Interactive world generation requires both accurate local control and stable scene consistency over long rollouts. Existing methods improve local generation through camera or geometry conditioning and extend historical context with memory mechanisms. However, long-horizon exploration remains challenging as target-view guidance becomes incomplete and inconsistencies accumulate across repeated scene updates. We present GSExplorer, a generation--reconstruction framework built around a persistent 3D Gaussian world state. At each step, the current 3DGS scene provides multimodal rendering conditions for a condition-driven video generator, while camera geometry is injected through a separate Relative Ray Encoding branch. Generated videos are then fed back to a SLAM-based reconstruction backend, where inconsistency filtering prevents conflicting content from corrupting the scene and cross-chunk bundle adjustment improves spatial alignment across generation steps. We further design a data-construction pipeline that reproduces realistic 3DGS degradation during scene extrapolation, yielding 293K training pairs of degraded multimodal renderings and real target videos. GSExplorer achieves the best overall performance among the compared methods on both single-step and three-step WorldScore evaluation, with strong camera control and consistency across both settings, including a Camera Control score of 99.07 in single-step extrapolation. On 481-frame leave-and-return rollouts, it also shows strong revisit consistency, demonstrating the value of maintaining a reliable 3D world state for continued exploration.
Figures & tables
Figure 1: Overview of GSExplorer. Starting from a 3DGS scene initialized from a single image, we render multimodal scene conditions along the planned camera trajectory at each step. These conditions, together with camera geometry and text prompts, guide the condition-driven generator to produce a completed video. The generated video is then fed back to the SLAM-based world update, where inconsistency filtering enables selective scene updates and cross-chunk bundle adjustment aligns current and historical views. The updated 3DGS world state then provides reliable conditions for the next step, forming a generation–reconstruction loop.
Figure 2: Architecture of the condition-driven video generator. Multimodal 3DGS renderings are encoded as scene conditions and injected into the diffusion backbone through one input projection and five residual projections. Camera intrinsics and poses define token-wise ray geometry, which is incorporated through a separate RRE attention branch in every DiT block using geometry-aware Q/K/V and output transforms. Text prompts follow the original Wan cross-attention pathway.
Figure 3: Training-pair construction pipeline. First, long sequences are filtered and sampled into 57-frame clips containing 8 context frames and 49 target frames. Next, we reconstruct the full sequence with 3DGS-SLAM while forcing Gaussian insertion at the first target frame. Together, these steps improve target-view alignment while preserving the desired context-limited degradation. We then retain only Gaussians born within the context frames, recovering the desired context-limited degradation. Finally, the retained 3DGS is rendered at the target poses to produce multimodal degraded conditions paired with the real target frames.
Method
3D Consist.
Photo Consist.
Style Consist.
Camera Ctrl.
Object Ctrl.
Content Align.
Subjective Qual.
Avg.
WonderWorld
82.85
67.86
55.79
92.32
47.63
79.09
69.03
70.65
AETHER
79.84
58.68
72.09
57.44
52.26
28.06
41.11
55.64
Uni3C
78.59
85.48
88.32
62.94
45.83
47.40
57.00
66.51
Voyager
56.00
80.68
72.89
45.92
57.69
48.36
44.74
58.04
FantasyWorld
83.31
86.11
94.22
57.05
34.46
38.45
57.40
64.43
Wan2.2-5B-Camera
78.19
83.07
84.16
75.83
62.59
63.12
47.46
70.63
Table 1: Results on WorldScore for single-step extrapolation. GSExplorer achieves the best overall performance with strong controllability and consistency. The first , second , and third best results are highlighted accordingly.
Method
3D Consist.
Photo Consist.
Style Consist.
Camera Ctrl.
Object Ctrl.
Content Align.
Subjective Qual.
Avg.
WonderWorld
63.70
3.22
35.95
96.28
38.61
97.10
72.46
58.19
AETHER
63.97
33.11
61.99
4.43
34.78
33.69
35.09
38.15
Uni3C
79.95
46.78
71.43
8.69
34.28
77.88
51.12
52.88
Voyager
13.82
9.52
61.34
0.00
49.23
64.10
39.21
33.89
FantasyWorld
74.83
60.61
82.02
11.24
31.96
77.20
50.46
55.47
Wan2.2-5B-Camera
64.73
80.05
71.07
49.38
51.62
39.57
34.97
55.91
Table 2: Results on WorldScore for three-step extrapolation. GSExplorer maintains the best overall performance under multi-step generation, with strong camera control and scene consistency.
Figure 4: Qualitative WorldScore results across one-, two-, and three-step extrapolation. P and C denote the text prompt and camera instruction. GSExplorer accurately follows both controls while maintaining high visual quality and coherent spatial structure over multiple generation steps.
Metric
UniWorld-View
DreamX-World
WonderWorld
Spatia
GSExplorer (Ours)
PSNR ↑
11.73
12.02
15.09
17.86
19.85
SSIM ↑
0.375
0.417
0.592
0.565
0.623
LPIPS ↓
0.544
0.649
0.427
0.415
0.369
Match Ratio ↑
0.210
0.131
0.431
0.423
0.481
Table 3: Long-horizon revisit evaluation. GSExplorer achieves the best revisit consistency.
Figure 5: Qualitative comparison of long-horizon revisit results. GSExplorer better preserves the content and details of the starting view after long-range exploration.
Method
3D Consist.
Photo Consist.
Style Consist.
Camera Ctrl.
Object Ctrl.
Content Align.
Subjective Qual.
Avg.
w/o RRE-Attn
88.14
84.51
93.80
96.03
65.84
73.70
65.56
81.08
w/o pose
88.64
85.74
93.68
95.58
65.72
73.41
65.75
81.22
w/o depth
87.13
88.02
93.84
97.65
65.53
74.30
66.20
81.81
w/o opacity
90.06
88.91
94.06
98.82
64.34
73.06
65.87
82.16
w/o pose & opacity
87.66
85.37
93.80
95.02
64.18
72.59
65.30
80.56
Ours
90.83
88.47
94.40
99.07
65.91
74.70
66.59
82.85
Table 4: Ablation of conditioning modalities and pose injection. All conditioning modalities contribute to the final performance, while separate pose encoding through RRE achieves the best overall result. “w/o RRE-Attn” injects pose together with the rendered conditions instead of using the separate RRE attention branch.
Method
WorldScore Three-step Extrapolation
Revisit Evaluation
3D Consist.
Photo Consist.
Style Consist.
Camera Ctrl.
Object Ctrl.
Content Align.
Subjective Qual.
Avg.
PSNR
Match Ratio
w/o filter
72.78
73.97
78.16
68.15
56.09
52.58
54.23
65.14
17.15
0.378
w/o BA
75.54
81.16
81.43
78.50
55.69
51.79
58.29
68.91
17.41
0.406
Ours
77.50
84.51
84.36
85.17
56.25
52.82
60.34
71.56
19.85
0.481
Table 5: Ablation of persistent world-state update. Both observation filtering and cross-chunk bundle adjustment improve multi-step extrapolation and revisit consistency.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Dataset
Training pairs
RealEstate10K
231,644
DL3DV
14,279
ACID
9,983
ScanNet
38,037
Total
293,943
Appendix
Table 6: Composition of the final training set.
Figure 6: Additional qualitative results on WorldScore. GSExplorer follows the specified text and camera controls while maintaining coherent scene appearance and geometry during extrapolation.
Figure 7: Additional qualitative comparison of long-horizon revisit results. After extended leave-and-return exploration, GSExplorer better preserves the scene content and spatial layout of the initial view.