I had an idea today to create a matching value with some machine learning. Here I wanted a value for similarity. For parts of a rendered image. With that of the same part of a higher detailed image. The idea was to speedup blender.

I’m not finished yet. Actually I’m not the disciplined programmer I want to be. However I came up with an idea that I want to share.

I had a target vector with little variation. Actually many zeros and a single one value. I don’t know if my idea works but what I figured out was that you could transform a single value to an object of many values and then use statistical methods to get the target value back.

So if you have a target vector with 5 elements. With zeros as values. Then you could see what happens when you insert 100 random samples with a mean value of zero as a replacement object. Then for each of those zeros you would suddenly have 100 values. Giving you the matrix size of 5×100. However taking the mean value of each of the rows gives you back the original value.

So the idea is to train the machine learning setup with more values then get the wanted value by taking the mean of the output.

Depending on factors like the amplitude of the random values. This would effect the outcome of the training.

An example of a universal sample generator. The sun.