Opinion
The artifice of art: how AI is shaking up the art market
by Michelle Brändle
AI creates images from texts in real time
30.11.2023
Translation: machine translated
There are now several AI generators that create an image from a text. Patience is always required when using them, as it often takes several minutes to get a result. Stability AI changes that.
Stability AI is a company that creates text, image and video generators based on deep learning models. They call their text-to-image generator Stable Diffusion. This spits out an image based on a descriptive text, called a prompt. The first version has been around for over a year, with an improved successor, the Stable Diffusion XL.
Source: Stability AI
However, both the first version and its successor take several minutes from text input to result. Stability AI has now released a new version of its image generator as a demo. The so-called Stable Diffusion XL Turbo (SDXL-Turbo) is based on the Stable Diffusion XL model and, unlike its bigger brother, delivers results in a matter of seconds.
I have already covered how the generators work in more detail in previous articles. The principle as a user is simple: you enter a text, called a prompt, and the programme spits out one or more images based on this description. This is great for inspiration for creative professionals, for example.
Since a programme like this has to access a fairly large database to produce a result, it requires enormous computing power. And it takes a lot of time. If you don't like the result afterwards, you have to amend, change or even rewrite your text. The process starts all over again. That can be tedious.
What's new with SDXL Turbo?
With the SDXL-Turbo, Stability AI has created a programme that delivers images from the prompts so quickly that they themselves call it a real-time result. The resulting images are not quite as detailed as the slower version, but are still of quite good quality. However, they are impressive for the speed of their creation alone.
Source: Stable Diffusion XL Turbo
Normally, the process of an image generator involves 20 to 50 steps. The innovation of the SDXL-Turbo is that the programme only requires up to four steps for the output. Stability AI calls this procedure "Adversarial Diffusion Distillation". It is quite complex, and it is not for nothing that Stability AI has explained the theory behind it in detail.
The exciting thing is that the result not only saves time, the process also requires less computing power according to Stability AI. So even my Macbook Air with M1 chip has enough power for some funny results. You can find the demo version on this website. After about 10 sentences and results, however, the free version is used up. The images can be edited even further with an editor button. For example, you can remove the background at the touch of a button or display three more similar results. Again, these edits take a little longer.
Source: Stable Diffusion XL Turbo
In my world, Super Mario chases Stormtroopers with a unicorn and Harley Quinn mixes cocktails for Eddie and Peter at the beach bar. Wherever I can live out my creativity, my fingers tingle. Or maybe it's because nothing flows through my veins but chocolate, glitter and coffee.