---
pretty_name: Game2World Benchmark
task_categories:
  - video-to-video
language:
  - en
license: other
size_categories:
  - n<1K
---

# Game2World Benchmark

[![GitHub](https://img.shields.io/badge/GitHub-Game2World-181717?logo=github)](https://github.com/Dongping-Chen/Game2World)
[![Training Data](https://img.shields.io/badge/Hugging_Face-HUD--Video-FFD21E?logo=huggingface)](https://huggingface.co/datasets/shuaishuaicdp/hud-video)

The benchmark contains 200 five-second gameplay clips at 1280x720 and 30 FPS:

- `synthetic.zip`: 100 clips with synthetic HUD overlays.
- `in_the_wild.zip`: 100 gameplay clips collected in the wild.

Every example includes the input video, a 1280x720 clean reference image, and
the annotation used by the Game2World evaluation. The clean references for the
in-the-wild split were produced with Codex ImageGen and reviewed as part of the
benchmark preparation.

The paired training set is released separately as
[HUD-Video](https://huggingface.co/datasets/shuaishuaicdp/hud-video). It contains
96,037 clean/HUD-overlaid gameplay pairs, HUD mask videos, and per-sample JSON
renderer annotations. HUD-Video is training data; the 200 clips in this
repository are the separate evaluation benchmark.

Use `scripts/run_benchmark.sh` from the
[Game2World repository](https://github.com/Dongping-Chen/Game2World) to download
the archives, run the model, and evaluate the outputs.