SynthLight
收藏资源简介:
SynthLight数据集由耶鲁大学和Adobe Research共同创建,旨在通过3D渲染技术模拟不同光照条件下的人像图像。该数据集包含约126万张图像,每张图像均由Blender的Cycles渲染器生成,分辨率为512×512。数据集通过350个主题的3D头部模型生成,每个主题包含多种外观变化,如不同发型、肤色、表情和服装。数据集的创建过程涉及使用高质量的PBR纹理贴图和环境光照图进行渲染。该数据集主要用于训练扩散模型,以实现肖像图像在不同光照条件下的重新照明,解决肖像摄影中光照控制的难题。
The SynthLight dataset was co-created by Yale University and Adobe Research, aiming to simulate portrait images under various lighting conditions via 3D rendering technology. It contains approximately 1.26 million images, each generated using Blender's Cycles renderer with a resolution of 512×512. The dataset is developed based on 3D head models of 350 subjects, with each subject featuring diverse appearance variations including different hairstyles, skin tones, facial expressions and clothing. The dataset creation process involves rendering with high-quality PBR texture maps and environment lighting maps. This dataset is primarily used for training diffusion models to achieve relighting of portrait images under different lighting conditions, addressing the challenge of lighting control in portrait photography.

- 1SynthLight: Portrait Relighting with Diffusion Model by Learning to Re-render Synthetic Faces耶鲁大学, Adobe Research · 2025年



