High-Fidelity 4× Neural Reconstruction of Real-time Path Traced Images

The growing demand for high-resolution, realistic rendering has significantly increased the workload for real-time path tracing, putting considerable strain on most graphics cards. One common approach to reduce the burden is rendering images at lower resolutions with fewer samples per pixel (spp) and then applying reconstruction techniques, such as denoising and upsampling, to achieve high-quality rendering at the desired resolution. Recovering fine details directly from noisy, low-resolution (LR) inputs in high-resolution (HR) images is challenging. There-fore, a two-stage reconstruction process is typically employed, treating denoising and upsampling as separate steps. We propose an integrated deep learning model based on a Generative Adversarial Network (GAN) to produce high-quality reconstructed images - i.e., denoised and up-sampled - for efficient, adaptive image reconstruction. We employ a self-adaptive data augmentation strategy to iteratively reduce residual noise and recover high-frequency details, continuously improving the output. Our method significantly reduces rendering overhead while minimizing quality loss. Experimental results show consistent performance in demanding 4 x upscaling scenarios, delivering real-time performance with enhanced quality and substantial improvements.