Using DALL-E to generate cover art

I tried that a while ago with a certain Infocom game, but all I got was nightmare fuel:

And here I thought it couldn’t get worse than when Telltale’s Sam & Max started glitching during the end credits when running it in Wine under Linux.

sam-and-max-nightmare-fuel

5 Likes

Terrifying! Bottom-middle is so dramatic!

Craiyon did a pretty good job with people for my latest work. :joy:

Vesper and Harris talk about Cost of Living

I just found Midjourney, which yields some very nice pictures too. This is for my WIP detective game.

3 Likes

The Dream by Wombo app can also produce some nice results. Here’s an alternative take on my fishing village game setting.

4 Likes

I hadn’t known about Midjourney. Very nice, better results than craiyon, I think.

Please keep in mind that those pictures are not REALLY computer generated… DALL-E is performing a google search in the background and distorts the results.

So you might still run into legal issues.

2 Likes

I recently tried this with classic Infocom titles. Here are the results:





I like the simple “Lurking Horror” best:

1 Like

This is not exactly how it works. Here is a more detailed technical data for anyone interested on it:

DALL-E is a 12 billion parameters AI used to search in google images what you type. Then these images are 3D printed with clay in the Open AI Clay Factory located in Oklahoma to generate a physical tridimensional mockup of the image.

After that the work is finally processed by a very specific machine to obtain the final results.

The company has been very careful to keep the details of this machine secret, but reverse engineering has revealed that the main module is made up of a monkey with a hammer and some crayons.

9 Likes

I’m no lawyer, so I can’t comment on the legal issues other than saying I think they’ll turn out to be complex. But i can say with a fair amount of confidence that your description of DALL-E’s algorithm is terribly misleading. It does not distort images, it builds images up from data it previously collected from images and their captions. It’s more akin to you painting a picture of an elephant based on all the pictures of elephants you’ve ever seen. It bears only a loose relationship to the individual pictures. Midjourney shows the successive approximations it uses — it’s very illuminating as to how it doesn’t merely distort existing images.

4 Likes

Photorealistic Dr. Sourpuss. Not bad. (From “orange cat with round glasses”)

image

3 Likes

Very cool results! AI image generation has improved leaps and bounds since I generated the images/cover of Universal Hologram (for which I used VQGAN+CLIP and bigGANxCLIP).

2 Likes

It is interesting to consider that if technology such as this becomes sufficiently reliable and efficient, it might be possible for a future Z-machine interpreter to be created, that draws images based on the description read out by the game on each turn!

1 Like

It missed both the boarded front door and the mailbox :frowning:

7 Likes

I love that you tried this out :smiley:

It looks like there’s a technical issue with using DALL-E in its current form to “illustrate” IF games. I’ve noticed from their research blog, that often the AI requires descriptions that emphasise certain parts of the description to get them represented satisfactorily in the output. This precludes the use of it to illustrate un-tested/unedited IF text…

I have this same problem with Midjourney (I don’t have access to DALL-E). I was trying to render Pikachu and Lenin together (for my ahistorical Twitter account https://twitter.com/AnachronisticI6), and depending on which way I worded it, sometimes I got them next to each other, sometimes it was Lenin with Pikachu’s ears, and sometimes it was Pikachu with Lenin’s beard.

1 Like

On the legal question, I just ran into this article: U.S. Copyright Office Rules A.I. Art Can't Be Copyrighted | Smart News| Smithsonian Magazine

…But if you start with non-copyrighted art and then modify it significantly, the result is your work and owned by you in the usual way. That’s legal.

5 Likes

So the question becomes: can you guarantee that a particular image base (such as that used by DALL-E) is free of copyrighted works? I noticed that Midjourney has a mechanism for copyright takedowns.

Yeah, I have no sense of what level of diligence is legally required there.

I don’t believe so, and this is a major point of contention for digital artists… artists with unique styles are having their names used as prompts to generate similar images (based on their actual work), and some of them are pretty upset about being included in the image databases.

2 Likes