Don’t Spend Too Much On Cover Art, But Beware AI
A comment from reader Grace reinforces the idea that you shouldn’t overspend on cover art:
“As someone who reads many, many books on Kindle Unlimited a week, I can confirm that the cover is only used as a quick way to guess the genre and read the title. The small size of the cover shown in Kindle recommendations means that small details aren’t as important as the broad picture. I can normally guess the genre of a book if I can see the title and cover.
Also, the print-book habit of making the author’s name really big, bigger even than title? I’m pretty sure that’s only good for the really big-name authors.”
That is a very good point. I know of people who have spent megabucks on cover art. There can be legitimate need for that – like you’re doing special-edition hardcovers, or most of your sales are print, and so forth. Scenarios where readers might spend a lot of time actually looking at the cover art. Like, to pick a concrete example, the four books that came out of Brandon Sanderson’s Kickstarter this year will probably have very expensive cover art because so many of the Kickstarter pledges were for the hardcovers.
However, most of us aren’t Brandon Sanderson, and Grace’s comment makes a good point. The majority of your cover’s actual utility will be people looking at it in thumbnail, and you don’t necessarily need to spend a ton of money for it to look good in thumbnail.
It’s a bit how I finally realized that most science fiction readers want covers with a spaceship and a planet on them. It clearly conveys the genre, and the sales of the SILENT ORDER series went up once I switched to covers like that.
However, you don’t want to go too cheaply, since that has its own set of perils. “Free” stock photo sites often don’t have proper attribution for their images.
I would also issue a strong caution against using the new generative AI image tools for anything, especially for something commercial like book covers. For all the talk and hype about the AI “creating” images, it isn’t creating anything. For that matter, calling it “intelligence” is deceptive – it’s just a very complicated mathematical formula and decision-making tree.
It doesn’t create anything new. What it does do, however, is scan thousands of related images, grabs bits and pieces of them, and smushes them together into a composite. That’s basically what I do when I make a book cover in Photoshop, just on a larger scale. The difference, however, is that I pay for all the stock images/3D textures I use for book covers. (The image included with this post was made with a lot of stock photos and 3D textures I’ve collected over the years.) The artificial “intelligence” tools, by and large, do not. There have been documented cases of photographers finding the watermarks they use for images on their sites in the images “created” by the AI tools.
So basically many of the AI tools are industrial-scale plagiarism engines, and there’s going to be a massive lawsuit about this at some point. You don’t want to be the test case!
The most important thing about your cover is that it doesn’t look ugly and it precisely conveys the genre. While trying to do it for free has its dangers, you can achieve what you need in a cover without spending a lot of money.
-JM
JM wrote: “So basically many of the AI tools are industrial-scale plagiarism engines, and there’s going to be a massive lawsuit about this at some point. You don’t want to be the test case!”
Wow! That’s quite an intense and sweeping allegation expressed with overwhelming certitude! Are you sure you understand how most of these tools work? It doesn’t look like it from your description above (I work in machine vision for robotics in agriculture and I’m intimately familiar with how these tools are trained and what their resulting models are and mean).
In any event, whether or not ALL of these tools are “plagiarism engines” (I, of course, disagree), it seems to me that the probability that one of your readers/indie authors who sells a few hundred or perhaps a few thousand copies of a book is going to singled out as a “test case” for a “massive lawsuit” for using a cover which was partially generated by an AI tool is zero (assuming the image user isn’t so stupid as to not check for a watermark). Logistically, how would that even work? If you see a book cover, how would you determine that it was generated by an AI tool (which was then probably run through Photoshop or equivalent)?
I suspect there’s a vast difference between machine learning for industrial/agricultural applications and image generation. There’s probably much less margin for error when programming for a $500k agricultural or industrial device than there is asking an AI image generator to spit out a picture of Santa Claus riding a tyrannosaur at McDonalds or something.
Overall, I’m very dubious about AI image generation, and it seems sketchy.
That said, the technology probably isn’t going away. A less-sketchy use would be a stock photo site that specifically licenses images for use in AI generation. That way, the AI generator could attribute which images it uses, and the original photographers could get a micropayment. Microsoft and Adobe are both rumored to be working on something like that.
However, right now it’s way too early. The enthusiasm for AI image generation reminds me a lot of the enthusiasm for Napster and Limewire and other file sharing services in the 2000s. P2P file sharing technology eventually had a bunch of uses, and most people who used Napster and Limewire didn’t get used, but a whole bunch of people did.
JM wrote: “A less-sketchy use would be a stock photo site that specifically licenses images for use in AI generation. That way, the AI generator could attribute which images it uses…”
That’s not possible. For example, consider DALL-E 2. It’s trained on 400 MILLION images. However, once trained, the images are not used again. There are no images, no image indices, no website information, and no pixels contained in the runtime program and network or within anything it has access to (which is not much).
The internal program consists of a frontend that turns the prompt that the user types in (for example, something like “magical woman casting firespell”) into a mathematical representation (vectors), a random (or specified) seed, and the 12 BILLION parameter network that was created by training an initially random set of parameters on the 400 MILLION images (now long gone) using backpropagation and other techniques. The 12 BILLION parameters are used to perform vector and matrix operations (mostly convolutions) on the input vectors (derived from the prompt and the seed) and the result is an image matrix which usually is at least somewhat related to what’s desired (as specified by the user prompt).
That’s it. The surprising thing about these sorts of “AI” tools is not how complex the software is, but how brutally simple the runtime portions of them are. It’s just a crapload of multiplies, adds, mins, maxes, and other simple arithmetic operations. No clever algorithms, no databases of images, no complicated mathematical formulas, no decision making trees, etc. Just a huge number of simple arithmetic operations.
As an example from my world, we recently started utilizing an AI deep net (YOLOv4) to find vegetables and weeds in real-time while moving through the vegetable beds in a field. This 30 million parameter network (less than 1% the size of DALL-E 2) replaced 50,000 lines of my beautifully crafted (I would say artistic) C code. All that effort replaced by a bunch of multiplies and it works significantly better too. Ah well, if I can adapt to this brave new world, so can artists in the world of AI tools like DALL-E 2.
Back to attribution. Let’s say I asked you to draw something, perhaps a face of a 30 year old unshaven Korean man, without doing a google search of related images, but rather from memory. Assuming you don’t have a good friend who happens to be a 30 year old unshaven korean man, you’d draw someone that didn’t look like anyone in particular. And depending how good you are at drawing, the picture may not look Korean and may not look 30 either.
Now I want a list of all images and imagery (in-person visuals) that formed the basis for your drawing. All the Koreans, all the men, all the unshaven folk, all the 30-year-olds, etc., that you’ve seen sometime in your lifetime that MAY have influenced what you drew. First, it’d be a pretty darn long list and second, you have no ability to produce that list because that information is long gone, never recorded/memorized by you, or both.
I asked DALL-E 2 to draw that very same thing. But DALL-E 2 is in a similar boat. All 400 MILLION images were used to teach it to create images and therefore it would have to list ALL 400 MILLION of them as attribution to each and every image it creates. That simply wouldn’t be useful. In case you’re curious, here’s the 30-year-old unshaven Korean DALL-E 2 created for me: https://drive.google.com/file/d/1pTjsBrZEoR73bUxiQHUZfffDP8yoX05p/view?usp=share_link . Note that this face is completely unique and no actual human looks like that. Yet, I think you’ll agree it’s very realistic. The lighting and shadowing are quite good and DALL-E 2 has learned how to do that from every single one of the 400 MILLION images.
In summary, no, AI image tools cannot provide attribution to a subset of images for a given created image, but if they did provide attribution, it would be a list 400,000,000 images long which wouldn’t be useful.
That’s a very cogent explanation of AI image generation, and I’ll refer people to it who want it explained to them.
That said, I’m still convinced it’s industrial-scale plagiarism. Calling it a “plagiarism engine” might have been slightly too broad a brush, but I think “plagiarism salami slicer” might be a more apt description. If it’s “learning” patterns, then it’s copying those patterns. One could say that’s the same thing humans do, but there’s a wide difference between a human learning to copy a painting and then applying those skills to a new composition
and putting the same painting on a photocopier. It might be five hundred million photocopiers pulling pieces of five hundred million pictures, but it’s still just copying.
But it doesn’t matter what anyone thinks. The AI image genie is thoroughly out of the bottle. Overall, I don’t see any potential benefit to the technology and a whole lot of obvious avenues for harm. But as destructive technologies with no upside go, at least it’s still not as massive of a folly as biological weapons or improving respiratory viruses.
Welp, you were (sorta) right about the massive lawsuit:
https://www.techspot.com/news/97276-artists-launch-copyright-lawsuit-against-ai-art-generators.html
The “sorta” part is they are not going after individual users, just the groups that host the AI art tools so writers who have generated images for book covers are still safe.
My prediction is that if the plaintiffs win, and assuming places like China, Russia, Iran, etc. don’t host and continue to develop these tools, lawyers will make a fabulous amount of money, 7+ billion people on the planet will be poorer (lives less enriched by visual beauty), and in the long run, even the artists themselves will be relatively worse off (because they won’t be able to use the tools either so their productivity will be much lower).
The lawsuit is unfortunate, but perhaps inevitable. But the technology (while I’m still not fond of it) isn’t going away. I suspect we’ll end up seeing an accelerated legal evolution of the sort that went from Napster P2P file sharing to Spotify becoming a publicly traded company.