Wonder: A Real-Time Camera-Controllable Video World Model Done Better

Forum topic · 小凯 · 2026-07-30

Summary

Wonder is a general-purpose video world model for real-time, camera-controllable world exploration, presented by researchers from JHU and Adobe Research (arXiv 2607.26037). Given an image or a conditional video, Wonder builds a playable world where users navigate interactively by moving the camera, discovering unseen regions and revisiting previously observed areas over long horizons. The system combines three co-designed components: (1) a novel camera conditioning scheme based on a dense coordinate field, whose renderings provide spatially aligned motion and orientation cues so the model interprets camera motion directly as visual evidence; (2) an efficient sparse-attention memory mechanism that selectively retrieves a small set of relevant context tokens regardless of growing context length; and (3) techniques to correct self-forcing-style distillation, improving the student model's control-signal fidelity while preserving the teacher's diverse generation modes and long-term memory. Wonder synthesizes diverse minute-long video at 16 FPS while maintaining coherent geometry, appearance, and dynamics, and naturally supports video-conditioned generation for re-shooting existing dynamic scenes in real time.

Overview

Field: Computer Vision Authors: Jiacong Xu, Hanwen Jiang, Zhixin Shu, Kalyan Sunkavalli, Vishal M. Patel, Yiqun Mei arXiv: 2607.26037

Abstract

We present Wonder, a general-purpose video world model for real-time, camera-controllable world exploration. Given an image or a conditional video, Wonder constructs a playable world where users can navigate interactively by moving the camera, discovering unseen regions, and revisiting previously observed areas in real time and over a long-term horizon.

Achieving this capability requires a system-level co-design of control method, memory mechanism, and training strategy:

Camera conditioning: A novel dense coordinate field whose renderings provide spatially aligned motion and orientation cues, allowing the model to interpret camera motion directly as visual evidence.
Sparse-attention memory: An efficient memory mechanism enabling fast and precise retrieval over a growing generation context, letting the model selectively attend to a small subset of relevant context tokens regardless of the actual context length.
Corrected self-forcing-style distillation: Multiple techniques that fix the distillation pipeline, improving the student model's ability to respect control signals while preserving the teacher model's diverse generation modes and long-term memory.

Together, these components allow Wonder to synthesize diverse, minute-long video at 16 FPS while maintaining coherent geometry, appearance, and dynamics over long-horizon rollouts. Beyond image-to-video generation, Wonder naturally supports video-conditioned generation, enabling existing dynamic scenes to be re-shot in real time.

Key Takeaways

Real-time, interactive world exploration from a single image or video
Dense coordinate field provides spatially aligned control cues
Sparse-attention memory scales to minute-long generations without slowdown
Distillation corrections preserve diversity and long-term memory
16 FPS generation of minute-scale, coherent video

---

*Auto-collected on 2026-07-30*