While existing camera-controllable video generation models can produce visually compelling sequences, preserving intrinsic 4D spatiotemporal coherence remains challenging. To address this limitation, we propose ChronoWorld, an "Observation--State--Reflection" framework that leverages spatiotemporal causal cues and reconstruction priors to generate globally consistent, free-view 4D scenes. Given a context video, we introduce a Spatiotemporal Epipolar Causal Attention mechanism that enforces multi-view epipolar constraints and temporal causality throughout the generation process. In addition, we develop a reconstruction-driven geometric reflection pipeline with a 4D retrieval strategy to enable dynamic self-assessment and correction of generated outputs, improving consistency and accuracy. Extensive experiments show that ChronoWorld achieves state-of-the-art performance in spatiotemporally consistent, cinematic-quality 4D scene generation, with strong generalization and high-fidelity geometry across diverse scenarios.
Figures & tables
Figure 1: Showcases of our ChronoWorld. Our method addresses the challenge of spatiotemporally consistent 4D scene generation by introducing a “Observation–State–Reflection” framework. This approach leverages spatiotemporal causal cues and reconstruction-driven geometric reflection to generate globally consistent, free-view 4D world.
Figure 2: Framework of ChronoWorld. Our method is trained to denoise target video conditioned on the camera trajectories with a Diffusion Transformer with Spatiotemporal Cues (STC-DiT), and a reconstructive multi-head decoder for predicting 4D attributes.
Figure 3: “Observation-State-Reflection” inference strategy. Leveraging a unified 4D memory, ChronoWorld enables dynamic self-assessment and correction of generated outputs, thereby enhancing the spatiotemporal consistency of 4D scenes.
Method
Type
Visual Quality
4D Visual Synchronization
Camera Accuracy
CLIP-V ↑
FID ↓
FVD-F ↓
FVD-V ↓
FVD-4D ↓
RPE-R ↓
RPE-T ↓
MotionCtrl [ 8 ]
2D
0.46
123.68
535.96
421.58
442.72
4.77
8.93
CameraCtrl [ 9 ]
2D
0.49
114.27
510.59
378.05
390.16
4.02
7.65
RecamMaster [ 38 ]
2D
0.64
101.58
403.18
214.33
198.52
3.01
6.24
TrajectoryCrafter [ 39 ]
3D
0.64
98.46
487.02
293.56
286.50
1.26
3.76
Vmem [ 40 ]
3D
0.66
113.41
389.64
252.08
264.73
2.59
6.42
Table 1: Quantitative evaluation of our method against state-of-the-art models in camera-controlled video generation, 3D-enhanced scene generation, 4D dynamic scene generation, measuring visual generation quality, and 4D visual synchronization. All tests are conducted under fair comparison settings, using identical video inputs and target camera trajectories.
Figure 4: Qualitative results of ChronoWorld. We compare ChronoWorld against SOTA baselines across diverse camera trajectories to validate its effectiveness (zoom in for details).
Method
Representation Consistency
Spatiotemporal Consistency
PSNR ↑
SSIM ↑
LPIPS ↓
Aesthetic Quality ↑
Imaging Quality ↑
Temporal Flickering ↑
Motion Smoothness ↑
Subject Consistency ↑
Background Consistency ↑
Free4D [ 13 ]
18.03
0.64
0.48
0.51
0.54
0.96
0.98
0.92
0.94
More4D [ 51 ]
18.32
0.67
0.49
0.48
0.56
0.97
0.98
0.91
0.95
DeepVerse [ 13 ]
16.37
0.56
0.61
0.50
0.51
0.92
0.95
0.93
0.91
Lyra [ 49 ]
17.01
0.58
0.72
0.41
0.43
0.88
0.81
0.87
0.90
Gen3R [ 52 ]
17.28
0.64
0.68
0.44
0.46
0.49
0.88
0.90
0.92
Table 2: Quantitative comparisons on spatiotemporal consistency of our method against state-of-the-art models in generation and reconstruction of 4D scenes.
Method
PSNR ↑
CLIP-V ↑
FVD ↓
RPEavg↓
w/o ST-ECA
18.03
0.71
184.27
1.75
w/o Plücker Emb
18.71
0.79
181.42
1.79
w/o RMD
18.59
0.77
212.79
1.68
w/o ST Reflection
17.65
0.75
196.53
1.72
Ours-Full
18.87
0.80
177.41
1.64
Table 3: The effect of core components in our method.
School of Artificial Intelligence and Automation, Huazhong University of Science and Technology · School of Computing, National University of Singapore · School of Computing, Macquarie University
MAIS&NLPR, Institute of Automation, Chinese Academy of Science, Beijing, 100190, China · University of Chinese Academy of Sciences, Beijing, 101408, China · Shanghaitech University, Shanghai, 201210, China