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

2026-08-14

AI Summary

This study addresses the challenges of interaction misalignment and high latency in few-step generation for action-conditioned video world models by proposing a progressive training framework and a dual-path deployment protocol. Leveraging domain adaptation, causal training, and distillation techniques, bidirectional generators are transformed into efficient few-step models that balance low-latency interaction with replay optimization. Experiments demonstrate that this approach achieves state-of-the-art image quality and control precision in tasks such as Minecraft. Furthermore, the proposed replay refinement mechanism matches four-step generation quality while reducing trajectory deviation by threefold, effectively enabling high-quality, controllable few-step video generation.

Abstract

Action-conditioned video world models require low-latency causal generation and reliable responses to game-native controls. Although causal distillation enables one- or few-step video synthesis, extending it to interactive world models remains challenging, as discrete keyboard states and continuous mouse motion must remain aligned with temporally compressed latent chunks during causal training and autoregressive rollout. We introduce ForgeWM, a progressive framework that transforms a bidirectional action-conditioned video generator into efficient few-step world models through domain adaptation, teacher-forced causal training, causal consistency distillation, and on-policy distribution matching with a bidirectional teacher. The resulting budget-specialized students operate at steady-state denoising budgets of 1, 2, and 4 steps. ForgeWM further supports a dual-path deployment protocol combining latency-critical interaction with optional replay-time refinement, where the one-step student re-noises and refines its saved draft. On paired Minecraft trajectories, ForgeWM leads the evaluated systems in Imaging Quality, reference-aligned motion-profile agreement, action-sign accuracy, and mouse-control accuracy, while achieving the lowest reference LPIPS; the same four-stage recipe transfers to gamepad-controlled FPS gameplay. Replay-time refinement matches four-step reference quality while remaining roughly three times closer to the experienced trajectory than regeneration from noise. These results demonstrate ForgeWM's effectiveness for controllable few-step video generation.

Problem

• Action-conditioned video world models
• Causal distillation
• Few-step generation
• Interactive control alignment
• Low-latency

Innovation

• Progressive Causal Training
• Few-Step World Models
• Causal Consistency Distillation
• Replay-Time Refinement
• Action-Conditioned Generation