Game2World

Model weights for Game2World Engine: Unlocking In-the-Wild Gameplay Videos for World Model Training.

The two main rank-64 DiT LoRAs for GameCleaner are:

File	Base checkpoint	Training
game2world-without-reference-pretrain.safetensors	GameCleaner without reference pre-train	3,001 steps, prompt-only
game2world-with-reference-pretrain-refdrop20.safetensors	GameCleaner with reference pre-train	3,001 steps, 20% reference dropout

GameCleaner with reference pre-train is the recommended release and is used for the Game2World benchmark. The older 1,300/1,400-step experiments are not included.

The code, download commands, runtime loading order, and offline merge script are in Dongping-Chen/Game2World. The normal runtime loads the matching GameCleaner checkpoint first, then applies the Game2World LoRA to its DiT with scale 1.

Reference dropout ablations

gamecleaner-refdrop-final.zip contains all six final step-3001 LoRAs for 0%, 5%, 10%, 20%, 50%, and 80% reference dropout, all with reference pre-train. These are unmerged adapters. The recommended 20% model is also available separately above.

Use the same with-reference-pretrain base and load the chosen LoRA with scale 1.

hf download shuaishuaicdp/Game2World gamecleaner-refdrop-final.zip --local-dir models/Game2World
unzip models/Game2World/gamecleaner-refdrop-final.zip -d models/Game2World/refdrop

Pass the extracted file to scripts/inference.py --adapter. Each filename includes its dropout percentage; the archive README includes an inference example.