Tri-DehazeGS Separates Scene and Haze in 3D Reconstruction from Photos

A group of researchers presented Tri-DehazeGS — a method for reconstructing clean 3D scenes from smoky or hazy images. The work was published on September 10, 2026 on arXiv. The authors propose separating the clean geometry of the scene from atmospheric haze: the scene is described by Gaussian primitives, while the medium is described by an independent tri-plane field shared across all views. Additionally, gradient compensation is introduced for low-transmittance regions so that distant and heavily smoky areas are not lost during optimization.

What it means

Tri-DehazeGS solves the task of reconstructing a scene from multiple images, not generating a model from a single shot. In our reference of AI tools for 3D, generators of a different class are currently featured: OpenLRM, Hunyuan3D 2.1 and TRELLIS — all of them work on the image-to-3d scheme, that is, they create a model from a single image. They have no direct overlap in task with Tri-DehazeGS. Nevertheless, the method touches on a problem common to all Gaussian approaches: reconstruction quality under poor visibility conditions. If a tool for reconstructing scenes from photos with smoke, fog or other atmospheric interference appears in the future, the approach of separating scene and medium could become its foundation. For now, the fact sheets of our tools have no such functionality: none of them claims to work with smoky or hazy sources. The Tri-DehazeGS code has been released as open source, but the authors do not specify hardware requirements or the export format in the abstract.