xiangbog/PISCO-Dataset

PISCO Training Set

Training data for PISCO: Precise Video Instance Insertion with Sparse Control (NeurIPS 2026) — the set used for training stages I–III at 832×480.

Paper · Code · Project Page · Models: PISCO-1.3B, PISCO-14B

Contents

20,590 samples, 36.6 GB:

Subset	Samples	Source
Env1 … Env27	11,816	rendered scenes from ROSE
data_v2	8,774	real videos from VPData

Each sample is six aligned videos (832×480, 49 frames, 15 fps):

Folder	Content	Training key
<subset>_Unedited/<name>.mp4	target video with the instance	video
<subset>_Edited/<name>.mp4	background video, instance removed	pisco_video
<subset>_Masked/<name>.mp4	spatial mask of the instance	pisco_video_mask
<subset>_Entity/<name>.mp4	segmented instance	pisco_reference_video
<subset>_Edited_Depth/<name>_viz.mp4	depth of the background video	pisco_depth
<subset>_Depth_Entity/<name>_viz.mp4	depth of the segmented instance	pisco_reference_depth

PISCO.json is the training index that lists all six files of every sample. Every file was fully decoded and checked. A sample appears in the index only if all six of its files are intact.

Download

The videos are packed as tar shards in shards/. Each shard is self-contained and holds all six files of its samples.

huggingface-cli download xiangbog/PISCO-Dataset --repo-type dataset --local-dir PISCO-Dataset
mkdir -p dataset/PISCO
for f in PISCO-Dataset/shards/*.tar; do tar -xf "$f" -C dataset/PISCO; done
cp PISCO-Dataset/PISCO.json dataset/PISCO/

Then follow the training instructions in the code repository.

License

The data is derived from ROSE and VPData, so their licenses and terms of use apply. It is released for research purposes only.

Citation

@article{gao2026pisco,
  title   = {PISCO: Precise Video Instance Insertion with Sparse Control},
  author  = {Gao, Xiangbo and Li, Renjie and Chen, Xinghao and Wu, Yuheng and Feng, Suofei and Yin, Qing and Tu, Zhengzhong},
  journal = {arXiv preprint arXiv:2602.08277},
  year    = {2026}
}