Video World Models · 2026-10-26
UniWorld: A Unified World Model for Real-Time Visual World Simulation
Xin Zhang, Yabo Chen, Shuaiwei Wang, Zixuan Duan, Haibin Huang, Feng Xu, Xuelong Li
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
UniWorld is a unified continuous visual world simulation framework that renders evolving scene geometry (image-aligned depth trajectories) together with temporally persistent appearance (reference RGB) using one shared video generator. It uses noise-aligned overlapping continuation and rolling depth latent state to preserve committed latent history exactly while bounding active depth state by window size, enabling real-time steady-state throughput.
Why it matters
The paper targets the simulator-to-renderer link needed for continuous visual world simulation, where geometry changes over time but appearance should remain consistent. UniWorld’s “shared temporal contract” makes state transitions explicit: previously generated video latents and their associated depth conditions are preserved across sampling windows while only newly requested geometry produces newly committed outputs. The authors also report bitwise-identical outputs across cached vs incremental execution paths.
Method
- Represent heterogeneous world sources via image-aligned depth (scene layout/occlusion) plus reference RGB appearance, both fed into a shared DiT-based video generation model.
- Noise-aligned overlapping continuation: project the inherited history prefix at every flow time to the prescribed noise-to-data trajectory, and commit only the newly generated suffix so latent history is exactly preserved.
- Rolling depth latent state: retain canonical depth latents, incrementally read only newly available depth positions, keep depth conditioning fixed within each sampling window, and bound active depth state by window size.
Abstract (from arXiv)
Continuous visual world simulation requires coupling evolving scene geometry with temporally persistent appearance. We present UniWorld, a unified framework that integrates depth trajectories and reference appearance within a shared video generation model. The central design maintains visual and geometric histories on a common timeline. Noise-aligned overlapping continuation places generated history at the prescribed sampling noise level and commits only newly generated latents. Rolling depth latent state preserves the corresponding geometric controls, incrementally reads newly available geometry, and keeps the conditioning fixed within each sampling window. Together, these mechanisms ensure exact preservation of latent history while bounding the active depth state by the window size. We evaluate continuous generation, paired history interventions, and long-horizon behavior across diverse scene categories. The same generator supports estimated, native, and on-demand simulator depth, while cached and incremental execution paths produce bitwise-identical outputs. Appearance changes and geometric control branches operate through the same interface without modifying committed history. The measured steady-state DiT execution path achieves real-time generation throughput, demonstrating the efficiency of UniWorld for continuous visual world simulation. Project page: https://zephipage.github.io/UniWorld.
Related papers
- WorldProbe: When Video World Models Fall Apart under Long-Horizon Spatial Exploration
- Persistent Planar Memory for Video World Models
- Momentum: Persistent Dynamics for Video World Models
- Orbia: Do Generated Videos Form a Consistent 3D World?
- L3G-World: Real-Time Infinite World Generation with Latent 3D Gaussian Memory
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: Date not supplied · Indexed by Pub-AI: 2026-10-26 UTC
UniWorld: A Unified World Model for Real-Time Visual World Simulation
Source date: Date not supplied · Original record ↗
- arxiv
- 2610.32381
Imported from the existing registry. No new publication milestone asserted.
Abstract · arxiv
Continuous visual world simulation requires coupling evolving scene geometry with temporally persistent appearance. We present UniWorld, a unified framework that integrates depth trajectories and reference appearance within a shared video generation model. The central design maintains visual and geometric histories on a common timeline. Noise-aligned overlapping continuation places generated history at the prescribed sampling noise level and commits only newly generated latents. Rolling depth latent state preserves the corresponding geometric controls, incrementally reads newly available geometry, and keeps the conditioning fixed within each sampling window. Together, these mechanisms ensure exact preservation of latent history while bounding the active depth state by the window size. We evaluate continuous generation, paired history interventions, and long-horizon behavior across diverse scene categories. The same generator supports estimated, native, and on-demand simulator depth, while cached and incremental execution paths produce bitwise-identical outputs. Appearance changes and geometric control branches operate through the same interface without modifying committed history. The measured steady-state DiT execution path achieves real-time generation throughput, demonstrating the efficiency of UniWorld for continuous visual world simulation. Project page: https://zephipage.github.io/UniWorld.