arXivGenerative ModelsMethodImportance 8arXiv1mo ago

DreamX-World 1.0: A General-Purpose Interactive World Model

DreamX Team, Yancheng Bai, Rui Chen, Xiangxiang Chu, Rujing Dang, Hao Dou, Bingjie Gao, Qiwen Gu et al.

0 CitationsPopularity 0.8

AI analysis

TL;DR

DreamX-World 1.0 is a text/image-to-video world model that controls camera paths and events, generating long videos across photorealistic, game, and stylized domains.

Problem

Existing video generation models lack camera control, suffer from color/style drift in long generations, and fail to generalize across diverse domains.

Approach

E-PRoPE: A lightweight variant of projective positional encoding that applies camera-aware attention to spatially reduced tokens.

Convert a bidirectional video generator into an autoregressive world model using causal forcing, DMD-style distillation, and long-rollout training.

Memory-Conditioned Scene Persistence: Reuses previous views via camera-geometry-based retrieval.

Event Instruction Tuning and reinforcement learning alignment improve control performance and visual quality.

Mixed-precision DiT, residual recycling, 75%-pruned VAE decoding, and asynchronous pipeline parallelism enable fast inference.

Results & Contribution

On the 5-second basic evaluation, DreamX-World 1.0 achieves a camera-control score of 73.75 and an overall score of 84.76, outperforming HY-WorldPlay 1.5 (80.79) and LingBot-World (80.45). It achieves up to 16 FPS real-time inference on eight RTX 5090 GPUs.