ForgeWM: Progressive Causal Training for Few-Step Action-Conditioned Video World Models

Summary

Action-conditioned video world models require low-latency causal generation, but extending causal distillation to interactive world models is challenging due to the need to align discrete keyboard and continuous mouse control with temporally compressed latent chunks. ForgeWM uses a progressive four-stage recipe—domain adaptation, teacher-forced causal training, causal consistency distillation, and on-policy distribution matching against a bidirectional teacher—to transform a bidirectional generator into few-step world models with steady-state denoising budgets of 1, 2, and 4 steps. It also supports a dual-path deployment with optional replay-time refinement, where the one-step student re-noises and refines its saved draft. On paired Minecraft trajectories, ForgeWM leads in imaging quality, reference-aligned motion-profile agreement, action-sign accuracy, and mouse-control accuracy, achieving the lowest reference LPIPS; the same recipe transfers to gamepad-controlled FPS. Replay-time refinement matches four-step reference quality while staying about three times closer to the experienced trajectory than regeneration from noise.