基于条件生成式对抗网络的高质量动态实时渲染方法  

High-quality dynamic real-time rendering method based on conditional generative adversarial networks

作　　者：江李铠 王国中 赵海武 JIANG Likai;WANG Guozhong;ZHAO Haiwu(Institute of Artificial Intelligence,School of Electronic and Electrical Engineering,Shanghai University of Engineering Science,Shanghai 201620,China)

机构地区：[1]上海工程技术大学电子电气工程学院人工智能实验室,上海201620

出　　处：《上海工程技术大学学报》2024年第4期451-457,共7页Journal of Shanghai University of Engineering Science

摘　　要：聚焦计算机图形学中的实时渲染挑战,通过结合光栅化技术和优化后的条件生成式对抗网络(conditional generative adversarial networks,CGANs),实现实时生成近似光线追踪图像,解决现有研究中生成的帧与帧之间不连贯的问题,实现实时性、真实感和视觉连贯性之间的优化平衡。基于Pix2PixGAN架构,对CGANs进行结构、数据输入和损失函数方面的改进,并利用Unity和Blender构建一套训练渲染数据集。结果表明,本研究提出的渲染方法在关键性能指标上优于传统方法,显著提升了图像生成的质量以及帧与帧之间的连贯性。Focusing on the challenge of real-time rendering in computer graphics,integrating rasterization techniques with optimized conditional generative adversarial networks(CGANs),real-time generation of approximate ray-traced images was achieved,the issue of discontinuity between frames in existing research was effectively addressed,and optimized balances among real-time performance,realism,and visual coherence were achieved.Based on Pix2PixGAN architecture,the structure,data input and loss functions of CGANs were improved,a training rendering dataset using by Unity and Blender was constructed.Experimental results demonstrate that our rendering method can surpass traditional approaches in key performance metrics,enhance the quality of image generation and the coherence between frames.

关 键 词：实时渲染 光栅化 光线追踪 生成式对抗网络 渲染管线 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]