It Took Me 125 Hours to Draw the First 8 Panes of My Webtoons Using an AI Workflow… About 3x Slower Than a Normal Artist…
A month or two ago I gave myself a mission to create a non-fiction Webtoon called Thoughts by Aaron, based on my life. The catch is that I’m not particularly an artist — I haven’t drawn in a long time, and I’m not nearly fast enough to produce regular episodes. This has been a roadblock for me in the past — in university I was the writer of a video-game-related comic inspired by Penny Arcade and drawn by my friend Noah Kroese. I’m very proud of that time, and have fond memories of it, but our styles never really meshed perfectly and we shut it down after a year or so.
I didn’t try something on my own afterward because I couldn’t draw fast enough.
In the past I’d give up on the idea of a webtoon every time the reality of my slow and untended artistic ability crashed home for me. This time, though, I was curious about the possibility of training AI on my own art, my own characters, and then using an AI workflow to speed myself up and maybe make the medium a viable place to tell stories for me.
As part of this I’ve been meticulously tracking the time I’ve put into each step, often recording my screen and coding it for how I’m spending my time. It looks like this:
These are not casual hours. I’m pretty meticulous about clocking in and out on tasks, so these are 125 hours of focused, “in the zone” work. No bathroom breaks counted. I read once that the average person gets about three hours of actual work done during an 8 hour day, and if that’s true the above represents about 41 working days.
Let’s go through the steps one-by-one:
Learning So I Could Teach:
Research and Prep (16 hours & 16 minutes):
This was a learning phase. Not just the AI systems, I spent a lot of time researching comics themselves. I read books on the conventions of comics, my favorite being Scott McCloud’s very innovative Understanding Comics. I began reading some really good Webtoons, my favorite of those being Spontaneous World Shifting by Sushiru-River, and read them until caught up.
Toilet time, which used to be dominated by doom scrolling on Instagram, was consumed by Webtoons, instead.
I also researched and experimented with different generative systems, installing and testing software, that sort of thing. This is on top of hundreds of hours experimenting with tools like Midjourney from before, which I didn’t count because I have no idea how to quantify that time. I ended up using Midjourney’s Character Reference and Style Reference ability as my place to start, simply because it’s the one I was most familiar with and ready to move on first. I’ll have to explore others over time.
Relearning to Draw (34 hours & 10 minutes):
The second step was buying an Apple Pencil, designing a character, and relearning to draw enough to create some training references. I decided to base the character on a 14-year-old version of myself… or at least somewhat. A more cool version of myself, because I was not cool at 14 years old. Over the 34 hours, I used a combination of vector and raster art to make three drawings of “Aaron” using a tool I love and hate called Affinity Designer 2 on my iPad.
Here are the three original drawings:
They were not fast, but I’m fairly happy with the character. Especially the one sitting down — that dude is adorable to me. My plan is to continue drawing out the character over time to enrich the data set, and to refine my ability for future characters, as well.
Training the Model (16 hours & 25 minutes):
When I say “train the model” I’m actually not really training anything in the traditional AI sense. Instead, I experimented with using Midjourney’s style and character reference settings with my character as the reference. At some point I’ll revisit some specific Stable Diffusion tools for this, but for the time being I figured my first goal was to see if I could generate a character closely enough that I could edit it into production.
Initially, the generations were pretty hit or miss, but had potential.
He was moody and misshapen, but at least it was clear that the images were influenced by my references. Eventually I got to a point where I could generate a character doing what I asked, and some percentage of the outputs looked like they would be close enough that I could edit them into what I needed.
It was time to begin attempting my first few actual panes, with the goal of creating images that aligned with my storyboard.
The First Attempt:
Image Generation (4 hours & 41 minutes):
When I actually draw the first official episode of Thoughts by Aaron, I’m planning on using a very human story about losing my grandfather when I was younger. One of the dangers of AI-enabled work is that it doesn’t reflect the human experience, and so it’s important to me that I start on something that is clearly an expression of something important to me. But I wanted to learn on an initial episode that had less emotional nuance — so I decided to start on Episode 0.5, instead of Episode 1, which is an episode about a company I work with called VinAI. It’s a non-fiction topic that lets me learn, and also experiment some with the idea of “comic as education” which I also find interesting.
I wanted to start on an image of young Aaron introducing himself in his home office. My vision of the comic is that — while Aaron is represented as a kid — he does so from a place later in life, like a young representation of who I am today. I started by asking Midjourney to generate images of Aaron sitting behind a desk, talking to the viewer.
Eventually I landed on an image I really liked — friendly, cute, not born from the unholy depth of my nightmares, and relatively accurate to the initial training…
The only problem is that the Aaron in my chosen image apparently lived in a blurred interdimensional space of human-desk hybrids. But, I started with this…
Editing & Drawing (Recorded 21 hours & 16 minutes):
What follows is an example of what consumed a huge portion of my life over the last few weeks.
I count editing and drawing as one category because often times one became the other and were hard to distinguish. Redoing hands, adding the table, adding personal elements that are unique to me, all that fell into this hybrid category of editing what was there and drawing what was not.
There were a lot of things about this image that I wanted to change. I didn’t like the desk he was at, even ignoring the fact it was spearing through his chest. I wanted the initial pane to be an establishing shot, so needed it zoomed out. His clothes are fairly wrong — he’s wearing something around his neck like he just got out of an E3 convention, and his eyes are definitely not blue. Mine are.
I created a time lapse of the 32-hours it took me to edit the very first pane, which you can watch if you want. If you do, please pay attention to the lyrics — I had a lot of fun making the song for what should have been a 5-minute side clip. haha.
Ultimately, I generated 268 images searching for the exact elements that I intended, in addition to editing them together. Here’s the final image:
At first glance they seem pretty similar, because the focal character hasn’t changed expressions. Watching the time lapse, it’s a bit hard to figure out what took 32 freakin’ hours, and maybe this should be a lesson for me to be less focused on details. But for better or worse, you’ll see that literally everything changed from the initial. None of the original background and setting remained beyond his chair legs. And the hands have more fingers, obviously.
I also added in a few things as little nods to my real-world environment. I actually do have a giant commissioned painting of Patapon on the wall of my house… two of them, actually. I tried to pull examples from around my office. A poster of Sword Art Online. A plastic toy Minecraft sword. A Lego set of both Star Wars and the house from Disney’s UP.
And a cat. Because. Cat.
These are actually fairly high resolution edits, but I blurred the background to keep focus on Aaron, so they’re a little blurry in the clips above, as well.
This first pane was by far the most time consuming of the ones I’ve worked on. Partly, that’s because I was learning as I went, and partly because I learned to pick my battles a bit more carefully in terms of how detailed to be on future panes. While the video covers 32 hours of time lapse for just one pane, the remaining 7 panes didn’t take nearly as long.
My goal is to start… and then become more efficient over time.
Dialog and Speech Bubbles (4 hours & 44 minutes):
I wrote the script before starting the comic, and so I don’t mean writing the dialog here. What I mean is that I had to figure out how to make speech bubbles from scratch, place them, and arrange the layout, etc. This portion will definitely get faster, and anyone that knows Photoshop and watches my video will tell you that I don’t know how to do things the most efficient way. I’ll get there, hopefully.
So, with all of that, I figure I should end by sharing the first 8 or 9 panes as they exist today, as a hint of what I hope to be able to do in the future. I’d link to the Webtoon page, but I’m not posting there yet until I’m actually ready with Episode 1, so you just sort of get it here.
Hope you enjoy…
1St 8 Panes of Thoughts by Aaron, Episode 0.5:
