendering grid per scene: one axis is the camera pose within its trajectory, the other is time. This structure enables unique operations impossible with standard video datasets — bullet-time, slow motion, fast motion, non-linear temporal, and full 4D free navigation.
All videos are rendered at 1080 × 1080 · 30 fps.
Standard video datasets give you one camera path per scene. CamXTime renders every camera pose × every time step — the complete 4D rendering grid.
Horizontal axis is time, vertical axis is camera position. Draw any path on the pad and play it back — or pick a predefined pattern below.
A curated selection of scenes. Pick a space-time pattern below — each tab traces a different path through the camera × time grid.
† Work done during internship at Adobe.
To request dataset access, please fill out the form below and email chunhaoh@adobe.com so we can track your application. We're currently sharing with research groups and institutes — working through your team lead helps us process requests faster.
For the parent project, see SpaceTimePilot.
Fill out the request form →For any questions, contact Zhening Huang or Chunhao Huang.
The CamXTime dataset is part of the SpaceTimePilot project. If you use the dataset in your research, please cite:
@inproceedings{huang2026spacetimepilot,
title={SpaceTimePilot: Generative Rendering of Dynamic Scenes Across Space and Time},
author={Zhening Huang and Hyeonho Jeong and Xuelin Chen and Yulia Gryaditskaya and Tuanfeng Y. Wang and Joan Lasenby and Chun-Hao Huang},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
year={2026},
}