Six Kinds of AI Image Tools, Reviewed by a Working Illustrator

I've been illustrating children's books for nine years. For the past two, I've tested every new category of AI image tool with the same question in mind: will this help me finish a book, or just help me procrastinate in a more expensive way?

To keep the comparison fair, I used one fictional commission throughout: a double-page spread for a picture book about a fox who keeps a lighthouse. It's a night scene in a storm, with warm window light and the fox on a ladder holding an oil can. I could do that brief by hand in my sleep, which made it easy to judge what each tool added or took away.

I'm reviewing kinds of tools, not brands. Products change every month, but the categories have been remarkably stable.

The scorecard

The six kinds, up close

Text-to-image generators

You type a description and get images back. For the lighthouse fox I got forty images in ten minutes, and about six were genuinely evocative, including a sickly green storm sky I'd never have chosen and a lighthouse leaning at an angle that made the whole scene feel precarious.

What I couldn't get was my fox. Consistency between images was poor, paws went strange, and every attempt to adjust one detail changed three others. I use these tools the way I use a stack of old magazines: to jolt my thinking, never as the work itself.

Sketch-to-image

Here you upload a drawing and the tool renders it in colour or in a chosen style while keeping your composition. This was the first category that felt built for people like me. My pencil rough of the fox on the ladder came back as a colour study that respected my layout, my perspective and, more or less, my character.

I wouldn't print the result. But it gave me four lighting options in the time it takes me to mix one palette, and I painted the final spread with one of them pinned beside my desk.

Generative fill in image editors

This is the least glamorous tool and the one I use most. I select an area of my own painting and ask the editor to extend the sky, remove a stray smudge or widen the canvas for a bleed I'd forgotten about. It saves me hours on every book.

Because it works inside my own image, the results usually match my brushwork closely, and when they don't, I paint over them. This is a tool, not a collaborator, and that's the highest compliment I can give it.

Upscalers and restorers

Upscalers increase resolution and restorers clean up scans. I tested them on an old ink drawing scanned at low resolution. The upscaled version was crisp enough to print, with only slight over-sharpening on the fine hatching. Publishers' print specs are unforgiving, and these tools have rescued at least two of my deadlines.

Vector and pattern generators

I wanted a repeating wave pattern for the book's endpapers. The vector tools produced clean shapes but struggled with seamless repeats, and the files were messy, full of hundreds of unnecessary points I had to clean up by hand. For now I'm faster drawing patterns myself.

Personal style models

Some tools let you train a small model on your own artwork so it generates in your style. Ethically, this is the cleanest option I tested, because the only work it learns from is mine. In practice, it needed around thirty of my finished pieces, careful captioning and an afternoon of fiddling, and the results captured my palette better than my line.

It's useful for variations, like the same fox in six poses for a character sheet. It's also unsettling. Seeing an imitation of your own hand is a strange mirror.

My actual stack

After all this testing, my real workflow is surprisingly small:

1. Sketch by hand. Always. This is where the book lives.
2. Use sketch-to-image for lighting studies on complex spreads.
3. Paint the final by hand, digitally or traditionally.
4. Use generative fill for extensions and fixes.
5. Run the upscaler for print prep.

Text-to-image generators stay in a separate folder labelled "weather," for days when I need a jolt. Everything I deliver includes a one-line note about which of these tools I used.

The tools that earned a place are the ones that respect my drawing, and that turns out to be a very good test for any of them.