VideoReveal · Motion and Disocclusion Control

Every movement is a revelation. Like a Jack-in-the-Box slowly turning its crank, anticipation builds around the hidden spaces inside the box (a). Although object motion inherently creates disocclusions, motion control (what moves and how) and disocclusion control (what appears in the newly revealed region) remain largely unexplored as a unified design space. Methods emphasizing disocclusion control may generate unrealistic motion (b, video generated by CogVideoX-Interp. [Zhe25]), whereas methods emphasizing motion control may generate unexpected disocclusion content (c, video generated by DiffusionAsShader [GYL∗25]).

From left to right: ground-truth video, method (b), and method (c).

In this work, we present a systematic study of the unified design space of motion and disocclusion.

01 · Benchmark

VideoReveal Benchmark

To systematically study motion and disocclusion control in image-to-video generation, we introduce VideoReveal. The benchmark contains 32 modeled scenes and 5 captured scenes spanning rigid/deformable motion and static/dynamic disoccluded content. Each sample provides the rendered sequence, tracking guidance, disocclusion masks, and associated prompts.

Modeled and captured examples from the VideoReveal benchmark.

Behind the benchmark

Building scenes
in Blender.

Take a closer look at the modeling workflow behind the synthetic scenes.

Download the modeling video

Scene modeling in Blender

Behind the benchmark

Annotating
the scenes.

Explore the annotation workflow in this video demonstration.

Download the annotation video

Scene annotation workflow

02 · Design Space

Evaluation Setup

Starting from an input image, we organize representative image-to-video methods in a 2×2 design space: motion is controlled by either text or tracking, while newly revealed content is controlled by either text or a reference image. We evaluate five representative methods across this space: CogVideoX, CogVideoX-Interp., DiffusionAsShader, DaS-Interp., and Pixel-CP.

03 · Video Results

Motion meets revelation.

Explore the same scene across guidance signals and five control methods. Use the player controls to inspect each result.

Full dataset & results

Citation

@article{Qi2026VideoReveal,
  title   = {A Study of the Design Space of Motion and Disocclusion Control in Video Generation},
  author  = {Qi, Anran and Li, Changjian and Bousseau, Adrien and Mitra, Niloy J.},
  journal = {Computer Graphics Forum},
  year    = {2026},
  note    = {Pacific Graphics 2026}
}

References

[GYL∗25] Gu, Zekai, Rui Yan, Jiahao Lu, Peng Li, Zhiyang Dou, Chenyang Si, Zhen Dong et al. , "Diffusion as shader: 3d-aware video diffusion for versatile video generation control.", Siggraph 2025.
[Zhe25] Zheng, F., "CogVideoX-Interpolation." , GitHub repository, 2025. Accessed June 1, 2026. Project Repository .