Diffusion Models

--

you have a lot sprite images

this is your training data.

You want: even more sprites

This is a neural network that generates more sprites for you.

the neural network learns to take different noisy images and turn them back into sprites

it learns to remove the noise you added

no idea of noise is important becuase it is normally distributed

so, when u ask the neural network for a new sprite

you can sample noise from the normal distribution

get a completely new sprite by using the net to remove the noise

Now you can get even more sprites, beyond your training data.

--

--

Sivanarayana Mamidi

Computer Vision Engineer, Master in artificial intelligence, enthusiastically to solve DSA problems