Orange-3DV-Team/SceneFly

CaR

Compression and Retrieval: Implicit Memory Retrieval for Video World Models

Zhan Peng^1,2, Jie Ma², Huiqiang Sun¹, Chong Gao^2,3, Zhijie Xue¹, Zhiyu Pan¹, Zhiguo Cao^1*, Jun Liang^2*, Jing Li²

¹Huazhong University of Science and Technology   ²HUJING Digital Media & Entertainment Group   ³Sun Yat-sen University

^*Corresponding author

SceneFly

SceneFly is a curated video dataset organized by synthetic 3D scenes. Each selected video contains the source video, camera annotations, text prompt, and sample-level context/ground-truth metadata.

This release contains a subset of the SceneFly dataset filter, including 58 scenes, 462 videos, and 45,173 sample annotations. The complete dataset will be provided in a future release.

Usage

The dataset is provided as a split tar.gz stream. Merge the split parts and extract the dataset with:

cat SceneFly_* | tar -xzvf -

The split parts are named:

SceneFly_aa
SceneFly_ab
SceneFly_ac
...

After extraction, the dataset root will be:

SceneFly/

A checksum manifest is provided in:

manifest.csv

It records each split part name, part size, source dataset size, and SHA256 checksum.

Dataset Structure

SceneFly/
├── metadata.csv
├── AlbertMansion/
│   ├── 0/
│   │   ├── video.mp4
│   │   ├── camera.json
│   │   ├── prompt.txt
│   │   └── samples/
│   │       ├── sample000.json
│   │       ├── sample001.json
│   │       └── ...
│   ├── 1/
│   │   └── ...
│   └── ...
├── AsianArchitecture/
│   └── ...
└── ...

File Description

• video.mp4: rendered scene video.
• camera.json: camera trajectory and intrinsic annotation for the video.
• prompt.txt: text description/prompt of the scene.
• samples/sampleXXX.json: sample-level annotation file.
• metadata.csv: global metadata index for all samples.
• manifest.csv: archive-level checksum and size manifest.

Sample Annotation Fields

Each samples/sampleXXX.json contains fields such as:

• scene_name: scene identifier.
• video_name: video identifier inside the scene.
• image_width, image_height: video resolution.
• focal_length: camera focal length.
• context_start_segment, context_end_segment: context segment range.
• context_length_segments: number of context segments.
• context_segments: list of context segment IDs.
• context_start_frame: starting frame of the context window.
• context_num_frames: number of context frames.
• gt_segment: ground-truth target segment.
• gt_start_frame: starting frame of the ground-truth segment.
• gt_num_frames: number of ground-truth frames.
• overlap_score, containment_score, final_score: sample matching scores.

Metadata

metadata.csv provides one row per sample and includes the relative sample path plus the context and ground-truth fields above.

Citation

If you find this dataset useful, please cite:

@article{peng2026car,
    title={Compression and Retrieval: Implicit Memory Retrieval for Video World Models},
    author={Peng, Zhan and Ma, Jie and Sun, Huiqiang and Gao, Chong and Xue, Zhijie and Pan, Zhiyu and Cao, Zhiguo and Liang, Jun and Li, Jing},
    journal={arXiv preprint arXiv:2606.23105},
    year={2026}
}