News - Sunday April 3rd, 2022

Disco Diffusion Experiments

I am dedicating most of my free time to image generation by running a notebook on Google Colab called Disco Diffusion: it is an artificial intelligence based on the clip guided diffusion.

It is difficult to feel completely the author of my generated images, yet if I had not written exactly that “prompt”, with that seed and those settings I would not have obtained that image.
It takes a long time to understand how AI interprets your words: the result is not obvious and sometimes something horrible comes out, while other times it leaves you speechless.

The generated image is unique and an identical one will never be generated.
The AI ​​did it by fishing from its dataset, but it did not copy it as many think: Disco Diffusion starts from a gaussian or perlin noise image from which it begins to remove noise for a number of times that is defined in the settings , until reaching a result that takes its cue from a dataset of millions of images accurately described during the training. Its purpose is to be as consistent as possible with the prompt.

It seems like a kind of game to me, at the same time I realize that I have a very powerful tool in my hand. I imagine infinite future developments: some excite me, others frighten me.

<<