KwaiVGI/Context-as-Memory-Dataset

Context as Memory: Scene-Consistent Interactive Long Video Generation with Memory Retrieval

SIGGRAPH Asia 2025

[Project page] [ArXiv] [Dataset]

File Structure

To prepare the dataset for use, merge the parts into a single zip file using the following command:

cat Context-as-Memory-Dataset_* > Context-as-Memory-Dataset.zip

After extracting Context-as-Memory-Dataset.zip, the dataset will be organized as follows:

Context-as-Memory-Dataset
├── frames
│   ├── AncientTempleEnv_0
│   │   ├── 0000.png
│   │   ├── 0001.png
│   │   ├── 0002.png
│   │   └── ...
│   ├── AncientTempleEnv_1
│   │   ├── 0000.png
│   │   ├── 0001.png
│   │   ├── 0002.png
│   │   └── ...
│   └── ...
│  
├── jsons
│   ├── AncientTempleEnv_0.json
│   ├── AncientTempleEnv_1.json
│   └── ...
│
├── overlap_labels
│   ├── AncientTempleEnv_0
│   │   ├── 0.json
│   │   ├── 1.json
│   │   ├── 2.json
│   │   └── ...
│   ├── AncientTempleEnv_1
│   │   ├── 0.json
│   │   ├── 1.json
│   │   ├── 2.json
│   │   └── ...
│   └── ...
│  
└── captions.txt

Explanation of Dataset Parts

• frames/: 100 subdirectories, each containing 7,601 video frame images.
• jsons/: 100 JSON files, each storing the camera pose (position + rotation) of every frame in the corresponding long video.
• overlap_labels/: 100 subdirectories, each containing 7,601 JSON files, where each file records the indices of overlapping frames corresponding to that frame.
• captions.txt: Captions annotated for a segment of a long video, from a given starting frame to an ending frame.
• We also provide a simple code file, tools.py, which can convert (x, y, z, yaw, pitch) into RT, and can also select a specific frame as the reference frame to align the RT of other frames to its coordinate system.