π-Light Dataset
收藏资源简介:
π-Light数据集由南洋理工大学S实验室与腾讯联合构建,专为全图像重光照任务设计,包含从Objaverse筛选的10,000个具有BRDF材质属性的3D对象,每个对象在10种视角和10种光照条件下渲染,总计生成百万级图像。数据通过BlenderProc工具生成,采用Poly Haven的700余种HDRI环境贴图进行光照模拟,并创新性地使用前半球灰球表示光照条件以提升控制精度。该数据集支持物理逆向渲染与神经前向渲染的联合训练,解决了传统方法在材质一致性、光照控制及跨场景泛化方面的瓶颈,为影视特效、增强现实等领域的可控重光照研究提供了关键基准资源。
π-Light Dataset is co-developed by the S-Lab at Nanyang Technological University (NTU) and Tencent, specifically designed for the full-image relighting task. It contains 10,000 3D objects with BRDF material properties selected from Objaverse. Each object is rendered under 10 viewing perspectives and 10 lighting conditions, generating a total of millions of images. The dataset is generated using the BlenderProc toolkit, adopts over 700 HDRI environment maps from Poly Haven for lighting simulation, and innovatively uses the front-hemisphere gray sphere to represent lighting conditions to improve control accuracy. This dataset supports joint training of physics-based inverse rendering and neural forward rendering, addressing the bottlenecks of traditional methods in material consistency, lighting control and cross-scene generalization, providing a critical benchmark resource for controllable relighting research in fields such as visual effects and augmented reality.
PI-Light 数据集概述
基本信息
- 数据集名称:PI-Light: Physics-Inspired Diffusion for Full-Image Relighting
- 相关论文状态:已接受至ICLR 2026
- 论文预印本:https://arxiv.org/abs/2601.22135
- 代码发布状态:代码将于三月发布
作者与机构
- 作者:Zhexin Liang (第一作者), Zhaoxi Chen, Yongwei Chen, Tianyi Wei, Tengfei Wang, Xingang Pan
- 所属机构:
- S-Lab, Nanyang Technological University
- Tencent
核心内容
- 研究主题:基于物理启发的扩散模型,用于全图像重照明任务。




