Key points
- Bako et al. (2017) and Vogels et al. (2018) introduce kernel-predicting networks that denoise Monte Carlo renders using predicted per-pixel kernels, making production-quality rendering far less compute-intensive.
- An April 2025 Academy Scientific and Engineering Award recognised the Disney Research, Intel OIDN and Weta FX denoisers as a category.
- ML denoising is the production standard across CG feature pipelines, and its assumptions and failure modes are core rendering-literacy topics.
Summary
Vogels et al. (2018) and the preceding Bako et al. (2017) KPCN work introduce kernel-predicting convolutional networks for Monte Carlo render denoising, predicting a spatially varying denoising kernel at each pixel from noisy input and auxiliary feature buffers such as normals and albedo. The approach allows production renders to use far fewer samples per pixel while achieving final-frame quality, with ML inference replacing expensive sample accumulation. An April 2025 Academy Scientific and Engineering Award recognised the Disney Research KPAL work, Intel Open Image Denoise (integrated into Blender, Arnold, and V-Ray), and the Weta FX denoiser as a category, confirming that ML denoising has become an invisible production standard in CG feature pipelines globally.
Related items
- DifFRelight: Diffusion-Based Facial Performance Relighting
- A deep learning framework for character motion synthesis and editing (SIGGRAPH 2016 / SIGGRAPH 2026 Test-of-Time award)
Source
Source: ACM Transactions on Graphics (SIGGRAPH 2018) ↗ (Research)
Cite this item
Vogels, T., Rousselle, F., McWilliams, B., Roethlin, G., Harvill, A., Adler, D., Meyer, M. & Novak, J. (2018). ‘Production Machine-Learning Denoising: Kernel-Predicting Networks’, ACM Transactions on Graphics (SIGGRAPH 2018). Available at: https://studios.disneyresearch.com/2018/07/30/denoising-with-kernel-prediction-and-asymmetric-loss-functions/
Your reference manager can also read this page directly: with the Zotero (or Mendeley) browser connector installed, save it straight to your library. Whole-collection exports: RIS, BibTeX, CSL-JSON.