arXiv Computer Vision

RetinexGS: Retinex-Inspired Gaussian Splatting for Low-Light RAW Reconstruction

RetinexGS is a RAW-domain 3D Gaussian Splatting framework that improves low-light novel-view synthesis by combining Retinex-inspired appearance modeling with sensor-informed optimization. It factorizes each Gaussian’s color into a view-independent base color and a shared, position- and view-dependent modulation field, regularizing them separately to handle noisy observations. Using DNG noise metadata to weight the reconstruction loss and depth-guided initialization, RetinexGS outperforms LE3D and RawNeRF by 2.18 dB and 2.71 dB in mean sRGB PSNR across five scenes while enabling real-time rendering.

arXiv Computer Vision
Oct 7

SteadySplats: Resampling of Low-Variance Gaussians for High-Fidelity Stochastic Rendering

SteadySplats introduces a method to reduce high‑frequency noise in stochastic rendering of 3D Gaussian Splatting models. The approach combines a history‑based spatial resampling scheme for faster image convergence, temporal importance resampling for camera‑movement coherence, and a color regularizer during training to lower variance along view rays. The resulting Vulkan‑based renderer achieves a 13 dB PSNR improvement over prior stochastic methods at one sample per pixel and converges quickly to a low‑error sorted 3DGS representation.

By Felix Windisch, Thomas K\"ohler, Lukas Radl, Chris Wyman, Georgios Kopanas, Bernhard Kerbl, Markus Steinberger
arXiv Computer Vision
Oct 16

Naka-GS: Naka-Guided Chroma Correction and Point-Prior Compaction for Low-Light 3D Gaussian Splatting

Naka-GS is a low‑light 3D Gaussian Splatting framework that jointly enhances observations and improves Gaussian initialization. It introduces Naka‑Guided Chroma Correction, which uses a fixed bionic response as a photometric anchor to learn bounded, frequency‑decoupled luminance and chromatic corrections, preserving fine structures while reducing color and boundary artifacts. A lightweight, non‑learned Point‑Prior Compaction Module refines dense scene priors via voxel pooling and distance‑adaptive pruning, cutting the number of points from 1.00 M to 59.52 K and shortening optimization time by 37.7 % while achieving state‑of‑the‑art PSNR, SSIM, and LPIPS on eight RealX3D scenes.

By Runyu Zhu, SiXun Dong, Zhiqiang Zhang, Qingxia Ye, Rongjun Qin, Zhihua Xu
arXiv Computer Vision
Aug 31

WilLaGS: Latent-Conditional 3D Appearance Fields for Robust Gaussian Splatting In-the-Wild

WilLaGS introduces a unified framework that enhances 3D Gaussian Splatting for in-the-wild scenes by learning a continuous global appearance manifold with a β‑VAE and generating dynamic Tri‑Plane features for spatially‑varying local illumination. It also incorporates a self‑supervised perceptual masking mechanism using a Teacher‑Student EMA architecture to suppress transient artifacts and identify inconsistent regions. Experiments on multiple datasets show that WilLaGS achieves state‑of‑the‑art reconstruction quality and novel view synthesis while preserving real‑time rendering efficiency.

By Yuhao Bai, Qianqiu Tan, Lilong Chen, Huanhuan Lv, Lijun Chen
arXiv AI
Sep 28

Spackle: Completing Large View Single Image NVS with Adaptive Gaussians

Spackle is a lightweight residual learning framework designed to improve large-view single-image novel view synthesis (NVS) by mitigating capacity competition in hybrid decoupled systems that combine 3D Gaussian Splatting (3DGS) and diffusion models. It operates in three stages: predicting base 3DGS attributes, automatically identifying poorly reconstructed regions, and learning a residual 3DGS focused on those areas. During inference, Spackle merges the baseline and augmented Gaussians to produce high-fidelity novel views, achieving state‑of‑the‑art performance on large-view-deviation cases.

By Xuanzhi Liu, Yuhe Zhou, Xinyi Wu, Zhenyao Wu, Jinghao Chen, Ruize Han, Song Wang
arXiv Computer Vision
1d ago

DIRT: Depth-completed Image-based Rendering using Two-band Blending for Live Novel-view Synthesis

The paper presents DIRT, a live novel-view synthesis method that uses depth-completed image-based rendering with a two-band blending strategy. It operates on a posed set of depth sensor observations and images, completing a sparsely filled depth buffer in screen space and fusing projected images without any neural networks or scene-specific optimization. The approach is non‑neural, training‑free, and works across heterogeneous sensors such as RGB‑D cameras, LiDAR, and thermal cameras, achieving lower latency and competitive performance in both live and conventional NVS benchmarks.

By Mathias Harrer, Laura Fink, Noah Lewis, Paul Himmler, Tim Weyrich