HUD-Video

Disclaimer — work in progress: We are actively scaling HUD-Video to approximately one million clips. The before_overlay source videos in the current release were collected and filtered on a best-effort basis to find gameplay footage without UI. However, they are not guaranteed to be completely UI-free: some clips may still contain original in-game HUD elements, subtitles, captions, watermarks, menus, or other screen-space overlays. Users who require strictly clean targets should apply additional filtering or manual verification.

HUD-Video is the paired gameplay training dataset released with Game2World Engine. It is designed for training GameCleaner to remove heads-up displays (HUDs), menus, and other screen-space overlays while preserving the underlying gameplay scene and motion.

Each sample starts from a clean gameplay clip. A synthetic HUD is rendered over it to produce the model input, together with a HUD mask and a JSON record describing how the overlay was generated. This Hugging Face repository mirrors the 17 source archives released on ModelScope; the archive byte sizes and SHA-256 checksums were verified against the source release.

Dataset at a glance

Item	Count / size
Paired samples	96,037
Clean videos (before_overlay)	96,037
HUD-overlaid videos (after_overlay)	96,037
HUD mask videos (masks)	96,037
Per-sample JSON annotations (annotations)	96,037
Total MP4 files	288,111 (899.305 GB)
Total JSON files	96,037 (2.893 GB)
ZIP archives	17 (902.294 GB including ZIP overhead)

The released clips are 1280×720, 30 FPS, and 150 frames / 5 seconds. The sample distribution is:

View type	Samples	Share
Egocentric	17,128	17.83%
Non-egocentric	78,909	82.17%

Volumes 001–002 contain egocentric samples, volume 003 is mixed, and volumes 004–017 contain non-egocentric samples.

Sample organization

The archives contain both videos and metadata. Every archive has the same four top-level directories:

before_overlay/<view_type>/<source_group>/<clip_id>.mp4
after_overlay/<view_type>/<source_group>/<clip_id>.mp4
masks/<view_type>/<source_group>/<clip_id>_mask.mp4
annotations/<view_type>/<source_group>/<clip_id>.json

The matching relative path identifies one sample:

• before_overlay: clean gameplay clip and training target.
• after_overlay: the same clip with a synthetic HUD, used as the model input.
• masks: video mask for the rendered HUD.
• annotations: renderer metadata for the same pair.

Each JSON annotation records the video dimensions and timing, taxonomy version and labels, random seed, requested and realized HUD layout, opacity, selected static and dynamic assets, transient events, aesthetic-matching settings, and renderer options. Some provenance fields contain the original producer-side absolute paths; use the matching archive-relative paths above when loading this release.

Download and extract

Install the Hugging Face CLI and download the repository:

pip install -U "huggingface_hub[cli]"
hf download shuaishuaicdp/hud-video \
  --repo-type dataset \
  --local-dir ./hud-video

Each ZIP is a separate shard and has no enclosing shard directory. Extract each archive into its own directory so that files from different shards are not merged:

mkdir -p data/HUD-Video

for archive in ./hud-video/HUD-Video-*-of-017.zip; do
  shard="$(basename "${archive%.zip}")"
  mkdir -p "data/HUD-Video/${shard}"
  unzip -q "${archive}" -d "data/HUD-Video/${shard}"
done

The resulting layout is:

data/HUD-Video/
├── HUD-Video-001-of-017/
│   ├── before_overlay/
│   ├── after_overlay/
│   ├── masks/
│   └── annotations/
├── ...
└── HUD-Video-017-of-017/

To build the paired training manifest used by Game2World, clone the official code repository and run:

python scripts/prepare_hud_pairs.py \
  --data-root data/HUD-Video \
  --output data/hud_pairs.csv

See the official training guide for the remaining preparation and training steps.

Intended use and limitations

HUD-Video is intended for research on gameplay video editing, HUD removal, paired video-to-video learning, and world-model data preparation. The HUDs are synthetically rendered, so their distribution does not cover every real game's interface or every recording condition. The release does not include player-action annotations.

The Game2World code repository is MIT-licensed. A standalone license for the HUD-Video media is not stated in the source dataset release, so users should consult the source release and the applicable rights for the underlying gameplay and UI assets before redistribution or commercial use.

Citation

@misc{shen2026game2worldengineunlockinginthewild,
  title        = {Game2World Engine: Unlocking In-the-Wild Gameplay Videos for World Model Training},
  author       = {Wenxuan Shen and Dongna Jin and Dongping Chen},
  year         = {2026},
  eprint       = {2608.24680},
  archivePrefix = {arXiv},
  primaryClass = {cs.CV},
  url          = {https://arxiv.org/abs/2608.24680}
}