XPENG GWM TEAM RESEARCH

X-World 1.0

A controllable multi-camera world model.

Generate future driving observations from multi-view history and ego actions, with optional control over the scene and its appearance.

01OVERVIEW

From actions to future observations.

X-World models how a driving scene evolves under a specified action sequence. It generates synchronized multi-camera video directly in observation space.

Observe
A short history from seven synchronized cameras provides the visual context.
Condition
Specify future ego motion. Add traffic agents, road elements or appearance prompts when needed.
Generate
Predict the corresponding multi-camera future, with joint modeling across views and time.

02CAPABILITIES

Control the conditions. See the result.

Four demonstrations of action following, scene control, temporal continuity and appearance editing.

01

Ego-action control

Starting from the same scene, vary ego-motion conditions such as speed and curvature to generate straight driving, acceleration and turns.

02

Dynamic & static scene control

Condition on traffic participants, road boundaries and lane markings. Compare the generated scene with these spatial controls across views.

03

Long-horizon multi-camera generation

Generate future video chunk by chunk. Follow the continuity of motion, geometry and appearance across time and camera views.

04

Appearance editing

Keep ego actions and scene conditions fixed while text prompts change weather and time of day. All frames are generated in this C2V setting.

03METHOD

One model, multiple conditions.

A multi-view latent video generator, built on WAN 2.2 5B, combines visual history with distinct control signals.

X-World architecture: multi-view frames are encoded into video tokens, processed through view-temporal self-attention with ego-action and environment conditioning via multi-head cross-attention, then decoded into multi-view videos.
X-World model architecture · Figure 2 from the technical report.
01 / ATTENTION

Connect views and time

Alternating attention across cameras and timesteps models shared geometry, object identity and motion.

02 / CONDITIONING

Keep controls distinct

Ego actions use adaptive normalization; camera parameters use additive embeddings. Separate cross-attention branches handle dynamic agents, static elements and text.

03 / REPRESENTATION

Work in video latent space

A 3D causal VAE encodes video into a compact representation. A multi-view DiT generates future latents that are decoded back into camera observations.

From short clips to streaming generation

STAGE I

Learn controllable multi-view generation

Bidirectional training on synchronized short clips establishes action following, scene control and cross-view modeling.

Bidirectional generation · typically ~50 denoising steps

STAGE II

Generate the future chunk by chunk

A causal model uses self-forcing and distribution matching distillation. A rolling KV cache carries recent context into subsequent chunks.

Causal generation · 4 denoising steps per chunk

04CITATION

Cite X-World.

If X-World supports your research, please cite our technical report.

BIBTEX
@misc{zheng2026xworld,
  title={X-World: Controllable Ego-Centric Multi-Camera World Models for Scalable End-to-End Driving},
  author={Chaoda Zheng and Sean Li and Jinhao Deng and Zhennan Wang and Shijia Chen and Liqiang Xiao and Ziheng Chi and Hongbin Lin and Kangjie Chen and Boyang Wang and Yu Zhang and Xianming Liu},
  year={2026},
  eprint={2603.19979},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2603.19979}
}