crates/optix/src/denoiser.md
Image areas that have not yet fully converged during rendering will often exhibit pixel-scale noise due to the insufficient amount of information gathered by the renderer. This grainy appearance in an image may be caused by low iteration counts, especially in scenes with complex lighting environments and material calculations.
The NVIDIA AI Denoiser can estimate the converged image from a partially converged image. Instead of improving image quality through a larger number of path tracing iterations, the denoiser can produce images of acceptable quality with far fewer iterations by post-processing the image.
The denoiser is based on statistical data sets that guide the denoising process. These data, represented by a binary blob called a training model, are produced from a large number of rendered images in different stages of convergence. The images are used as input to an underlying deep learning system. (See the NVIDIA Developer article “Deep Learning” for more information about deep-learning systems.)
Because deep-learning training needs significant computational resources—even obtaining a sufficient number of partially converged images can be difficult—a general-purpose model is included with the OptiX software. This model is suitable for many renderers. However, the model may not yield optimal results when applied to images produced by renderers with very different noise characteristics compared to those used in the original training data.
Post-processing rendered images includes image filters, such as blurring or sharpening, or reconstruction filters, such as box, triangle, or Gaussian filters. Custom post-processing performed on a noisy image can lead to unsatisfactory denoising results. During post-processing, the original high-frequency, per-pixel noise may become smeared across multiple pixels, making it more difficult to detect and be handled by the model. Therefore, post-processing operations should be done after the denoising process, while reconstruction filters should be implemented by using filter importance-sampling.
In general, the pixel color space of an image that is used as input for the denoiser should match the color space of the images on which the denoiser was trained. However, slight variations, such as substituting sRGB with a simple gamma curve, should not have a noticeable impact. Images used for the training model included with the NVIDIA AI Denoiser distribution were output directly as HDR data.