📄 中文摘要
GR3EN方法旨在解决大型房间级别三维重建环境的重照明问题。当前三维场景重照明方案常面临欠定或病态逆渲染问题，难以在复杂真实场景中生成高质量结果。虽然近期生成式图像和视频扩散模型在重照明领域取得了显著进展，但这些技术通常受限于特定条件。GR3EN通过引入一种新颖的生成式方法，克服了传统逆渲染方法的局限性，特别针对大规模、高复杂度的真实世界三维环境。该方法不依赖于精确的物理模型或复杂的优化过程，而是利用深度学习的生成能力，直接从场景几何和少量照明信息中合成逼真的重照明效果。其核心在于构建一个能够理解场景几何、材质属性以及光照分布之间复杂关系的生成模型，该模型能够学习并内化这些关系，从而在新的照明条件下生成一致且高质量的图像。GR3EN利用多模态输入，包括几何数据（如网格或点云）和初始的低质量重照明结果，通过扩散模型迭代优化，逐步细化光照和阴影细节，最终生成与目标照明环境高度匹配的视觉效果。这种生成式框架允许在保持场景几何和材质特性的同时，灵活地改变光照方向、颜色和强度，实现对整个环境的全局重照明。GR3EN的优势在于其处理复杂真实场景的能力，能够应对多光源、复杂遮挡和多样材质等挑战，生成具有真实感的光照和阴影效果，显著提升了三维环境重照明的质量和效率。
📄 English Summary
GR3EN: Generative Relighting for 3D Environments
GR3EN introduces a novel method for relighting 3D reconstructions of large, room-scale environments. Existing solutions for 3D scene relighting frequently encounter under-determined or ill-conditioned inverse rendering problems, consequently struggling to produce high-quality results in complex real-world scenarios. While recent advancements in utilizing generative image and video diffusion models for relighting have shown promise, these techniques are often constrained by specific limitations. GR3EN addresses these challenges by presenting a generative approach that circumvents the shortcomings of traditional inverse rendering methods, particularly for large-scale, highly complex real-world 3D environments. This method does not rely on precise physical models or intricate optimization processes; instead, it leverages the generative capabilities of deep learning to directly synthesize photorealistic relighting effects from scene geometry and minimal lighting information. At its core, GR3EN constructs a generative model capable of understanding the intricate relationships between scene geometry, material properties, and light distribution. This model learns and internalizes these relationships, enabling it to generate consistent and high-quality images under novel lighting conditions. GR3EN utilizes multi-modal inputs, including geometric data (such as meshes or point clouds) and initial low-quality relighting results, to iteratively refine lighting and shadow details through a diffusion model. This process progressively generates visual effects that closely match the target illumination environment. Such a generative framework allows for flexible manipulation of light direction, color, and intensity while preserving scene geometry and material characteristics, thereby achieving global relighting for the entire environment. The primary advantage of GR3EN lies in its ability to handle complex real-world scenes, tackling challenges such as multiple light sources, intricate occlusions, and diverse materials to produce realistic lighting and shadow effects, significantly enhancing the quality and efficiency of 3D environment relighting.