The picture on the right actually works surprisingly well given that only one sample is taken. It seems that resampling this computed shadow on the surface by averaging nearby samples will give a reasonable outcome in this case.
Just for clarification, "sample center of light" means literally sampling the center of the area light source, and "sample random point on light" means sampling a random point of the area light source?
Is there ever a situation where someone might prefer the picture on the right? For example, if someone wanted to use it in a certain art style would the number of samples be a hyperparameter to control? Or would the approach be to still take many samples, average it out, and then downsample to get some desired effect?
@hangg7 yes, this is correct
@selinafeng this type of sampling is very non-deterministic, which may make it difficult to produce the desired effect. However, there are also efficiency and other tradeoffs that you may need to take into account.
The picture on the right also reminds me of how something looks like through a rainy car window or a shallow puddle in a gravely area. I wonder if there is some connection to small amounts of water diffusing light and having 1 sample per pixel.
If you told me the ground was say smooth cobblestone the right picture looks realistic.