Abstract Computer vision increasingly uses synthetic data from physically based rendering to supplement limited real‐world datasets. In industrial inspection, defect data is scarce and the rendering pipeline is explicitly controlled, and synthetic defect generation therefore becomes a dataset design problem. In this setting, building a dataset means choosing points in a rendering parameter space: defect shape, material, illumination, viewpoint, and sampling define the training distribution, but their effect on downstream learning is often hard to judge from images alone. We therefore study how these factors change defect features, where their relation to downstream learning is easier to inspect. Our results show that the rendering factors do not matter equally: defect shape, material, illumination, and viewpoint often affect downstream behavior much more than the number of samples per pixel. Synthetic subsets that transfer better downstream tend to stay close to real defect features, cover the observed defect modes, and stay separated from the defect‐free (OK) region. Based on these observations, we build a simple feature‐space screening heuristic for selecting subsets from large candidate pools. The selected subsets often outperform matched random selection for downstream segmentation on real data.

Denoising is an important post-processing step in physically based Monte Carlo (MC) rendering. While neural networks are widely used in practice, statistical analysis has recently become a viable alternative for denoising. In this paper, we present a general framework for statistics-based error reduction of both estimated radiance and variance. Specifically, we introduce a novel denoising approach for variance estimates, which can either improve variance-aware adaptive sampling or provide additional input for image denoising in a cascaded manner. Furthermore, we present multi-transform denoising: a general and efficient correction scheme for non-normal distributions, which typically occur in MC rendering. All these contributions combine to a robust denoising pipeline that does not require any pretraining and can run efficiently on current GPU hardware. Our results show distinct advantages over previous denoising methods, especially in the range of a few hundred samples per pixel, which is of high practical relevance. Finally, we demonstrate good convergence behavior as the number of samples increases, providing predictable results with low bias that are free of hallucinated neural artifacts. In summary, our statistics-based algorithms for adaptive sampling and denoising deliver fast, consistent, low-bias variance and radiance estimates.