How we make our pictures: local AI, sorting by hand, what we turn down
143 of the 147 pictures in our gallery are generated by artificial intelligence. We would rather say so up front and explain how they are made: on which computer, with which software, how we sort them, and what we refuse to publish.
Software on our own computer, not an online service
The pictures are made on a desktop computer with a gaming graphics card (an RTX 3070), using Fooocus, free image generation software that runs on the machine itself. The model it uses weighs 6.5 GB and was downloaded once. So the pictures are made entirely on that computer: no picture is ordered from an online service, and no subscription is paid to produce them.
In practice, each picture starts from a sentence in English describing the scene, for example an owl with stained-glass feathers perched on a branch under the full moon. The software takes about a minute to come up with a picture. A second pass then doubles its size while redrawing the details, which takes another four to six minutes: that is what makes it sharp enough for a wallpaper.
Sorting: every picture looked at, every picture rated
A generated picture is never published without being looked at. Each one gets a rating: 3 when it is truly beautiful, 2 when it is beautiful, 1 when it has a flaw or simply does nothing for us. Pictures rated 1 leave the gallery. Those rated 3 become favourites, flagged with a star and shown first. Today the gallery holds 112 favourites and 35 pictures rated 2. Only pictures rated 2 or 3 go into the HD wallpaper collection.
There is nothing scientific about this sorting: it comes down to our own eye and taste. But it stops the gallery from filling up with everything the machine produces.
What we reject
The software has known weaknesses, and we watch for them. It cannot write: a shop sign or a poster turns into a string of fake letters. So we ask it not to include text, and a retro cinema picture with unreadable signs was removed. It draws hands and close-up crowds badly: we avoid those subjects, or reject the result. It sometimes adds details that have no business being there, such as a national flag on a picture that never asked for one: rejected too.
For our escape games the rule is stricter: an illustration must never contradict a puzzle. For the electrical panel puzzle in Professor Volt's Lab, the software kept drawing four switches when the puzzle has three. A player might have gone looking for the fourth, so we used a picture that gives nothing away instead.
Why use AI at all?
Because we are a very small team with no illustrator. Without these tools, the gallery would not have 147 pictures, and probably no puzzle of the day either. We also know that the question is a sensitive one, especially for artists. That is one more reason to say it clearly, picture by picture, and never to present these pictures as an illustrator's work. If you would rather avoid AI images altogether, the 4 NASA photos in the Space collection are a good place to start.
What we do not do
We do not pass off a generated picture as a photo or a hand-drawn illustration: every picture in the gallery is labelled. The 4 nebula photos come from NASA, with their credit; on the site and in the HD collection, we brightened their colours and added a few points of light. So they are retouched photos, not scientific documents. Nor do we promise exclusive rights to the pictures: our terms of sale say so. Finally, the pictures shown on the site are reduced to 1,200 pixels wide and carry a small watermark; the HD version without a watermark is the one in the paid collection.
These articles themselves are written with the help of AI tools. If a picture or a text seems wrong to you, get in touch: we read every message.