Comparative Study of Path Tracing Rendering Parameters and Visual Quality in Blender

Abstract. This study investigates the impact of path tracing samples per pixel count on image quality in Blender, within the framework of 4D reconstructions of partly destroyed castles. These 4D models integrate both spatial and temporal dimensions. High-quality rendering is essential for both research and communication purposes, but it comes with significant computational costs. The study focuses on identifying a balance between visual quality and rendering time by analysing how image quality evolves with the number of rendering iterations. The evaluation uses the Mean Structural Similarity Index Measure (MSSIM), a perceptual metric that reflects human visual sensitivity more effectively than traditional methods such as the Root Mean Square Error (RMSE) or the Peak Signal to Noise Ratio (PSNR). Each test image is compared to a reference image used as the ground truth. Images are rendered with increasing numbers of samples per pixel while maintaining all other scene parameters fixed to ensure comparability. The study shows a clear MSSIM convergence, indicating that visual quality improves significantly with more iterations, but converges after a certain threshold. Pixel-wise SSIM maps are also generated to provide local information into the spatial distribution of similarity across the images. In addition, the study examines the role of Blender’s built-in denoising algorithms, evaluating their effectiveness in enhancing perceived image quality and their potential to reduce necessary iteration counts. By quantifying the relationship between samples per pixel and image quality, this research aims to define a rendering strategy for heritage applications. The goal is to minimise rendering time without compromising the visual standards required for documentation, analysis, and public dissemination of digital reconstructions.