Organizations: State Key Laboratory of Complex and Critical Software Environment, Beijing, China · School of Computer Science and Engineering, Beihang University, China · School of Artificial Intelligence, Beihang University, China
We present TokenSplat, a feed-forward framework for joint 3D Gaussian reconstruction and camera pose estimation from unposed multi-view images. At its core, TokenSplat introduces a Token-aligned Gaussian Prediction module that aligns semantically corresponding information across views directly in the feature space. Guided by coarse token positions and fusion confidence, it aggregates multi-scale contextual features to enable long-range cross-view reasoning and reduce redundancy from overlapping Gaussians. To further enhance pose robustness and disentangle viewpoint cues from scene semantics, TokenSplat employs learnable camera tokens and an Asymmetric Dual-Flow Decoder (ADF-Decoder) that enforces directionally constrained communication between camera and image tokens. This maintains clean factorization within a feed-forward architecture, enabling coherent reconstruction and stable pose estimation without iterative refinement. Extensive experiments demonstrate that TokenSplat achieves higher reconstruction fidelity and novel-view synthesis quality in pose-free settings, and significantly improves pose estimation accuracy compared to prior pose-free methods. Project page: https://kidleyh.github.io/tokensplat/.
Figures & tables
Figure 1: Overview of TokenSplat. TokenSplat performs feed-forward 3D Gaussian reconstruction and camera pose estimation from unposed images. A shared ViT encoder extracts image tokens, which are processed by the Canonical Scene Decoder and the Asymmetric Dual-Flow Decoder (ADF-Decoder). The fused tokens are then used by the Token-aligned Gaussian Prediction module and the camera pose head to generate dense 3D Gaussians and accurate poses.
Method
RE10K (4 views)
RE10K (8 views)
ScanNet (4 views)
ScanNet (8 views)
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
Pose- required
MVSplat
23.82
0.792
0.201
23.96
0.802
0.164
24.48
0.854
0.225
23.11
0.761
0.296
FreeSplat
24.99
0.814
0.162
25.20
0.829
0.179
27.23
0.873
0.196
24.64
0.819
0.241
Pose- free
NoPoSplat
24.87
0.813
0.169
25.01
0.832
0.163
27.10
0.864
0.189
24.26
0.796
0.269
VicaSplat
24.65
0.795
0.206
24.50
0.806
0.164
26.46
0.837
0.214
22.95
0.765
0.306
SPFSplat
24.98
0.819
0.166
25.18
0.828
0.169
27.17
0.873
0.185
24.15
0.793
0.263
Table 1: Quantitative results of NVS on RE10K with varying reference views (left) and cross-dataset generalization to ScanNet (right). The best and second-best values are highlighted.
Method
3 views
10 views
28 views
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
Pose- required
MVSplat
24.98
0.784
0.228
23.12
0.768
0.236
22.26
0.714
0.336
FreeSplat
25.63
0.804
0.218
25.38
0.815
0.189
24.30
0.798
0.249
Pose- free
NoPoSplat
25.55
0.803
0.213
24.21
0.811
0.198
23.21
0.764
0.268
VicaSplat
24.54
0.786
0.226
23.73
0.760
0.268
22.37
0.709
0.316
SPFSplat
25.85
0.788
0.229
24.63
0.773
0.226
23.54
0.775
0.261
Table 2: Quantitative results of NVS on ScanNet under varying numbers of views. The best and second-best values are highlighted.
Figure 2: Qualitative comparison on RE10K and ScanNet under varying numbers of reference views.
Figure 3: Cross-dataset generalization from RE10K to ScanNet.
Method
RE10K (4 views)
RE10K (8 views)
ScanNet (4 views)
ScanNet (8 views)
ATE ↓
RPE-t ↓
RPE-r ↓
ATE ↓
RPE-t ↓
RPE-r ↓
ATE ↓
RPE-t ↓
RPE-r ↓
ATE ↓
RPE-t ↓
RPE-r ↓
NoPoSplat
0.032
0.070
1.923
0.123
0.135
1.757
0.139
0.259
2.135
0.208
0.268
4.098
VicaSplat
0.027
0.057
1.392
0.021
0.031
0.793
0.095
0.234
2.735
0.163
0.197
3.162
SPFSplat
0.036
0.078
3.113
0.037
0.047
2.117
0.075
0.201
2.658
0.121
0.173
3.676
Anysplat ∗
0.028
0.051
1.515
0.020
0.025
0.578
0.092
0.238
1.902
0.098
0.150
2.016
Anysplat
0.033
0.055
1.624
0.024
0.029
0.605
0.099
0.268
1.995
0.101
0.157
2.114
Table 3: Quantitative results of pose prediction on RE10K with diverse views (left) and cross-dataset generalization to ScanNet (right).
Method
3 views
10 views
28 views
ATE ↓
RPE-t ↓
RPE-r ↓
ATE ↓
RPE-t ↓
RPE-r ↓
ATE ↓
RPE-t ↓
RPE-r ↓
NoPoSplat
0.055
0.218
2.041
0.167
0.175
5.537
0.187
0.197
4.551
VicaSplat
0.075
0.268
1.803
0.171
0.231
3.768
0.243
0.189
3.725
SPFSplat
0.104
0.336
2.351
0.155
0.167
3.249
0.207
0.216
4.428
Anysplat ∗
0.057
0.198
1.567
0.135
0.181
1.165
0.097
0.094
1.239
Anysplat
0.062
0.213
1.611
0.148
0.196
1.192
0.098
0.091
1.362
Table 4: Quantitative results of pose prediction on ScanNet with diverse views. The best and second-best values are highlighted.
Figure 4: Scene-level visualizations and multiple novel viewpoints renderings.
Method
PSNR ↑
SSIM ↑
LPIPS ↓
ATE ↓
RPE-t ↓
RPE-r ↓
(a) Ours
26.15
0.858
0.135
0.012
0.019
0.458
(b) w Pixel Head
25.33
0.832
0.159
0.018
0.028
0.496
(c) w Any. Fusion.
25.77
0.847
0.148
0.020
0.027
0.489
(d) w/o ADF-Dec.
25.88
0.845
0.146
0.023
0.029
0.504
(e) w/o intrin. emb.
25.54
0.835
0.157
0.016
0.022
0.471
Table 5: Component ablations on RE10K (8 view).
Figure A1 : Structure of our Gaussian Prediction Head.
Figure A2 : Network details in our Gaussian Prediction Head.
Figure A3 : Number of Gaussians with increasing input views for pixel-aligned GS, Gaussian fusion methods, and our Token Fusion operation.
Figure A4 : Inference time with increasing number of input views.
Figure A5 : Visualization Results with increasing number of input views.
Figure A6 : More qualitative comparisons on RE10K with 4 input images.
Figure A7 : More qualitative comparisons on RE10K with 8 input images.
Figure A8 : More qualitative comparisons on ScanNet with 3 input images.
Figure A9 : More qualitative comparisons on ScanNet with 10 input images.
Figure A10 : More qualitative comparisons on ScanNet with 28 input images.
Department of Electrical and Computer Engineering, Sungkyunkwan University · Department of Artificial Intelligence, Sungkyunkwan University · Department of Artificial Intelligence, Yonsei University