MissFortune post is very interesting, but may lead to confusion. Denoising indeed is a low-pass filter (like a gaussian blur), but it is only applied to outlier pixels. An outlier pixel is supposed to be significantly different form all its surrounding pixel. The first denoisers used a more or less large neighborhood and an empiric threshold to decide if a pixel is an outlier or not. Recent ones use an AI training on a large region and are extremely accurate. For instance, they would not consider skin pores as outliers.
But they can still do errors and misclassify a pixel.
To minimize these errors, here is how I proceed.
1/ I render the image at twice the desired resolution, but with a limited number of iteration (around 300-400 depending on the lighting). I generate the image as a png, to avoid lossy compression artifacts (that are basically low pass filters).
2/ I denoise in post process using an external denoiser. As the image is twice the final size, and that only single pixel outliers are suppressed by the denoiser, even if a pixel is wrongly denoised, its impact on the final image will be very weak. Again, I save the image as png.
There are two main (free) denoisers: intel and nvidia. I prefer intel, because I mostly postprocess on my laptop (without a GPU), but both denoisers are very good (and better than daz integrated denoiser).
3/ I do the downsizing and generate the final webp image.
All is done with command line tools (denoiser or ImageMagick) and to limit all manual operations, I have a script that applies steps 2/ and 3/ to all images in a directory if denoised images are non-existent or older that the rendered ones.
Besides its interest in terms of denoising, having a double sized image is very useful. For instance, if you are not completely happy with the framing and need to do some cropping, or want to do a close-up on an image. And of course, if I need to do some post processing (retoning, darkening, blurring, blending/adding images, etc), I also do that before downsizing.
For animations, I proceed similarly, but I generally reduce the number of iterations (say 250), and I downsize more aggressively (generally x3). If I want to add some video effects (zooming, pan, etc), it is obviously better to do that on the full-scale image. Ditto if you want to retime your frames (for instance for slow motion).
The main drawback of this method is that you can have to keep several large png images on your disk.
