Internet photos of a scene differ in exposure, illumination and transient objects such as passers-by, so one set of Gaussian colours cannot explain all photos (the in-the-wild setting). Recent in-the-wild Gaussian Splatting methods mostly report hours of per-scene optimization. We present FastWild-GS, which describes each photo by a 128-dimensional Photo Appearance Token and a per-channel transform, a scale and an offset applied to the rendered image (six parameters); a small MLP predicts the colour of each Gaussian from this token and a per-Gaussian feature, replacing the spherical-harmonics colour. The same transform is applied to the renders that score Gaussians for densification and pruning, so appearance differences between photos are not counted as reconstruction error. On PhotoTourism, FastWild-GS trains in 8.8 minutes per scene with the lowest LPIPS among the compared methods, 23x faster than AsymGS, the next best in LPIPS. On NeRF-OSR, it has the best average PSNR, SSIM and LPIPS among the compared methods in 9.7 minutes per scene.
Figures & tables
Figure 1: Right halves of PhotoTourism test views rendered by FastWild-GS and recent methods, with the number of training iterations and the training time per scene, measured on one RTX 5090. FastWild-GS trains in minutes, while the other methods train for up to several hours. Below each image: the PSNR of the scene as in Tab. 1 ; for gsplat, GS-W and K-Planes, as reported by NerfBaselines [ 18 ] .
Figure 2: Overview of FastWild-GS. (1) Training: the PAT Decoder predicts the colour cg,i of every Gaussian from its Splat Descriptor [fg,dg,i] and the Photo Appearance Token (PAT) ei of photo Ii , and the channel transform exp(si)⊙Ri+bi is applied to the render before the loss; a monocular depth prior regularizes the rendered depth. (2) Colour-matched densification: Gaussians are grown and pruned from the error of renders of 10 training views that use the stored SfM colour and the same per-image transform (sj,bj) ; the PAT is not used there. (3) Test-time fitting: the Gaussians and the decoder are frozen; a test-time PAT e∗ , the transform (s∗,b∗) and a radial gain v are fitted on the left half of each test photo, and the right half is scored. Blue marks the per-photo appearance state and the operations that use it; flames mark trained parts. The Gaussians are drawn as ellipsoids coloured by the trained decoder; all renders are from FastWild-GS.
Brandenburg Gate
Sacré-Cœur
Trevi Fountain
Mean
Method
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
Train
GPU
K-Planes [ 9 ]
23.65
0.859
0.253
19.96
0.762
0.299
19.70
0.662
0.388
21.10
0.761
0.313
15.9 min
RTX 5090
GS-W [ 49 ]
24.35
0.903
0.155
19.73
0.824
0.210
20.06
0.724
0.272
21.38
0.817
0.213
43.0 min
RTX 5090
gsplat [ 47 ]
26.36
0.931
0.118
22.05
0.876
0.154
22.56
0.765
0.213
23.66
0.857
0.162
1.06 h
RTX 5090
WildGaussians [ 17 ]
27.77
0.927
0.133
22.56
0.859
0.177
23.63
0.766
0.228
24.65
0.851
0.179
5.65 h
RTX 5090
NexusSplats [ 39 ]
27.76
0.922
0.141
23.13
0.859
0.174
23.97
0.766
0.237
24.95
0.849
0.184
3.39 h
RTX 5090
Table 1: PhotoTourism per scene. All rows evaluate the right half of each test image. Rows above the line give the metrics reported in each paper, and for K-Planes, GS-W and gsplat those reported by NerfBaselines [ 18 ] . Training times marked † are taken from the papers; the others are measured by us per scene on one RTX 5090, that of WilLaGS with our reimplementation. The values of FastWild-GS are means over five seeds; the last row gives their standard deviation.
Figure 3: Enlarged regions of further PhotoTourism test views. The first column shows the right half of the test photo with the region in blue; the other columns show that region in the ground truth and in the renders of each method, with FastWild-GS outlined and the other methods ordered by training time. Views and regions are chosen by a fixed rule given in the supplementary material.
Figure 4: Quality against training time per scene on PhotoTourism. All curves are those of intermediate checkpoints; methods that train longer than 20 minutes are marked at the right edge with their final value and training time. NexusSplats, AsymGS and WilLaGS are shown only by these marks, with the values and training times of Tab. 1 ; the training time of WilLaGS is that of our reimplementation on one RTX 5090. The label of FastWild-GS gives the values and the training time of Tab. 1 .
Method
PSNR ↑
SSIM ↑
LPIPS ↓
Train
NeRF-MS [ 22 ]
21.61
0.716
0.317
2.11 h
SWAG [ 5 ]
23.00
0.801
0.270
49.8 min ∗
GS-W [ 49 ]
23.79
0.792
0.217
1.77 h
WE-GS [ 42 ]
23.91
0.818
0.224
3.38 h
WilLaGS [ 1 ]
25.79
0.855
0.207
27.6 min
FastWild-GS
26.00
0.856
0.152
9.7 min
Table 2: NeRF-OSR, mean over the four scenes. Rows above the line give the metrics reported in each paper; their means are computed by us. ∗ Source code not available; numbers from the paper, computed on an unspecified “high-tier GPU”. The other training times are on one RTX 5090; for NeRF-MS, WE-GS and WilLaGS they are those of our reimplementations. The values of FastWild-GS are means over five seeds; the last row gives their standard deviation. Per-scene results are in the supplementary material ( Tab. 10 ).
Figure 5: Right halves of NeRF-OSR test views rendered by FastWild-GS and GS-W, with the number of training iterations and the training time per scene. Below each image: the PSNR of the scene as in Tab. 10 . GS-W is rendered with its official code and takes the appearance from the test photo itself. For each scene, the view is the one with the median LPIPS of FastWild-GS, a rule fixed before rendering.
Setting
PSNR ↑
SSIM ↑
LPIPS ↓
Train
FastWild-GS
25.52
0.874
0.1461
8.8 min
w/o colour-matched scoring
24.80
0.873
0.1491
8.9 min
w/o channel transform
24.94
0.874
0.1491
8.8 min
w/o Ldepth
24.86
0.873
0.1498
7.0 min
w/o PAT
21.22
0.834
0.1882
7.6 min
w/o Splat Descriptor
23.53
0.861
0.1660
11.4 min
Table 3: Ablations on PhotoTourism (protocol as in Tab. 1 ); each row removes one component of FastWild-GS. Train: training time per scene. Values are means over five seeds; their standard deviations are in the supplementary material ( Tab. 11 ). w/o colour-matched scoring : the scoring renders do not use the channel transform. w/o channel transform : no (si,bi) in training or in the scoring renders; test-time fitting still fits (s∗,b∗) . w/o PAT : the PAT Decoder takes [fg,dg,i] only, and test-time fitting adjusts only the channel transform and the radial gain. w/o Splat Descriptor : the PAT Decoder takes a trained spherical-harmonics colour of the Gaussian in direction dg,i instead of [fg,dg,i] . w/o PAT Decoder : each Gaussian is coloured by trained spherical-harmonics coefficients, without PAT.
Test-time fitting
#param.
PSNR ↑
SSIM ↑
LPIPS ↓
FastWild-GS: PAT, s,b , v ; 512 steps
137
25.52
0.874
0.1461
without radial gain v
134
25.20
0.874
0.1467
without channel transform s,b
131
24.81
0.872
0.1504
128 steps
137
25.13
0.874
0.1474
256 steps
137
25.45
0.874
0.1464
1024 steps
137
25.51
0.875
0.1461
Table 4: Test-time fitting on PhotoTourism with the trained models of FastWild-GS fixed (protocol as in Tab. 1 ).
Own colours
With our appearance module
Backbone
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
2DGS [ 13 ]
17.65
0.784
0.2618
24.68
0.866
0.1604
Mip-Splatting [ 48 ]
16.94
0.748
0.3258
24.62
0.867
0.1540
3DGS-MCMC [ 16 ]
18.40
0.810
0.2168
24.76
0.869
0.1591
Taming-3DGS [ 29 ]
18.33
0.801
0.2235
24.72
0.868
0.1539
StopThePop [ 34 ]
18.42
0.794
0.2382
24.91
0.868
0.1529
Table 5: The appearance module of FastWild-GS attached to other backbones on PhotoTourism (mean of the three scenes). Own colours : the backbone with its own colours, scored without test-time fitting; the public Scaffold-GS result in Fig. 4 (23.50 dB) instead uses appearance embeddings and 100k iterations. With our appearance module : PAT, PAT Decoder and channel transform, with the test-time fitting of Sec. 3.4 . All backbones are trained for 30k iterations. Values are means over five seeds; their standard deviations are in the supplementary material ( Tab. 12 ).
Scene
Training time (min)
Brandenburg Gate
6.1
Sacré-Cœur
8.5
Trevi Fountain
11.8
Mean
8.8
Table 6: Training time of FastWild-GS per scene on PhotoTourism on one RTX 5090.
Item
Value
Densification
iterations 600 to 14,900, every 100
Densification thresholds
FastGS defaults ( 0.0002 , 0.0012 , 0.1 ), multiplied by 1 before 12k iterations and by 0.651 from 12k
Additional pruning
at 5.9k, 8.9k, 11.9k and 14.9k iterations
Weight of Ldepth
0 until 500; linear to 0.05 at 2k; 0.05 until 15k; linear to 0.005 at 30k
Learning rates, Gaussians
position 1.6×10−4→1.6×10−6 (times the scene extent), opacity 0.025 , scaling 0.005 , rotation 0.001 , fg0.0025
Learning rates, appearance
PAT 5×10−4 , PAT Decoder 10−3 (Adam, constant)
Table 7: Hyperparameters of FastWild-GS, identical for all scenes and both datasets.
Setting
PAT
fg
Fitted
Δ PSNR
Δ SSIM
Δ LPIPS
Train
FastWild-GS
128
32
137
8.8 min
PAT size 64
64
32
73
− 0.26
− 0.0014
+0.0022
8.5 min
PAT size 256
256
32
265
− 0.30
+0.0001
+0.0007
9.7 min
feature fg size 16
128
16
137
− 0.06
− 0.0002
+0.0007
8.4 min
feature fg size 64
128
64
137
− 0.29
− 0.0004
+0.0015
9.5 min
Table 8: Size of the Photo Appearance Token (PAT) and of the per-Gaussian feature fg on PhotoTourism (mean of the three scenes). Δ : difference to FastWild-GS. Fitted: number of values fitted per test image.
Brandenburg Gate
Sacré-Cœur
Trevi Fountain
Setting
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
Train
FastWild-GS
28.87
0.940
0.1086
23.72
0.893
0.1357
23.96
0.790
0.1941
8.8 min
w/o colour-matched scoring
27.57
0.940
0.1116
23.01
0.890
0.1393
23.84
0.790
0.1963
8.9 min
w/o channel transform
27.87
0.940
0.1116
22.93
0.890
0.1410
24.02
0.791
0.1945
8.8 min
w/o Ldepth
28.12
0.939
0.1108
22.60
0.890
0.1421
23.86
0.790
0.1964
7.0 min
w/o PAT
25.02
0.918
0.1341
18.36
0.846
0.1815
20.29
0.738
0.2491
7.6 min
Table 9: Per-scene results of the ablations in Tab. 3 on PhotoTourism. The last two rows colour the scoring renders with the PAT Decoder instead of the stored SfM colour, without and with the channel transform.
europa
lwp
st
stjohann
Mean
Method
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
NeRF-MS [ 22 ]
21.03
0.721
0.294
21.90
0.719
0.336
20.68
0.630
0.402
22.84
0.793
0.235
21.61
0.716
0.317
SWAG [ 5 ]
23.74
0.845
0.242
22.07
0.783
0.303
22.29
0.713
0.364
23.91
0.864
0.172
23.00
0.801
0.270
GS-W [ 49 ]
22.04
0.758
0.231
24.44
0.827
0.208
22.45
0.690
0.315
26.23
0.893
0.113
23.79
0.792
0.217
WE-GS [ 42 ]
24.73
0.873
0.157
24.33
0.821
0.197
22.45
0.720
0.341
24.12
0.858
0.202
23.91
0.818
0.224
WilLaGS [ 1 ]
25.50
0.830
0.230
27.07
0.879
0.200
24.45
0.788
0.255
26.13
0.925
0.144
25.79
0.855
0.207
Table 10: NeRF-OSR per scene. Rows above the line give the metrics reported in each paper; their means are computed by us over the four scenes. The last row gives the standard deviation of FastWild-GS over five seeds.
Setting
PSNR
SSIM
LPIPS
Ablations ( Tab. 3 )
FastWild-GS
0.15
0.0003
0.0017
w/o colour-matched scoring
0.12
0.0005
0.0012
w/o channel transform
0.07
0.0004
0.0005
w/o Ldepth
0.00
0.0006
0.0010
w/o PAT
0.02
0.0000
0.0010
Table 11: Standard deviation over five seeds of the PhotoTourism means in Tabs. 3 , 4 and 8 .
Own colours
With our module
Backbone
PSNR
SSIM
LPIPS
PSNR
SSIM
LPIPS
2DGS
0.09
0.0086
0.0150
0.02
0.0009
0.0016
Mip-Splatting
0.16
0.0091
0.0134
0.01
0.0021
0.0014
3DGS-MCMC
0.09
0.0015
0.0013
0.06
0.0006
0.0015
Taming-3DGS
0.03
0.0009
0.0035
0.06
0.0010
0.0012
StopThePop
0.20
0.0043
0.0080
0.04
0.0003
0.0008
Table 12: Standard deviation over five seeds of the values in Tab. 5 .
Brandenburg Gate
Sacré-Cœur
Trevi Fountain
Setting
PSNR
SSIM
LPIPS
PSNR
SSIM
LPIPS
PSNR
SSIM
LPIPS
FastWild-GS
0.39
0.0001
0.0029
0.20
0.0009
0.0017
0.04
0.0010
0.0021
w/o colour-matched scoring
0.31
0.0012
0.0030
0.12
0.0015
0.0010
0.04
0.0008
0.0011
w/o channel transform
0.05
0.0003
0.0010
0.18
0.0009
0.0020
0.05
0.0007
0.0004
w/o Ldepth
0.17
0.0003
0.0006
0.12
0.0010
0.0021
0.06
0.0005
0.0005
w/o PAT
0.01
0.0000
0.0004
0.05
0.0004
0.0021
0.01
0.0004
0.0006
Table 13: Standard deviation over five seeds of the per-scene values in Tab. 9 .
Figure 6: Per-pixel error on the right halves of the test views of Fig. 1 : mean absolute RGB error against the ground truth, on one colour scale capped at 0.35. Below each map: the mean absolute error of the shown half.
We propose WildSplatter, a feed-forward 3D Gaussian Splatting (3DGS) model for unconstrained images with unknown camera parameters and varying lighting conditions. 3DGS is an effective scene representation that enables high-quality, real-time rendering; however, it typically requires iterative optimization and multi-view images captured under consistent lighting with known camera parameters. WildSplatter is trained on unconstrained photo collections and jointly learns 3D Gaussians and appearance embeddings conditioned on input images. This design enables flexible modulation of Gaussian colors to represent significant variations in lighting and appearance. Our method reconstructs 3D Gaussians from sparse input views in under one second, while also enabling appearance control under diverse lighting conditions. Experimental results demonstrate that our approach outperforms existing pose-free 3DGS methods on challenging real-world datasets with varying illumination.
Feed-forward 3D Gaussian Splatting (3DGS) removes the need for time-consuming per-scene optimization required by traditional 3DGS. However, existing feed-forward approaches struggle with real-world photo collections that include diverse lighting conditions and transient objects. In this paper, we present Wild3R, a feed-forward approach for unconstrained sparse photo collections. The main bottleneck is the lack of training data that provides multiple viewpoints, a variety of illuminations, and transient variations necessary for learning robust scene representations. To address this, we introduce the WildCity dataset, which comprises 200 scenes, 170 lighting conditions, and transient objects, resulting in 337,500 images in total. By leveraging the dataset, our model learns appearance consistency across viewpoints conditioned on reference views, while removing transient content. Extensive experiments demonstrate that our method outperforms existing feed-forward approaches and achieves results competitive with prior per-scene optimization-based methods.
While feedforward 3D reconstruction excels at efficient novel view synthesis, it typically falters when faced with scenes under varying illumination. To this end, we introduce WildSplat, the first feedforward 3D Gaussian Splatting framework capable of appearance-conditioned novel-view synthesis for unposed in-the-wild images. To handle inconsistent photometric conditions, we propose a dual-branch architecture that explicitly decouples geometry from appearance. The geometry branch extracts an appearance-invariant 3D structure and jointly predicts camera poses. To govern the rendering appearance, the appearance branch injects target appearance cues into the content features via a globally pre-modulated cross-attention mechanism. To further prevent feature entanglement, we introduce a joint multi-reference training strategy that stabilizes the training process. Extensive experiments show that WildSplat surpasses existing optimization-based and feedforward methods, achieving state-of-the-art performance in in-the-wild novel view synthesis and appearance editing from sparse inputs in a single forward pass.
Xiyu Zhang, Jingyu Zhuang, Hongjia Zhai +4
State Key Lab of CAD&CG, Zhejiang University, China · vivo BlueImage Lab, China