ReCamMaster/ReCamMaster-WebVid-Results

ReCamMaster: Camera-Controlled Generative Rendering from A Single Video

This directory contains the comparison results of ReCamMaster and baseline methods on in-the-wild videos. Specifically, we randomly selected 1000 videos from WebVid-10M [1] as source videos and used 10 basic trajectories as model inputs. The results are stored as 0.mp4, ..., 999.mp4. The specific correspondence between videos and trajectories is as follows:

Video Range	Trajectory
0.mp4 - 99.mp4	Pan Right
100.mp4 - 199.mp4	Pan Left
200.mp4 - 299.mp4	Tilt Up
300.mp4 - 399.mp4	Tilt Down
400.mp4 - 499.mp4	Zoom In
500.mp4 - 599.mp4	Zoom Out
600.mp4 - 699.mp4	Translation Up (with rotation)
700.mp4 - 799.mp4	Translation Down (with rotation)
800.mp4 - 899.mp4	Arc Left (with rotation)
900.mp4 - 999.mp4	Arc Right (with rotation)

We have stored the source videos in the "Source Videos" folder and the generated results of three state-of-the-art baselines in the "GCD" [2], "Trajectory Attention" [3], and "DaS" [4] folders.

[1] Bain, Max, et al. "Frozen in time: A joint video and image encoder for end-to-end retrieval." Proceedings of the IEEE/CVF international conference on computer vision. 2021.

[2] Van Hoorick, Basile, et al. "Generative camera dolly: Extreme monocular dynamic novel view synthesis." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024.

[3] Zeqi Xiao, et al. "Trajectory attention for fine-grained video motion control." The Thirteenth International Conference on Learning Representations, 2025.

[4] Gu, Zekai, et al. "Diffusion as Shader: 3D-aware Video Diffusion for Versatile Video Generation Control." arXiv preprint arXiv:2501.03847 (2025).