Stable Diffusion is a really big deal

[Ed P]

Stable Diffusion merges AI,  and a written description  to produce graphics output.  I’ll confess my first reaction was that this is overblown hype of the ‘programming without code’ school. Some links from HackerNews have made me revise my estimation, and it appears that it could become a useful tool for graphics artists and game designers.

First the pretty picture aspect is covered in this link, where an image is generated from a text description such as the following:

“A distant futuristic city full of tall buildings inside a huge transparent glass dome, In the middle of a barren desert full of large dunes, Sun rays, Artstation, Dark sky full of stars with a shiny sun, Massive scale, Fog, Highly detailed, Cinematic, Colorful”

So far, quite impressive but then someone thought of animating between two scenes, and came up with the impressive transition shown in the following link.

What is even better is that the code has been open sourced and you can play with it on your own PC.

 

[Ed P]

Just to expand on this a little. Graphical AI programs need zillions of labelled images to use as training datasets.

As an aside this is the key area where hidden racial and gender bias comes into play. As an exaggerated example, if a Police criminal dataset has a disproportionate number of green people labelled as criminals, it is highly likely that all green people will then be labelled as potentially suspicious on Police CCTVs.  Hidden bias goes further than this as it has been found that the AI frequently uses texture rather than object boundaries or other humanly obvious characteristics. This could mean that all green people with acne could be doubly suspicious in such an imaginary police computer system!  The nasty side of AI is that currently AI does not have to explain which characteristics are most relevant too its classifications. (It is actually quite a difficult and laborious task to find this out.)

But back to topic – the 2.3 billion web image dataset initially used in Stable Diffusion is obviously too large for any single human to review, but a 12 million subset  has been grabbed and can be explored by you.

This link gives some details on how you can waste a couple of days viewing some of these images!

 

 

[Author]

Posts