Create images with Easy Diffusion AI
As a member of Generation X, it’s not often I find out something new, but over the past 24 hours I’ve learned more about AI than I have in the past…well, since ever. More specifically, I’ve learned how to create images using Easy Diffusion. Easy Diffusion is the same thing as Stable Diffusion, but it’s kinder and gentler to install and run. A few months ago I tried installing regular Stable Diffusion and wasn’t able to get it to work correctly. Yesterday I read a blog post at JuicyBomb that mentioned Easy Diffusion, and my life hasn’t been the same since.
Generative AI and Machine Learning
Over the past year, it’s been impossible to visit technology websites and not see articles about AI. The articles are everywhere, and if we go by the Gartner Hype Cycle we are probably on the way to the Peak of Inflated Expectations. We’ll see how it goes, but I suspect we won’t see the hype peak until 2024, mainly because I expect it will become a political football during the 2024 US presidential election.
Not wanting to be left behind, I thought I’d contribute my thoughts to the hype cycle to talk about what it is and what it isn’t. Most of the recently written articles about AI aren’t really about AI as in Artificial Intelligence. This isn’t HAL or Skynet, and we shouldn’t need to worry about evil robots flinging nukes. What the articles are really discussing is Machine Learning (ML), which is being used with Generative AI to create images, texts, music, and more.
Machine Learning is the term used to describe computer models that have been trained on something. Generative AI takes these models and uses them to create something new. In the case of Easy Diffusion, it takes models that have been trained on images; from a few dozen to hundreds of thousands, and it can create a new image when prompted.
This isn’t all that different than how people learn to do things. If people want to learn how to write, they look at what others have written. If they want to learn how to paint, they begin by looking at paintings created by others. With humans, we also tend to practice a lot, but I think “practice” for ML and Generative AI depends a lot on the programs used to create them.
To get Easy Diffusion to create an image, it needs to be given a prompt. The default prompt for Stable Diffusion and Easy Diffusion is “a photograph of an astronaut riding a horse”. Doing an online image search returns a plethora of images of astronauts riding horses. Not wanting to create anything so mundane, the first prompt I gave it was “a photograph of an astronaut riding an axolotl” and the results weren’t what I expected.
For the images above and the ones show below, the image on the left was created on a Dell Latitude 7480 laptop with a Core i7 CPU and 32 GB of memory (it was a refurb I bought a few years ago), while the image on the right was created on my home PC with an AMD Ryzen 7 CPU and 64 GB of RAM. The laptop is a basic laptop, not a gaming laptop, so there’s nothing fancy in the graphics department, but the PC has an older Radeon RX 580. Still nothing fancy, but better than the laptop. Both computers are running Linux Mint 21.1 Cinnamon.
As you can see, neither image actually shows an axolotl. Axolotls are really strange and somewhat unknown, so maybe I shouldn’t be surprised that Easy Diffusion doesn’t know how to create an image of one. I still didn’t want to show an astronaut riding a horse, so the next prompt I gave it was “a photograph of an astronaut riding a camel”. I figured that camels are well-known enough that the program should know what one is and how to draw one. The images shown below depict exactly what I prompted Easy Diffusion to create.
These images looked promising and are more or less what I would have imagined. Getting bored with astronauts pretending to be cowboys, I decided to changes things up a little. In addition to changing the size of the images, the next prompt I gave it was “A portrait of Hermione Grainger”. I was expecting something along the lines of Emma Watson in a school uniform, so the images that it came up with were unexpected.
I suspect that prompting Easy Diffusion for a portrait causes it to think of old photographs or paintings as ancestral portraits. That would explain the look of the “painting” on the left and the vintage photo looks of the one on the right. One of the primary issues with images generated by AI is that they frequently depict things that are impossible in the real world, as demonstrated by the image on the right which appears to show a woman with two heads stacked upon each other.
Being denied images of Emma Watson, and hoping for more realistic images, the next prompt I gave Easy Diffusion was simple and direct; “Smirking Emma Watson”. The results below didn’t disappoint.
These images were pretty good. While I wouldn’t mistake them for photographs of the actor, they’re good enough that it would be difficult to mistake them for anyone else.
Maybe it’s just because of the novelty of something new, but I’ve been all in on this technology since the first images of astronauts began showing up on my computers yesterday afternoon. I even dreamed about Easy Diffusion last night and was disappointed when I woke up and found the images in my head were gone forever. I’ve been playing with it some more and may write some more about it, but for now I think it’s at the stage where people should start playing with it on their own. It’s free, it’s easy to install, and fairly easy-to-use after some trial and error. This seems to eliminate any risks and encourages people to play with it and see what it can do.