Search papers, labs, and topics across Lattice.
DelowlightSplat addresses the challenge of 3D Gaussian reconstruction from sparse images under low-light conditions by introducing a feed-forward Gaussian splatting framework. The method employs a Lowlight Adapter for residual enhancement to improve feature matching and integrates it with cost-volume-based multi-view inference for direct prediction of clean 3D Gaussians. Experiments on a newly introduced low-light benchmark demonstrate that DelowlightSplat significantly outperforms existing feed-forward and two-stage pipelines in generating clean novel-view renderings.
Low-light 3D scene reconstruction from sparse views is now possible with feed-forward Gaussian splatting, thanks to a new method that directly predicts clean 3D Gaussians even when context views are noisy.
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.
