Standard 3D Gaussian Splatting (3DGS) pipelines struggle with severely underexposed images, producing unstable colors, floating artifacts, and incomplete reconstructions in dark regions. To better account for noise in these low-light observations, RAW images offer a natural basis for noise-aware reconstruction by preserving linear sensor measurements and providing access to noise metadata. We present RetinexGS, a RAW-domain 3DGS framework that combines Retinex-inspired appearance modeling with sensor-informed optimization for low-light novel-view synthesis. We factorize each Gaussian's color into a view-independent base color and a shared, position- and view-dependent modulation field, regularizing the two components separately to constrain appearance learning from noisy observations. We use DNG noise metadata to weight the reconstruction loss for signal-dependent noise, while distortion-aware rendering avoids spatially resampling RAW targets. To address sparse geometry in dark regions, we introduce depth-guided initialization. We evaluate RetinexGS on three public RawNeRF scenes and two custom-captured low-light RAW scenes. With all methods trained on the same RAW inputs for each scene, RetinexGS outperforms LE3D and RawNeRF in mean sRGB PSNR by 2.18 dB and 2.71 dB respectively, across the five scenes, while supporting real-time novel-view rendering.
Figures & tables
Figure 1: RetinexGS enables high-quality and efficient low-light RAW novel-view synthesis. Left: On a noisy held-out view, RetinexGS suppresses noise-driven artifacts while preserving thin structures and fine details; the crops compare RawGS, LE3D, RetinexGS, and the observed noisy target. Right: RetinexGS achieves the highest affine-sRGB PSNR on each of the five evaluated scenes and the highest five-scene mean after only 15,000 training iterations, compared with 30,000 for the Gaussian baselines. RetinexGS, LE3D, and Raw3DGS receive identical half-pack inputs; HDRSplat retains its native PMRID preprocessing. The five-scene Raw3DGS control is reported in Appendix F . RetinexGS renders at 91 – 145 FPS at 2016×1512 resolution on a single RTX 4090.
Figure 2: RetinexGS overview. Given noisy low-light RAW images and their DNG metadata, RetinexGS optimizes a Gaussian scene representation through three components. (1) A per-Gaussian base color Ri and a shared, position- and view-conditioned modulation Li(μi,di) define the primitive color Ci=Ri⊙Li(μi,di) . Separate regularization constrains the two factors. (2) Distortion-aware rasterization renders these colors directly onto the packed RAW grid, where training combines DNG-informed noise weighting, a log-domain reconstruction loss, and appearance regularization. (3) A frozen monocular depth prior supplies additional initialization points through dark-region sampling and distortion-aware unprojection. The DNG ISP branch is used only for visualization and evaluation.
Method
Iterations
PSNR ↑
SSIM ↑
LPIPS ↓
LDR-NeRF [ 42 ]
–
20.04
0.5541
0.5669
LDR-3DGS [ 1 ]
30k
20.29
0.5477
0.5344
HDR-3DGS [ 5 ]
30k
20.33
0.5286
0.6563
RawNeRF [ 4 ]
–
24.08
0.6100
0.4952
RawGS [ 5 ]
30k
23.35
0.5843
0.5472
LE3D [ 5 ]
30k
24.70
0.6076
0.5071
Table 1: Native-pipeline comparison on RawNeRF. Mean sRGB metrics over the three test scenes. Baseline values except HDRSplat+PMRID are taken from LE3D’s unified comparison [ 5 ] ; HDRSplat+PMRID is reproduced with its native preprocessing (Appendix E ). Bold and underline indicate the best and second-best values.
Scene
Method
Iterations
PSNR ↑
SSIM ↑
LPIPS ↓
miningtunnel
LDR-3DGS [ 1 ]
30k
17.29
0.3878
0.6209
Raw3DGS (our impl.)
30k
11.85
0.3415
0.7327
LE3D [ 5 ]
30k
21.27
0.4250
0.6017
HDRSplat+PMRID [ 16 ]
30k
15.78
0.3340
0.7330
RetinexGS
15k
23.47
0.4897
0.5053
housecorner
LDR-3DGS [ 1 ]
30k
18.95
0.6051
0.6264
Table 2: Native-pipeline comparison on self-collected scenes. All methods are trained by us. RawNeRF on these scenes is evaluated under the matched-input protocol (Table 3 ). Bold and underline indicate the best and second-best values.
Figure 3: Qualitative comparison on RawNeRF scenes. Held-out novel views, shown without per-channel affine correction since clean references are unavailable at these poses. Compared with RawGS, LE3D, and LDR-3DGS, RetinexGS produces fewer floating artifacts in dark regions and preserves finer details.
Method
sRGB
Linear RAW
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
RawNeRF [ 4 ]
22.31
0.484
0.550
60.92
0.9977
LE3D [ 5 ]
22.83
0.550
0.553
61.14
0.9977
RetinexGS
25.01
0.600
0.471
63.39
0.9987
Table 3: Matched-input comparison. Five-scene means for methods retrained on identical half-packed RAW tensors at the same resolution. sRGB metrics use per-channel affine alignment in linear RAW before ISP conversion; linear-RAW metrics use neither. RetinexGS attains the highest PSNR on every scene in both domains; per-scene values are given in Table 10 . Best values are shown in bold.
Variant
pianotest
miningtunnel
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
SH appearance
19.20
0.4651
0.5829
19.53
0.4019
0.6322
RawNeRF weighting [ 4 ]
25.59
0.6238
0.4800
23.47
0.4878
0.5117
Log-L1 only
25.42
0.6168
0.4801
22.19
0.4566
0.5485
w/o factor reg.
25.77
0.6213
0.4630
23.27
0.4894
0.5107
w/o depth init.
25.80
0.6209
0.4725
22.81
0.4477
0.5837
Table 4: Component ablations. sRGB metrics. RawNeRF weighting replaces the DNG-informed weighting; log-L1 only removes the heteroscedastic term. Best values are shown in bold.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Quantity
Input / storage
Output transform
Ri
learnable ci∈R3 per Gaussian
exp(clamp(−∞,3](ci))
Li
encoded μi and di ; shared MLP
clamp[0.05,10](softplus(⋅))
Ci
Ri and Li(μi,di)
Ri⊙Li
Appendix
Table 5: Factor architecture shared by the two composition variants. The output bias initializes Li near one. The base anchor is directly optimized, and the shared modulation field is parameterized by ϕL . The final row is the primitive-level composition used by RetinexGS; the image-space variant instead rasterizes Ri and Li separately and multiplies the resulting maps.
Loss term
Weight
Active from
λanc
0.001
iter 1000
λL (modulation magnitude)
0.02
iter 2000 (3k warm-up)
λR (base anchor)
0.005
iter 4000
λchr
0.005
iter 4000
λlog (iter ≤ 2000)
0.4
iter 0
λlog (iter > 2000)
0.8
iter 2000
Appendix
Table 6: Loss weights and activation schedule.
Scene
HDRSplat Gaussians
RetinexGS Gaussians
Reduction
pianotest
505k
340k
33%
officetest
579k
273k
53%
yuccatest
199k
133k
33%
miningtunnel
302k
225k
25%
housecorner
460k
138k
70%
Mean
409k
222k
46%
Appendix
Table 7: Optimized primitive counts for HDRSplat+PMRID (30k iterations, degree-3 SH) and RetinexGS (15k iterations). The mean reduction is computed from the mean counts.
Scene
HDRSplat+PMRID
RetinexGS
Difference
pianotest
22.84 / 0.565 / 0.490
25.84 / 0.623 / 0.461
+3.00 / +0.058 / −0.029
officetest
23.97 / 0.587 / 0.475
25.58 / 0.635 / 0.440
+1.61 / +0.048 / −0.035
yuccatest
19.73 / 0.396 / 0.638
23.91 / 0.560 / 0.509
+4.18 / +0.164 / −0.129
miningtunnel
15.78 / 0.334 / 0.733
23.47 / 0.490 / 0.505
+7.69 / +0.156 / −0.228
housecorner
16.57 / 0.517 / 0.755
26.27 / 0.692 / 0.440
+9.70 / +0.175 / −0.315
Appendix
Table 8: Per-scene direct comparison in sRGB PSNR / SSIM / LPIPS under the common output-domain evaluation. The public RawNeRF scenes support the primary comparison; the larger collected-scene gaps are interpreted only as end-to-end robustness results.
Scene
Iter.
PSNR ↑
SSIM ↑
LPIPS ↓
pianotest
30k
19.23
0.5675
0.6252
officetest
30k
21.85
0.5932
0.5506
yuccatest
30k
18.53
0.4973
0.6467
miningtunnel
30k
11.85
0.3415
0.7327
housecorner
30k
24.61
0.6255
0.5222
Mean
30k
19.21
0.5250
0.6155
Appendix
Table 9: Per-scene results for our vanilla Raw3DGS control: SH appearance with log-L1 and weighted-L2 supervision [ 4 ] , trained for 30,000 iterations. Metrics are computed in sRGB under the RawNeRF-style per-channel affine protocol; housecorner is averaged over three reference views.
Scene
Method
sRGB PSNR
sRGB SSIM
sRGB LPIPS
RAW PSNR
RAW SSIM
pianotest
RetinexGS
25.84
0.623
0.461
62.74
0.9987
LE3D (half-packed)
24.67
0.593
0.501
60.05
0.9979
RawNeRF (half-packed)
22.84
0.476
0.562
59.69
0.9978
HDRSplat (native)
22.84
0.565
0.490
59.19
0.9970
officetest
RetinexGS
25.58
0.635
0.440
61.30
0.9981
LE3D (half-packed)
24.44
0.607
0.496
59.78
0.9973
Appendix
Table 10: Complete five-scene values used in the matched-input and common-output analyses. RetinexGS, LE3D, and RawNeRF receive identical half-packed tensors; HDRSplat retains its native PMRID pipeline and is included only as a common output-domain comparison. sRGB reports PSNR/SSIM/LPIPS under the RawNeRF-style affine protocol; RAW reports PSNR/SSIM without affine fitting.
Method
RAW PSNR ↑
RAW SSIM ↑
HDR-3DGS
56.50
0.9926
RawNeRF
58.69
0.9969
RawGS
59.28
0.9971
LE3D
61.08
0.9983
RetinexGS
61.72
0.9984
Appendix
Table 11: Method-specific RAW-domain diagnostics on the RawNeRF test scenes. These values are documented but not used as the primary cross-method ranking.
Scene
Composition
sRGB PSNR ↑
SSIM ↑
LPIPS ↓
RAW PSNR ↑
RAW SSIM ↑
pianotest
Primitive-level
25.84
0.6230
0.4610
62.74
0.9987
Image-space
25.72
0.6215
0.4635
62.59
0.9986
officetest
Primitive-level
25.58
0.6350
0.4400
61.30
0.9981
Image-space
25.48
0.6345
0.4437
61.02
0.9980
yuccatest
Primitive-level
23.91
0.5600
0.5090
61.12
0.9983
Image-space
23.77
0.5596
0.5149
60.94
0.9983
Appendix
Table 12: Full composition comparison. Each model is rendered with its requested operator, and the two arms otherwise use the same factor parameterization, data, optimization, and evaluation. Primitive-level entries are the canonical RetinexGS checkpoints used throughout the main results and ablations; image-space entries are the independently verified comparison runs. Primitive-level composition requires approximately half the wall-clock training time (Sec. 4.5 ).
Scene
# Gaussians
RetinexGS
Matched SH control
FPS ↑
Memory (MB) ↓
FPS ↑
Memory (MB) ↓
pianotest
340k
91
602
148
614–661
housecorner
138k
145
471
176–187
475–494
miningtunnel
225k
138
427
207–215
434–464
Appendix
Table 13: Rendering efficiency. Throughput and peak memory at 2016×1512 resolution on an NVIDIA RTX 4090. SH ranges span degrees 1 and 3. Geometry, opacity, cameras, Gaussian counts, and the distortion-aware rasterizer are fixed to isolate appearance-evaluation cost.
Figure 4: Qualitative comparison on our self-collected benchmark. Test-view renderings on miningtunnel and housecorner , where long-exposure tripod-stabilized reference views are available. From left to right: noisy RAW input (same pose as GT, shown for reference), LDR-3DGS [ 1 ] , Raw3DGS (our impl.), LE3D [ 5 ] , RetinexGS (ours), and GT reference. All RAW-based renderings are aligned to GT via per-channel affine correction in linear RAW; LDR-3DGS is aligned in sRGB. Red boxes indicate zoomed regions. Our method produces fewer floating artifacts, more accurate color, and better details in dark regions.
Figure 5: Replacing our DNG-informed weighting with RawNeRF’s weighted L2 loss introduces a visible purple cast in low-signal regions, despite nearly identical PSNR/SSIM in Table 4 .
Full 360∘ novel view synthesis under low-light conditions remains challenging. Insufficient illumination, noise amplification, and view-dependent photometric inconsistencies prevent existing methods from jointly preserving geometric consistency and photorealism. Unsupervised approaches often exhibit color drift under large viewpoint variations, while supervised low-light enhancement models, though effective for 2D tasks, struggle to generalize to new scenes and typically require retraining. To address this issue, we propose MERID-GS, a Multi-Scale Explicit Retinex Illumination-Decoupled Gaussian framework for low-light 360∘ synthesis. Based on Retinex theory, the method explicitly separates illumination and reflectance, and suppresses noise propagation while enhancing dark-region structures via a learnable gain and Illumination-State-Guided Frequency Gating. Combined with lightweight Reflection Head and 3D Gaussian Splatting, MERID-GS adapts to new scenes with only a few shots and enables stable low-light novel view synthesis from sparse-view observations. In addition, we construct a low-light multi-view dataset covering full 360∘ scenes for joint evaluation. Thorough experiments across multiple datasets in this area demonstrate that MERID-GS achieves SOTA performance, exhibiting superior cross-scene generalization and view consistency. The source code and pre-trained models are available at https://github.com/YhuoyuH/MERID-GS..
YuHao Yin, Zongji Wang, Yuanben Zhang +3
Aerospace Information Research Institute, Chinese Academy of Sciences · Key Laboratory of Target Cognition and Application Technology (TCAT) · University of Chinese Academy of Sciences +2
Novel-view synthesis and 3D reconstruction from sparse posed images are central to robotics and AR/VR. Yet, feed-forward 3D Gaussian reconstruction fails under lowlight due to noise, color shifts, and unreliable correspondence. We propose DelowlightSplat, a lowlight-aware feed-forward Gaussian splatting framework for clean novel-view rendering. We build a controllable multi-view lowlight benchmark by degrading only context views while keeping target views clean. We introduce a lightweight Lowlight Adapter for residual enhancement to improve matchability, and couple it with cost-volume-based multi-view inference to directly predict clean 3D Gaussians. Experiments show that DelowlightSplat significantly outperforms previous feed-forward method and two-stage pipeline under lowlight conditions.
Fuzhen Jiang, Zengtian Xie, Zhuoran Li
Hangzhou Dianzi University · Hangzhou, China · Zhuhai College of Science and Technology +1
Reconstructing 3D scenes under real-world low-light conditions remains challenging due to severe sensor noise, low signal-to-noise ratios, and degraded photometric consistency, which destabilize geometry estimation and novel view synthesis. Existing approaches often rely on well-lit reference data for reliable Structure-from-Motion (SfM) initialization under degraded inputs or apply per-view enhancement methods that introduce cross-view inconsistencies. To address these limitations, we propose NOVA-GS, a unified noise-aware framework for low-light 3D Gaussian Splatting that subsumes enhancement, denoising, and geometry optimization within a single process. Our method leverages VGGT-based feed-forward estimation to obtain robust camera poses and geometry directly from degraded inputs, eliminating the need for SfM. Building on this initialization, NOVA-GS integrates three coupled components: a structure-aware enhancement module for exposure correction, a self-supervised denoising module with blind-spot masking for pseudo-supervision, and a consistency-driven Gaussian Splatting optimization enforcing cross-view geometric coherence. We further introduce a noise-guided spherical harmonic regularization to suppress view-dependent artifacts in noisy regions. Extensive experiments on diverse real-world low-light datasets demonstrate improved geometric fidelity, color consistency, and robustness without requiring paired supervision or well-lit references. https://shaurya2524.github.io/nova-gs/
Shaurya Pavan A, Vemunuri Divya Madhuri, Yash Pradeep Gawande +1