Search papers, labs, and topics across Lattice.
This paper introduces Smoke-GS, a 3D Gaussian Splatting framework that integrates visual priors from a pre-processing network (Nano-Banana-Pro) and a view-dependent medium branch to reconstruct 3D scenes from smoke-degraded multi-view images. By explicitly modeling the view-dependent appearance variations caused by smoke, Smoke-GS improves robustness to smoke-induced degradation while maintaining rendering efficiency. Experiments demonstrate the method's ability to generate consistent and visually clear novel views in smoky environments.
Smoke-GS lets you see through the haze, reconstructing 3D scenes from smoky images with surprising clarity by explicitly modeling view-dependent smoke appearance.
Reconstructing 3D scenes from smoke-degraded multi-view images is particularly difficult because smoke introduces strong scattering effects, view-dependent appearance changes, and severe degradation of cross-view consistency. To address these issues, we propose a framework that integrates visual priors with efficient 3D scene modeling. We employ Nano-Banana-Pro to enhance smoke-degraded images and provide clearer visual observations for reconstruction and develop Smoke-GS, a medium-aware 3D Gaussian Splatting framework for smoke scene reconstruction and restoration-oriented novel view synthesis. Smoke-GS models the scene using explicit 3D Gaussians and introduces a lightweight view-dependent medium branch to capture direction-dependent appearance variations caused by smoke. Our method preserves the rendering efficiency of 3D Gaussian Splatting while improving robustness to smoke-induced degradation. Results demonstrate the effectiveness of our method for generating consistent and visually clear novel views in challenging smoke environments.
