ActionRoPE: action as coordinates for spatially consistent control in planar-camera game world models
ActionRoPE: action as coordinates for spatially consistent control in planar-camera game world models
ActionRoPE 將動作視為座標而非特徵,並把攝影機位移寫入 RoPE 位置編碼,令同一世界位置在不同影格保持一致的位置編址,控制路徑不增加參數。在 Wan2.2-5B 上,與四種常見動作注入方法相比,角色走遠再返回時的漂移減少 66%,PSNR 提升 2.8 dB。其介面可直接遷移至 1.3B 至 33B 的四種骨幹模型,並能在未見過的美術風格中維持控制。
🎮 ActionRoPE: Actions as Coordinates for Game Worlds That Move Precisely and Stay Consistent
In top-down and isometric games, the camera follows the player,
so how far each action shifts the view is a fixed geometric quantity.
Yet existing game world models feed actions in as features
and leave the direction and distance of motion to be learned from data.
🧐 The motion is imprecise, and when you walk back the same way, the world has changed.
We propose ActionRoPE: actions as coordinates, not features.
📍 The camera displacement is written into the RoPE positional encoding, turning screen coordinates into world coordinates.
The same world location keeps the same positional address in every frame, with zero extra parameters on the control path.
🎯 We also propose CNS (Coordinate-Conditioned Noise Scheduling):
regions already visible in the first frame are denoised first, and newly revealed regions are generated with reference to them.
📊 On the same Wan2.2-5B backbone, compared with four common action-injection methods:
⦁ Walk away and come back: 66% less drift than the strongest baseline, +2.8 dB PSNR
🧩 The same interface transfers directly to four backbones from 1.3B to 33B,
and control holds in unseen art styles (oil painting, ukiyo-e, cyberpunk).
🔗
📄 arXiv: https://arxiv.org/abs/2610.25861
🏠 Project page: https://teawhite.cn/ActionRoPE/
💻 Code: https://github.com/TeaWhiteBro/ActionRoPE
🤗 Model: https://huggingface.co/teawhite/ActionRoPE
來源:HuggingFace Daily Papers(社區熱門論文) · huggingface.co