Despite impressive progress in video generation, existing models remain limited to surface-level plausibility, lacking a coherent and unified understanding of the world. Prior approaches typically incorporate only a single form of world-related knowledge or rely on rigid alignment strategies to introduce additional knowledge. However, aligning the single world knowledge is insufficient to constitute a world model that requires jointly modeling multiple heterogeneous dimensions (e.g., physical commonsense, 3D and temporal consistency). To address this limitation, we introduce DreamWorld, a unified framework that integrates complementary world knowledge into video generators via a Joint World Modeling Paradigm, jointly predicting video pixels and features from foundation models to capture temporal dynamics, spatial geometry, and semantic consistency. However, naively optimizing these heterogeneous objectives can lead to visual instability and temporal flickering. To mitigate this issue, we propose Consistent Constraint Annealing (CCA) to progressively regulate world-level constraints during training, and Multi-Source Inner-Guidance to enforce learned world priors at inference. Extensive evaluations show that DreamWorld improves world consistency, outperforming Wan2.1 by 2.26 points on VBench. Code will be made publicly available at \href{https://github.com/ABU121111/DreamWorld}{\textcolor{mypink}{**Github**}}.
Figures & tables
Figure 1: Limitations of extending VideoREPA to multi-source knowledge, leading to structural implausibility and unnatural distortion. The physics score(PC) on the Videophy benchmark dropped from 29.7 to 24.1 .
Figure 2: Overview of DreamWorld. (a) Training: Expert models extract multimodal features, which are noised, concatenated, and fused via a linear layer Win+ . A subsequent layer Wout+ jointly predicts appearance and world knowledge, optimized by a World Loss with Consistency Constraint Annealing (CCA) to ensure generation fidelity. (b) Inference: We introduce Multi-Source Inner-Guidance , leveraging inherent noise features to direct the final video generation.
Method
Temporal
Semantic
Spatial
Summary
Overall Score
Subject Consistency
Background Consistency
Temporal Flickering
Motion Smoothness
Dynamic Degree
Object Class
C olor
Human Action
Multiple Objects
Scene
Relationship
Quality Score
Semantic Score
Wan2.1-T2V-1.3B
91.83
94.71
99.17
96.51
65.00
76.09
89.93
74.60
53.66
20.03
62.37
79.81
65.43
76.93
Wan2.1-T2V-1.3B(FT.)
93.59
95.81
99.36
97.21
54.08
79.90
88.57
78.98
58.20
28.55
63.31
81.26
68.47
78.71
VideoJAM(Reimpl.)
91.51
96.01
99.13
96.05
73.88
79.22
88.92
79.20
59.66
28.34
66.17
81.18
69.08
78.76
DreamWorld (Ours)
93.62
94.95
98.81
98.07
79.16
81.32
92.61
81.20
65.03
29.71
70.47
83.49
70.89
80.97
Table 1: Quantitative comparison on VBench.FT. denotes the fine-tuned version, and Reimpl. indicates our re-implementation of the method.
Table 4Table 5
Figure 3: Qualitative comparison. DreamWorld outperforms baselines by maintaining semantic realism, spatial integrity, and temporal precision. In contrast, competitor models frequently exhibit geometric penetrations and uncanny distortions
Figure 7
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Method
Camera Control
Object Control
Content Alignment
3D Consistency
Photometric Consistency
Style Consistency
Subjective Quality
Wan2.1-T2V-1.3B
23.50
52.34
66.83
72.31
74.40
34.08
52.21
Wan2.1-T2V-1.3B(FT.)
23.66
49.66
65.35
71.73
74.62
39.04
53.83
VideoJAM
22.48
57.34
59.32
71.53
75.39
25.90
44.03
DreamWorld (Ours)
22.81
51.94
64.95
73.16
77.55
35.75
52.45
Appendix
Table C.1: Quantitative comparison on WorldScore (Static Attributes). The best and second-best results are highlighted in bold and underlined .
Method
Dynamic Attributes
Summary
Overall Score
Motion Accuracy
Motion Magnitude
Motion Smoothness
Static Score
Dynamic Score
Wan2.1-T2V-1.3B
32.36
34.14
68.18
53.66
44.89
51.03
Wan2.1-T2V-1.3B(FT.)
35.49
31.35
64.76
53.98
43.86
50.95
VideoJAM
33.38
32.51
71.95
50.86
45.94
49.38
DreamWorld (Ours)
34.75
32.44
69.02
54.09
45.41
51.48
Appendix
Table C.2: Quantitative comparison on WorldScore (Dynamic Attributes, Summary and Overall Score).The best and second-best results are highlighted in bold and underlined .
Method
Subject Consistency
Background Consistency
Temporal Flickering
Motion Smoothness
Dynamic Degree
Aesthetic Quality
Imaging Quality
Object Class
Multiple Objects
Human Action
Wan2.1-T2V-1.3B
91.83
94.71
99.17
96.51
65.00
57.24
59.14
76.09
53.66
74.60
Wan2.1-T2V-1.3B(FT.)
93.59
95.81
99.36
97.21
54.08
61.13
63.27
79.90
58.20
78.98
VideoJAM
91.51
96.01
99.13
96.05
73.88
58.41
62.68
79.22
59.66
79.20
DreamWorld (Ours)
93.62
94.95
98.81
98.07
79.16
59.59
66.76
81.32
65.03
81.20
Appendix
Table C.3 : Results of VBench (Part 1/2): Consistency, Motion, and Object-level Metrics.
Method
Human Identity
Dynamic Spatial Relationship
Complex Landscape
Instance Preservation
Multi-View Consistency
Human Clothes
Dynamic Attribute
Complex Plot
Wan2.1-T2V-1.3B
64.21
26.57
16.22
80.41
14.63
99.55
22.34
10.53
Wan2.1-T2V-1.3B (FT.)
72.43
28.50
21.66
88.24
8.24
90.13
22.71
11.77
VideoJAM
72.14
24.08
15.66
90.41
6.79
97.74
24.37
10.33
DreamWorld (Ours)
71.68
24.15
14.66
86.14
12.41
99.11
26.73
10.44
Appendix
Table C.5 : VBench-2.0 results (part 1/3).
Figure 12Table 13
Figure 8: Additional qualitative comparisons. The red rectangles highlight anomalies(e.g., object penetration, unnatural disappearance, and limb distortion).