Action-conditioned video world models are crucial for interactive applications, but they demand low-latency causal generation and reliable responses to game-native controls. While causal distillation has enabled one- or few-step video synthesis, extending it to interactive settings poses challenges, particularly in keeping discrete keyboard states and continuous mouse motion aligned with temporally compressed latent chunks during training and rollout.
To address this, Hugging Face researchers introduce ForgeWM, a progressive framework that transforms a bidirectional action-conditioned video generator into efficient few-step world models. The process involves four stages: domain adaptation, teacher-forced causal training, causal consistency distillation, and on-policy distribution matching with a bidirectional teacher. This yields budget-specialized students operating at steady-state denoising budgets of 1, 2, and 4 steps.
ForgeWM also supports a dual-path deployment protocol, combining latency-critical interaction with optional replay-time refinement. In this mode, the one-step student re-noises and refines its saved draft, achieving quality comparable to four-step reference while remaining roughly three times closer to the experienced trajectory than regeneration from noise.
Evaluated on paired Minecraft trajectories, ForgeWM leads in imaging quality, reference-aligned motion-profile agreement, action-sign accuracy, and mouse-control accuracy, while achieving the lowest reference LPIPS. The same recipe transfers to gamepad-controlled FPS gameplay, demonstrating broad applicability.
ForgeWM is fully open and reproducible, supporting keyboard, mouse, and gamepad control, making it a practical tool for training real-time, playable video world models.