PAM-ToD: Plug-and-Play Appearance Modeling for Cross-Time-of-Day 3D Gaussian Splatting

Kota Shimomura; Sungho Moon; Tsubasa Hirakawa; Takayoshi Yamashita; Sunghoon Im; Hironobu Fujiyoshi

cs.CVOct 8, 2026

PAM-ToD: Plug-and-Play Appearance Modeling for Cross-Time-of-Day 3D Gaussian Splatting

Authors: Kota Shimomura, Sungho Moon, Tsubasa Hirakawa, Takayoshi Yamashita, Sunghoon Im, Hironobu Fujiyoshi

Organizations: Chubu University, Japan · Elith Inc., Japan · DGIST, Daegu, Republic of Korea · KAIST, Daejeon, Republic of Korea

Abstract

Adapting a pre-trained 3D Gaussian Splatting (3DGS) road scene to a new time of day requires learning appearance changes from a few anchor images while preserving consistent, real-time rendering. We propose PAM-ToD, a lightweight plug-in that learns color corrections while keeping the pre-trained 3DGS parameters fixed. PAM-ToD scales each Gaussian's existing color to model illumination changes and uses an additive term for additional brightness, such as when street lamps turn on at night. Under a simplified image formation model, unchanged surface albedo can be eliminated from the relation between source and target appearances, allowing us to learn these corrections without separately estimating albedo and illumination. The model corrects colors across the scene while allowing the corrections to vary by location and by Gaussian. To guide learning from a few anchor images, it discourages abrupt spatial changes in these corrections. We also introduce CARLA-ToD, a benchmark with matching geometry, camera poses, and moving-object trajectories across three times of day. A few target-time anchor images are used to train each plug-in, while separate views are used for evaluation. Across the static and dynamic settings, PAM-ToD achieves higher PSNR and lower LPIPS than the baselines, even when the anchor images come from a single synchronized capture across multiple cameras.

Figures & tables

Figure 1: Overview of PAM-ToD .Given a frozen 3DGS scene from source time $s$ and sparse anchor images at target time $t$ , PAM-ToD adapts scene appearance in real time using a multiplicative illumination term and an additive emission term.

	Reconstruction			Novel View Synthesis			FPS $\uparrow$
Methods	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$
GS-W	0.748	16.76	0.312	0.729	16.58	0.324	22.9
GS-IR	0.663	16.42	0.372	0.671	16.57	0.361	61.8
GI-GS	0.663	16.37	0.353	0.667	16.60	0.342	16.3
LumiGauss	0.491	11.29	0.597	0.467	10.89	0.622	152.7
StreetGS	0.471	10.17	0.487	0.463	10.13	0.486	66.7

Table 1: Quantitative relighting comparison on the CARLA-ToD Static Dataset. The first, second, and third best performances are highlighted in First , Second , and Third , respectively.

Figure 2: Qualitative comparison on the CARLA-ToD Static Dataset.

	Reconstruction			Novel View Synthesis			FPS $\uparrow$
Methods	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$
LumiNet	0.569	14.44	0.422	0.560	14.45	0.421	0.08
DiffusionRenderer	0.541	14.49	0.496	0.529	14.49	0.566	0.08
UniRelight	0.597	15.89	0.484	0.560	15.49	0.577	0.07
StreetGS	0.467	10.54	0.486	0.462	10.51	0.481	59.3
+Ours ( $K=1$ )	0.715	21.29	0.318	0.706	21.15	0.311	49.4

Table 2: Quantitative relighting comparison on the CARLA-ToD Dynamic Dataset.

Figure 3: Qualitative comparison on the CARLA-ToD Dynamic Dataset.

Table 6

	$K=1$			$K=8$
Method	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$
Base	0.471	10.17	0.487	0.471	10.17	0.487
w/o Illumination term	0.565	12.83	0.431	0.659	16.73	0.375
w/o Emission term	0.717	20.86	0.329	0.783	22.69	0.311
Global Only	0.597	16.91	0.439	0.786	23.86	0.303
Residual Only	0.714	20.60	0.399	0.790	23.58	0.289

Table 5: Ablation on the correction channels and hierarchy levels with

K{=}1,8

Figure 4: Qualitative comparison on the Waymo Open Dataset.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Loss weights	$\lambda_{\mathrm{anc}}=20$ , $\lambda_{\mathrm{ssim}}=0.2$ , $\lambda_{\mathrm{atp}}=1$ , $\lambda_{\mathrm{unseen}}=0.1$ , $\lambda_{\mathrm{rch}}=1$ , $\lambda_{\mathrm{ch}}=1$ , $\lambda_{\mathrm{chu}}=2$
Regularizers	$\lambda_{\mathrm{tv}}=10^{-2}$ , $\lambda_{\mathrm{knn}}=10^{-2}$ , $\lambda_{E}=10^{-4}$
Spatial field	3 levels ( $n_{l}\in\{16,32,64\}$ ), TV weight $n_{l}/16$
Residual graph	$k=8$ neighbors
Robust anchor	$\sigma=2$ pixels, $\rho=10\%$
ATP prior	$Q=32$ bins, classes with $<2{,}000$ anchor px or $<Q$ Gaussians skipped, $65{,}536$ samples/iter
Seen-unseen	edges $\{2,4,8,16,32\}$ m, min. 500 Gaussians/cell, update every 200 iters

Appendix

Table 6: Hyperparameters of PAM-ToD, shared by all experiments.

Sensor	Loc $(x,y,z)$	Rot $(\mathrm{roll},\mathrm{pitch},\mathrm{yaw})$
Front camera (ID 0)	$(0,0,0)$	$(0,0,0)$
Front-left camera (ID 1)	$(-0.045,0.116,-0.001)$	$(-0.05,-0.58,55.00)$
Front-right camera (ID 2)	$(-0.049,-0.070,0.001)$	$(-0.55,-0.44,-55.00)$
LiDAR	$(-1.541,0.021,-0.300)$	$(-0.70,0.42,-0.01)$

Appendix

Table 7: Sensor mounting relative to the front camera, in a camera basis with

x

forward,

y

left, and

z

up. Location in m, rotation in deg.

Method	Gaussians	Params	Size
GS-IR	1.44 M	122.9 M	492 MB
GI-GS	1.58 M	125.9 M	504 MB
GS-W	1.91 M	173.7 M	695 MB
StreetGS	1.48 M	106.2 M	425 MB
+ Ours	1.48 M	113.9 M (+7.7 M)	456 MB (+31 MB)

Appendix

Table 8: Model size on CARLA-ToD Static.

ToD	Metric	GI-GS	GS-IR	LumiGauss	GS-W	Ours ( $K{=}1$ )	Ours ( $K{=}8$ )
Mo $\rightarrow$ No	PSNR $\uparrow$	11.91	12.37	7.50	19.47	20.93	24.03
	SSIM $\uparrow$	0.697	0.689	0.428	0.815	0.777	0.827
	LPIPS $\downarrow$	0.362	0.382	0.609	0.265	0.251	0.237
No $\rightarrow$ Mo	PSNR $\uparrow$	19.81	18.65	7.66	15.63	25.30	27.00
	SSIM $\uparrow$	0.736	0.707	0.441	0.823	0.798	0.858
	LPIPS $\downarrow$	0.306	0.318	0.587	0.287	0.278	0.234

Appendix

Table 9: Quantitative comparison of reconstruction on the CARLA-ToD Static Dataset. Mo, No, and Ni denote Morning, Noon, and Night, respectively.

ToD	Metric	GI-GS	GS-IR	LumiGauss	GS-W	Ours ( $K{=}1$ )	Ours ( $K{=}8$ )
Mo $\rightarrow$ No	PSNR $\uparrow$	12.74	12.98	7.51	16.41	20.70	23.56
	SSIM $\uparrow$	0.719	0.715	0.400	0.780	0.757	0.806
	LPIPS $\downarrow$	0.339	0.352	0.634	0.310	0.252	0.235
No $\rightarrow$ Mo	PSNR $\uparrow$	20.42	19.11	7.99	15.65	25.04	27.00
	SSIM $\uparrow$	0.744	0.731	0.437	0.780	0.788	0.848
	LPIPS $\downarrow$	0.289	0.301	0.601	0.310	0.273	0.223

Appendix

Table 10: Quantitative comparison of novel view synthesis (NVS) on the CARLA-ToD Static Dataset. Mo, No, and Ni denote Morning, Noon, and Night, respectively.

Figure 5: Qualitative comparison on of the number of anchors.

Figure 6: Qualitative comparison on our key illumination and emission term.

Figure 7: Additional qualitative results for NVS on CARLA-ToD static scenes.

Figure 8: Additional qualitative results for NVS on CARLA-ToD dynamic scenes.

Figure 9: Additional qualitative results for NVS on WOD.

Explore similar work

WildSplatter: Feed-forward 3D Gaussian Splatting with Appearance Control from Unconstrained Images
Apr 23, 2026Yuki Fujimura, Takahiro Kushida, Kazuya Kitano +23D Gaussian Splatting 3D Reconstruction
NAIST · Ritsumeikan University · Kyoto University
P2GS: Physical Prior-guided Gaussian Splatting for Photometrically Consistent Urban Reconstruction
May 16, 2026Kota Shimomura, Hidehisa Arai, Tsubasa Takahashi +2Multi-View Consistency 3D Gaussian Splatting
Chubu University1 · Turing Inc.2
AdpSplit: Error-Driven Adaptive Splitting for Faster Geometry Discovery in 3D Gaussian Splatting
May 7, 2026Yongjae Lee, Jingxing Li, Abhay Kumar Yadav +23D Gaussian Splatting 3DGS Acceleration
Arizona State University · Johns Hopkins University

Apr 23, 2026, cs.CVJ/K move · Enter open · S save

WildSplatter: Feed-forward 3D Gaussian Splatting with Appearance Control from Unconstrained Images

Yuki Fujimura, Takahiro Kushida, Kazuya Kitano, Takuya Funatomi+1

NAIST · Ritsumeikan University · Kyoto University

3D Gaussian Splatting 3D Reconstruction Sparse-View 3D Reconstruction Feed-Forward 3D Reconstruction

Open paper AskPDF arXiv

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.

Abstract

Explore similar work

WildSplatter: Feed-forward 3D Gaussian Splatting with Appearance Control from Unconstrained Images

P2GS: Physical Prior-guided Gaussian Splatting for Photometrically Consistent Urban Reconstruction

AdpSplit: Error-Driven Adaptive Splitting for Faster Geometry Discovery in 3D Gaussian Splatting