The Flood of AI Coloring Books
An AI-enthusiast explains how to recognize bad AI illustrations
Publishers of low-content books are discovering AI image generators. Recently, I watched and independent publisher in a vlog as they created a coloring book. My growing sense of dread rose in step with their own enthusiasm. The images were grainy and full of artifacts that didn’t make logical sense. The pictures were unsuitable for coloring.
The creator either couldn’t see how unsuitable the images were, or didn’t care. It broke my heart to imagine someone buying that book.
Other creators have found these AI tools and rejected them in their reviews, and this seemed like it would be the default reaction. But that creator opened my eyes. People need money to live and may be willing to overlook obvious flaws in order to make a paycheck.
In other words, people are greedy.
Print-on-demand publishing lowers the bar for people to create cheap coloring books. There’s nothing wrong with cheap coloring books. But, it can be difficult to tell when ordering a print-on-demand book exactly what you’re getting.
Will AI art flood the market with crappy coloring books? And is there an upside to the technology?
The Market for Cheap Books
I’ve spent a few months working with AI image generators trying to create coloring pages. I’ve documented my progress on my design website, and even published a coloring book on Amazon.
AI enthusiasts are starting to experiment with the technology. It takes work and technique but good illustrations can be made.
An industry already exists to create cheap coloring books. These coloring books are often made using clip art and premade materials that may be used by several different publishers. The people working in this niche focus on quantity over quality, for the most part, uploading hundreds of coloring books in the hopes that a few will become popular.
AI tools are now being integrated into the apps used by these independent publishers. The design sites, Canva and Creative Fabrica, have both launched AI image tools based on the open source Stable Diffusion platform.
For a coloring book enthusiast who regularly colors, or has children who color, this can be a boon. If coloring is a daily activity, one may need to buy a few coloring books each week. When publishers use the same clip art and interiors, this sometimes leads to duplicate images. AI images are much, much more likely to be unique.
Unfortunately, a new problem will arise.
The Pain and Promise of AI Art
Expect a lot of coloring books to be published with awful images. Most of the initial images produced by the current AI models are not suitable for printing. They’re too small. The lines are inconsistent. The composition doesn’t always make sense, and some areas of the image simply cannot be colored.
I’ve spent a few months figuring out how to correct those problems. The tools keep improving over time, but one also has to learn some techniques.
- Create an interesting composition using prompts, settings, new models, and applications.
- Recognize problems and fix the image using a technique called inpainting and image editing software like Krita.
- Upscale images while keeping the image sharp using an img2img function and a vector graphics program like Inkscape.
Using these techniques I can create a printable coloring page illustration in 20 to 30 minutes, although it took me a few months to perfect my technique. For example, here is an image I posted a half hour after noticing the hashtag #Caturday was trending. (Feel free to print it out for coloring.)
A skilled artist would have spent a good part of their day creating this image, and they wouldn’t be able to just throw it away on a social media post. Illustrators need to sell their art for a decent price to make a living wage. Which is why, if you can afford the work of a real artist like Hannah Karlzon or Kanoko Egusa, please buy their beautiful books.
This is the best way to find good quality coloring books.
But there will be a wave of good, cheap books made with AI art that can feed the growing demand for coloring books. These books will be able to cover particular niches and themes that might otherwise be ignored as unprofitable.
Unfortunately, this wave of new books created with AI art will also include books that just aren’t made with much care or technique.
Signs of Unsuitable AI Illustrations
Some people buy a lot of coloring books. Because of this they need cheaper books and are okay with slightly lower quality if the price is right. This demand needs a supply, and there is now a secondary market that supports these suppliers.
It is possible to buy whole interiors for coloring books that can be matched with a cover and resold. Let that sink in. Whole interiors of coloring books can be purchased for a few dollars.
This is why one can find the same images in different books by different publishers.
Most of the producers who create these low-cost books try to make the spreads different with some light editing. But they are accustomed to using vector images (SVG files) that scale easily or high-quality bitmap images. The low-quality images that AI-art generators produce need special handling.
Here are some red flags to watch for that might indicate the producer is using AI art without putting much work into fixing common problems.
- Square books. The AI models are trained on square images and will, by default, generate square images of 512x512 pixels.
- Books without preview images. If you can’t see a good quality preview of the interior, this was a conscious choice that the producer made. They don’t want potential buyers to see the interior.
- Multiple faces, arms, and legs. Because the models are trained on small squares, any attempt to use a larger frame might generate strangely mutated compositions.
- Blurry images. Because the images are being upscaled, they may be pixelated. There are ways to upscale an image with AI, but these tools aren’t usually included in the basic software.
- Large dark patches. When creating black and white images it can be hard to specify a minimalist line art style. These patches can be fine occasionally— especially for adult coloring books — but most professionally-made illustrations avoid solid black patches. (See the second image below.)
- Incomplete, cut-off forms. Notice the ears on the cat below are cut-off. Again, the model is trained on square images, which means many of the training images were cropped.
- Mottled, garbled lines. Notice the lines on the first cat’s dress below. This “crackled paint” texture is common on AI images that haven’t gone through enough iterations to fill out details.
- Watermarks. Some of the images used to train the models included watermarks, so in some cases the AI has learned to add these watermarks to all new images that it creates.
Here are two of the better examples of images picked from 20 that I created. They might look fine offhand, but still need improvement to become a good illustration.
Overall, I would warn coloring book enthusiasts to look carefully at sample images. Many of the disappointing images I have seen look fine at first glance. On closer inspection the details don’t hold up: A butterfly’s wing is suddenly misshapen at one end, or a pattern turns into a gray, garbled mess.
But don’t fear the coming wave of AI coloring books. AI has the potential to create a never-ending supply of unique and interesting images. Today’s AI models can create the base images for an entire book of cats dressed in Tudor-period dress with only about 10 minutes of processing time.
Getting those images into a printable form, however, takes more effort to prepare. Some publishers will try to skip that extra work. So, be wary. Inspect books closely for signs of bleh AI art. And take heart, the models are improving day by day.
That being said, a good quality book by a professional illustrator will always be worth the extra cost.
