ImageWAM: Do World Action Models Really Need Video Generation, or Just Image Editing?

2026-06-17

AI Summary

This work addresses the limitations of conventional world action models that rely on video generation, which suffer from high computational overhead, redundant modeling, and error accumulation in long-horizon predictions. The authors propose ImageWAM, a novel framework that leverages a pretrained image editing model as the core mechanism for world action modeling, bypassing explicit video generation by directly modeling visual transitions from the current frame to a target frame to support action prediction. ImageWAM incorporates a denoising KV cache, flow-matching action experts, and task-instruction-guided local editing priors to focus on task-relevant visual changes. Experiments demonstrate that ImageWAM outperforms existing vision-language-action (VLA) and world action modeling (WAM) approaches in both simulated and real-world environments, achieving a sixfold reduction in FLOPs, a fourfold decrease in latency, and more focused attention on task-critical regions.

Abstract

World Action Models (WAMs) commonly rely on video generation to bridge visual world modeling and robot control. However, video-based WAMs face three coupled limitations: dense multi-frame future tokens make inference costly, full video prediction spends capacity on action-irrelevant temporal and appearance details, and long-horizon future imagination may introduce errors that mislead action prediction. These issues raise a simple question: Does world action model really need video generation? We propose ImageWAM, a simple WAM framework that repurposes pretrained image editing models for robot action prediction. In contrast to video generation, image editing provides a better-matched prior: it only needs to model a target-frame transformation, focuses on action-relevant current-to-target visual differences, and grounds task instructions to localized visual changes through edit pretraining. In practice, ImageWAM does not decode the target frame at inference time; instead, it conditions a flow-matching action expert on the KV caches produced by image-editing denoising, using them as a compact world-action context. ImageWAM outperforms standard VLA baselines and matching competitive WAMs without additional policy pretraining across different simulator and real-world experiments. It also reduces FLOPs to 1/6 and latency to 1/4 of video-based WAMs. Attention analysis further shows that editing caches focus on task-relevant change regions, supporting image editing as an effective alternative to video-based world-action modeling.

Problem

• World Action Models
• video generation
• robot control
• image editing
• action prediction

Innovation

• World Action Models
• Image Editing
• KV Cache Conditioning
• Flow-Matching Action Expert
• Efficient Robot Control