Recent feed-forward 3D Gaussian Splatting (3DGS) methods have enabled effective multi-view 3D reconstruction by predicting Gaussian primitives in a single forward pass. However, existing approaches predominantly adopt a pixel-aligned paradigm that predicts one Gaussian per input pixel, coupling the Gaussian count to the input resolution and number of views. This also leads to redundant primitives in textureless regions and insufficient coverage in geometrically complex areas. We present Anchor3DGS, a pose-free feed-forward 3DGS framework that replaces pixel-aligned prediction with a compact anchor-based representation. Guided by information entropy, our method places a set of 3D anchors in the scene, concentrating them in regions of higher visual complexity while maintaining broad spatial coverage. Each anchor performs occlusion-aware aggregation of multi-view image features, exchanges information with other anchors, and is decoded into a set of local Gaussians. A lightweight feature enhancement module further refines the coarse RGB output from Gaussian splatting and improves rendering fidelity. Extensive experiments on diverse benchmark datasets demonstrate that our Anchor3DGS achieves state-of-the-art novel view synthesis quality while only using an order of magnitude fewer Gaussian primitives.
Figures & tables
Figure 1: Anchor3DGS reconstructs 3D scenes from uncalibrated multi-view images. Our anchor-based representation achieves better rendering quality with fewer Gaussians, and avoids the empty holes that pixel-aligned methods ( e.g., AnySplat [ 11 ] ) commonly leave in poorly observed regions.
Figure 2: Overview of Anchor3DGS. Given uncalibrated multi-view images, (1) a VGGT-based [ 28 ] geometry backbone first predicts the scene’s geometric scaffold. (2) A fixed budget of 3D anchors is then sampled from the scaffold, combining uniform sampling with entropy-guided sampling. (3) Each anchor is projected back into all input views to gather features via occlusion-aware aggregation, and then interacts with other anchors through self-attention. (4) Finally, each anchor is decoded into a set of local Gaussians, which are rasterized and refined to yield the final RGB output.
Figure 3: Uniform sampling ensures broad spatial coverage but wastes anchors (red dots) on uninformative regions ( e.g., textureless sky). In contrast, our entropy-guided sampling allocates more anchors to visually complex regions while still maintaining coverage of the scene.
Setting
Method
VR-NeRF
MipNeRF360
DL3DV
PSNR
SSIM
LPIPS
#GS
PSNR
SSIM
LPIPS
#GS
PSNR
SSIM
LPIPS
#GS
3 views
NoPoSplat [ 38 ]
20.52
0.686
0.433
0.60M
16.89
0.449
0.493
0.60M
17.29
0.509
0.451
0.60M
FLARE [ 44 ]
20.64
0.692
0.452
0.60M
16.92
0.447
0.498
0.60M
16.08
0.468
0.498
0.60M
DA3 [ 15 ]
19.87
0.656
0.464
0.60M
12.84
0.330
0.634
0.60M
18.24
0.510
0.376
0.60M
AnySplat [ 11 ]
21.31
0.729
0.340
0.49M
15.65
0.624
0.331
0.38M
17.74
0.625
0.308
0.51M
Ours
23.69
0.764
0.320
0.26M
21.36
0.678
0.307
0.26M
21.22
0.683
0.280
0.26M
Table 1: Quantitative comparison of the sparse-view novel view synthesis (3 and 6 input views) on VR-NeRF [ 35 ] , MipNeRF360 [ 2 ] , and DL3DV [ 16 ] . We report PSNR, SSIM, LPIPS, and the number of predicted Gaussian primitives (#GS). Bold indicates the best.
Method
VR-NeRF
DL3DV
PSNR
SSIM
LPIPS
#GS
Storage
Time
PSNR
SSIM
LPIPS
#GS
Storage
Time
VGGT+3DGS [ 13 ]
23.08
0.762
0.301
0.44M
109MB
6min
19.34
0.627
0.312
1.16M
288MB
8min
DA3 [ 15 ]
22.38
0.751
0.350
6.42M
1.03GB
1.9s
19.38
0.544
0.341
6.42M
1.03GB
2.0s
AnySplat [ 11 ]
24.05
0.819
0.309
4.75M
779MB
2.8s
18.76
0.565
0.371
3.98M
653MB
2.8s
Ours
24.56
0.815
0.297
0.26M
35MB
2.9s
19.52
0.593
0.364
0.26M
35MB
3.1s
Table 2: Dense-view rendering results ( 32 input views) on VR-NeRF [ 35 ] and DL3DV [ 16 ] . We also report the storage size of the predicted Gaussians and the inference time on a single H800 GPU.
Figure 4: Qualitative comparison of novel view synthesis. Our method recovers sharper textures and finer details while mitigating artifacts in poorly observed regions.
Configuration
3 views
32 views
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
w/o enhancement
22.50
0.754
0.345
24.13
0.807
0.324
w/o self-attention
23.03
0.750
0.323
24.38
0.808
0.303
w/o occ-aggregation
23.28
0.756
0.330
24.29
0.813
0.305
w/o entropy-sample
22.16
0.733
0.355
24.17
0.804
0.318
Full model (Ours)
23.69
0.764
0.320
24.56
0.815
0.297
Table 3: Ablation study results on VR-NeRF [ 35 ] dataset. Bold indicates the best.
Figure 5: Rendering metrics under varying anchor count N and decoded Gaussians K ( left ), and qualitative comparisons of removing the enhancement module and entropy-guided sampling ( right ).
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Method
AUC@ 10↑
AUC@ 20↑
AUC@ 30↑
VGGT [ 28 ]
50.4
67.2
74.9
AnySplat [ 11 ]
53.1
73.1
80.8
Ours
54.8
74.0
81.5
Appendix
Table A1 : Camera pose estimation results on Co3Dv2 [ 21 ] dataset. We report AUC at angular thresholds of 10 , 20 , and 30 degrees. Bold indicates the best.
Method
VR-NeRF
DL3DV
PSNR
SSIM
LPIPS
#GS
Storage
PSNR
SSIM
LPIPS
#GS
Storage
VGGT+3DGS [ 13 ]
23.74
0.780
0.289
0.46M
114MB
20.69
0.667
0.270
1.39M
345MB
DA3 [ 15 ]
21.91
0.725
0.365
12.8M
2.10GB
19.64
0.564
0.328
12.8M
2.10GB
AnySplat [ 11 ]
23.63
0.797
0.318
8.97M
1.47GB
18.71
0.567
0.369
7.43M
1.22GB
Ours
24.01
0.799
0.311
0.26M
35MB
19.33
0.585
0.367
0.26M
35MB
Appendix
Table A2 : Novel-view synthesis with 64 context views on VR-NeRF [ 35 ] and DL3DV [ 16 ] . We also report the number (#GS) and storage size of the predicted Gaussians.
Method
6 views
32 views
64 views
#GS
FPS ↑
#GS
FPS ↑
#GS
FPS ↑
DA3 [ 15 ]
1.20M
361.1
6.42M
146.3
12.8M
85.9
AnySplat [ 11 ]
1.03M
375.4
4.75M
176.7
8.97M
120.5
Ours
0.26M
179.4
0.26M
190.6
0.26M
187.4
Appendix
Table A3 : Rendering speed analysis on VR-NeRF [ 35 ] dataset. #GS denotes the number of predicted Gaussians; FPS is the per-frame 3DGS rendering speed averaged over 100 iterations.
Warping RMSE ↓ (traj)
with UNet
without UNet (reference)
6 input views
0.01187
0.01018
32 input views
0.00974
0.00970
Appendix
Table A4 : Multi-view consistency analysis on VR-NeRF [ 35 ] dataset. We report the RMSE metric.
Configuration
3 views
6 views
32 views
PSNR
SSIM
LPIPS
PSNR
SSIM
LPIPS
PSNR
SSIM
LPIPS
w/o enhancement
18.81
0.647
0.334
19.62
0.622
0.358
18.66
0.564
0.424
w/o self-attention
20.88
0.673
0.292
20.81
0.648
0.312
19.43
0.588
0.375
w/o occ-aggregation
21.00
0.675
0.295
20.72
0.642
0.317
19.03
0.576
0.390
w/o entropy-sample (uniform)
20.93
0.669
0.290
20.67
0.634
0.309
19.21
0.580
0.373
w/o entropy-sample (voxel)
20.32
0.630
0.345
19.73
0.594
0.376
18.34
0.537
0.446
Appendix
Table A5 : Ablation study on DL3DV [ 16 ] dataset under 3 , 6 , and 32 input views.
Figure A1 : Rendering metrics under varying anchor count N and decoded Gaussians per anchor K on DL3DV [ 16 ] and MipNeRF360 [ 2 ] datasets.
Dataset
Setting
3 views
6 views
32 views
PSNR
SSIM
LPIPS
PSNR
SSIM
LPIPS
PSNR
SSIM
LPIPS
VR-NeRF
Backbone + Pixel Decoder
21.59
0.736
0.339
22.48
0.760
0.319
24.24
0.818
0.304
AnySplat [ 11 ]
21.31
0.729
0.340
22.45
0.758
0.321
24.05
0.819
0.309
Anchor3DGS (Ours)
23.69
0.764
0.320
24.03
0.781
0.312
24.56
0.815
0.297
DL3DV
Backbone + Pixel Decoder
17.89
0.632
0.305
18.81
0.620
0.312
19.02
0.576
0.369
AnySplat [ 11 ]
17.74
0.625
0.308
18.42
0.613
0.316
18.76
0.565
0.371
Appendix
Table A6 : Ablation study on the anchor representation. We compare our anchor-based head with a standard pixel-aligned decoder under the identical fine-tuned VGGT backbone.
Figure A2 : More qualitative comparisons against existing feed-forward methods.