YAML Metadata Warning:The task_categories "video-generation" is not in the official list: text-classification, token-classification, table-question-answering, question-answering, zero-shot-classification, translation, summarization, feature-extraction, text-generation, fill-mask, sentence-similarity, text-to-speech, text-to-audio, automatic-speech-recognition, audio-to-audio, audio-classification, audio-text-to-text, voice-activity-detection, depth-estimation, image-classification, object-detection, image-segmentation, text-to-image, image-to-text, image-to-image, image-to-video, unconditional-image-generation, video-classification, reinforcement-learning, robotics, tabular-classification, tabular-regression, tabular-to-text, table-to-text, multiple-choice, text-ranking, text-retrieval, time-series-forecasting, text-to-video, image-text-to-text, image-text-to-image, image-text-to-video, visual-question-answering, document-question-answering, zero-shot-image-classification, graph-ml, mask-generation, zero-shot-object-detection, text-to-3d, image-to-3d, image-feature-extraction, video-text-to-text, keypoint-detection, visual-document-retrieval, any-to-any, video-to-video, other

CineScene: Implicit 3D as Effective Scene Representation for Cinematic Video Generation

CVPR 2026: Arxiv | Project Page

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

.
├── 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.

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.
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

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.

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

  python extract_scene_from_panorama.py