I'm currently trying to understand image generation a bit better. I'm working on a DDPM to generate samples from the MNIST set. My question doesn't really have anything to do with that, it's more just out of interest.

Now I was wondering if it is possible to have a divergence between the number of input and output channels. So for instance, if I have 6 channels, RGB and 3 other values, how would I generate an output that is only the RGB image, so 3 channels? I know that you can use labels to do supervised learning but what if these other 3 channels are floats such as recording time.

My initial thought was to use a table as the input with (R, G, B, x, y, z) being the columns and the rows being each pixel. But that wouldn't make a lot of sense with x, y and z being the same for each pixel in an image.

And now what if I train on these 6 channels but I only want to generate one? So for instance only an R band.

And how would the kernel for such an operation look like?

I would apprechiate any thoughts on this!



You must log in to answer this question.