File size: 4,424 Bytes
9828c7b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
21b9000
 
 
 
ae70909
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3e8a25f
ae70909
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9574f78
ae70909
 
21b9000
ae70909
 
 
 
 
 
 
 
 
21b9000
ae70909
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
---
task_categories:
- video-generation
- text-to-video
language:
- en
tags:
- video
- synthetic
- cinematic
- panoramic image
pretty_name: Scene-Decoupled Video Dataset
size_categories:
- 150G<n<200G
arxiv: 2602.06959
---

# CineScene: Implicit 3D as Effective Scene Representation for Cinematic Video Generation 
CVPR 2026: [Arxiv](https://arxiv.org/pdf/2602.06959) | [Project Page](https://karine-huang.github.io/CineScene/) 

## Scene-Decoupled Video Dataset

TL;DR: The Scene-Decoupled Video Dataset, introduced in CineScene, is a large-scale synthetic dataset for **video generation with decoupled scene**, which encompasses diverse scenes, subjects, and camera movements. This dataset contains camera trajectories, equirectangular panorama (scene image), and videos with/without dynamic subject. The data is organized into "With Human" (whuman) and "Without Human" (wohuman) categories, while panoramas are scene-decoupled and shared across both.

## 1. Directory Tree

```text
.
├── camera/                           # Camera trajectories and metadata
│   ├── whuman/                       # Sequences containing human characters
│   │   └── <scene_id>/               # e.g., scene1_3x3_loc1_scene_AncientTempleEnv/
│   │       └── <scene_id>_cam.json   # Camera parameters
│   └── wohuman/                      # Sequences with environment only
│       └── <scene_id>/
│           └── <scene_id>_cam.json

├── panorama/                         # Scene-decoupled environment maps
│   └── <scene_id>/                   # Shared between whuman and wohuman
│       └── <scene_id>_pano.jpeg      # 360° Equirectangular panoramic image

└── video/                            # Rendered video sequences (MP4)
    ├── whuman/                       # Videos with human characters
    │   └── <scene_id>/
    │       ├── <scene_id>_01_24mm.mp4 # Sub-sequences (01, 02, etc.)
    │       ├── <scene_id>_02_24mm.mp4
    │       └── ...
    └── wohuman/                      # Videos without human characters
        └── <scene_id>/
            ├── <scene_id>_01_24mm.mp4
            ├── ...

```



## 2. Dataset Statistics

*   **Total Scale**: 46,816 videos.
*   **Scenes**: 3,400 scenes (comprising both *whuman* and *wohuman* scenes) across 35 high-quality 3D environments.
*   **Trajectories**: 46,816 camera paths (7 distinct camera trajectories per scene).
*   **Panorama**: 360° Equirectangular images for every scene, providing a complete background reference for scene conditioning.

| Property | Value |
| :--- | :--- |
| **Video Resolution** | 672 x 384 |
| **Frame Count** | 81 frames per video |
| **Frame Rate** | 15 FPS |
| **View Change Range** | Up to 75° |
| **Decoupled Scene** | 360° Equirectangular (Panorama) |
| **Panorama Resolution** | 2048 x 1024 |


## 3. Dataset Construction

We follow the asset collection pipeline established by **RecamMaster**, but introduce three significant enhancements to support more complex generative tasks:

1.  **Decoupled Scenes**: We provide static 360° panoramic images (Equirectangular) for every scene. This allows for explicit background conditioning and facilitates novel view synthesis from any angle.
2.  **Extended Camera Range**: Our dataset covers significantly larger view changes (approx. **75°**) compared to the 5–60° range provided in [previous datasets](https://huggingface.co/datasets/KlingTeam/MultiCamVideo-Dataset).
3.  **Paired Subject/Background Data**: Every scene includes both "with-subject" (*whuman*) and "background-only" (*wohuman*) video sequences. This paired data is ideal for training models on subject-background decoupling, motion transfer, and cinematic composition.

## 4. useful script

- download

```bash
sudo apt-get install git-lfs
git lfs install
git clone https://huggingface.co/datasets/KlingTeam/Scene-Decoupled-Video-Dataset
cat Scene-Decoupled-Video-Dataset.part* > Scene-Decoupled-Video-Dataset.tar.gz
tar -xvf Scene-Decoupled-Video-Dataset.tar.gz
```

- camera visualization

  To visualize the camera, please refer to [here.](https://huggingface.co/datasets/KlingTeam/MultiCamVideo-Dataset/blob/main/vis_cam.py)

- Perspective Projection
To extract perspective frames from the panoramic images:

  ```
  python extract_scene_from_panorama.py
  ```