EvenSplat: Coupled 2D-3D Decomposition for Gaussian Splatting under Exposure and Illumination Variation
Tongyu Wu, Jacob Edwards, Ziteng Cui, Caigui Jiang, Cheng Wang
cs.CV
2026-10-01
EvenSplat couples 2D illumination decomposition with a Gaussian illumination field, gaining 3.17 dB PSNR over the strongest baseline under high-contrast light on a real-world benchmark.
High-fidelity reconstruction with 3D Gaussian Splatting (3DGS) and NeRF rests on one assumption: the same surface shows the same brightness and color in every photo. Real captures break it constantly. The camera re-meters between shots, a flash or a window lights one side of an object more than the other, and a harsh local light splits a single frame into blown-out and crushed regions. Standard 3DGS has no concept that any of this is lighting. It fits each image faithfully, and the capture-time illumination ends up baked into the recovered geometry and color.
The paper sorts the inconsistency into three modes: cross-view exposure variation (CEV), spatial illumination variation within one image (SIV), and high-contrast illumination (HCI). Current fixes have blind spots. Appearance embeddings in the NeRF-W lineage and exposure-correction methods such as Luminance-GS and PPISP all operate at the whole-image level. When illumination splits inside a single frame, these methods either keep the shadow or flatten the bright region.
EvenSplat borrows the Retinex factorization, observation = base appearance × illumination, and couples a 2D decomposition with a 3D one so each constrains the other during training.
Training runs 10,000 iterations; the illumination field and CRN freeze after 3,000, and ILEC activates at 5,000. The ablations show the two branches need each other: removing the 2D decomposition costs 3.60 dB on CEV and 4.94 dB on HCI, while the 2D branch alone is also far worse, because single-view priors cannot produce a cross-view consistent explanation.
The real-world benchmark consists of nine self-captured scenes (3 CEV, 4 SIV, 2 HCI), each paired with a second capture under uniform illumination as the evaluation reference.
| Setting | Best baseline PSNR | EvenSplat |
| CEV (real) | PPISP 17.79 | 19.53 |
| SIV (real) | PPISP 18.11 | 19.51 (+1.40) |
| HCI (real) | 3DGS+CHROMA 13.78 | 16.95 (+3.17) |
| CEV (simulated) | PPISP 19.22 | 20.98 (+1.76) |
| HCI (simulated) | PPISP 15.17 | 16.74 (+1.56) |
On the real SIV and HCI splits EvenSplat leads all three metrics. Two results need a discount: on simulated HCI it takes the best PSNR but trails PPISP on SSIM (0.501 vs 0.589), and on real CEV it again has the best PSNR while PPISP keeps SSIM and LPIPS. Correcting heavily clipped inputs trades pixel agreement against structural similarity.
In the cross-lighting experiment, the same scene is captured twice with only the light direction changed (left- versus right-dominant), each set is fitted independently, and the two base-appearance renderings are compared at matched poses. EvenSplat reaches 15.57 dB, 2.84 dB above Luminance-GS, the strongest baseline, indicating the recovered representation is less tied to capture lighting. On HDR-NeRF it maintains 37.6 dB full-scene rendering while its base-appearance output lifts mean PSNR against uniform-light ground truth from 13.23 to 17.91 over the strongest baseline. A control baseline, Tone-Matched 3DGS (global tone normalization only), underperforms vanilla 3DGS on every SIV and HCI scene, which rules out global brightening as the source of the gains.
Heritage digitization, indoor scanning, and robotic inspection cannot stage their lighting. This route needs no radiometric calibration and no RAW files; ordinary sRGB photos go straight in. Training cost stays at 3DGS scale (10,000 iterations on two RTX 3090s) and rendering stays real-time, and the ceiling-digitization case in the paper is a direct demo. The nine-scene benchmark with uniform-light references is itself a contribution in a subfield that mostly relies on digital perturbations. In scope this is a solid systems paper: the gains concentrate exactly where spatially uneven lighting breaks global correction, and 3DGS rendering itself is untouched.
The authors list four: the sRGB-domain decomposition is not guaranteed unique (base color is not physical albedo), colors lost to clipped highlights are physically unrecoverable, ILEC parameters are tied to training images, and specularities, motion blur, and sensor noise are unmodeled.
A few more from reading the paper. Real HCI covers only two scenes, so the headline 3.17 dB rests on a thin base. The simulated SIV and HCI operators (two-region blended exposure, contrast multiplied by three) fit the method's assumptions neatly; real lighting is rarely that tidy. The cross-lighting experiment measures consistency between two reconstructions, not accuracy, as the paper itself concedes. On HDR-NeRF base-appearance recovery, SSIM and LPIPS are mixed (CatRoom SSIM sits below plain 3DGS), so the PSNR lead deserves a discount.