markov chain monte carlo post processing


I have tried the mcmc of the optix examples wit my own obj files. It seems that regardless of how long I leave the simulation running the image remains noisy and does never completely converge to a clear image.

I have tried even with tripling the image resolution and then to resize the image to its original size but the results still appear quite noisy.

Which possibilities are there to either improve the convergence to a better image or to filter the image afterwards to remove noise?

Many thanks,

The way mcmc works is that it creates a predetermined number of samples. Running it longer only moves the samples about.

If you want a less noisy image, you will need to increase the number of samples. There are different ways to go about this. You could change the initial number of samples, you could render the same image multiple times and then combine them (make sure to use different seeds), you could also do what you tried and render a larger image and down sample it.

There could also be problems where the chain gets stuck in a local minima. I’m not sure what to do about that. You should probably check out the literature.

Another issue might be related to tone mapping. If you have a few bright samples they can blow out a pixel. A tone mapping pass will help with that.