Three AI graphic novels, one research question: where does the line between slop and quality actually live? What changed across 255 agents, 16M tokens, and three complete pipelines.
Effort matters, but it is picky about where it lands.
A kid can spend two hours rewriting prompts and come out with no more taste than when he started. That is real effort in the wrong place. What built your critic was not the hours, it was reading your own comic and admitting it was boring. Judgment against your own output, with nothing riding on it but honesty.
That move is the one nobody assigns. We assign the work and grade the product. The part that made you able to say round two was lifeless never appears on a rubric.
The cleanest result here is the one that costs nothing to reproduce: rounds one and two ran nearly the same token budget and landed a league apart. Scale was never the knob. Judgment was.
I teach high school chemistry and physics, and your three levers are the exact three my students do not have yet. Notice what round three actually required of you. You had to read your own comic and know, honestly, that it was boring. You had to write a brief that specified a credible antagonist. You had to look at a page and say redraw. Every one of those moves is taste, and taste is what a fifteen year old has not built.
So a student hands the machine a one-line prompt, gets round one, and cannot tell. That is the whole problem in my classroom, and you just ran the controlled experiment for it. You had a critic inside you that told you round two was lifeless. Build that critic in the kid first and the machine amplifies a mind. Skip it and the machine amplifies an absence.
Taste is the scarce part is right, and it is also a curriculum.
100% that. The more “human-input”, direction and creativity you give, the better results are. But TBH - it’s just the universal rule of life. The more effort you do, the better!
Fascinating experiment Pawel. I read all three (or attempted to)
1. Afterlight - Unreadable. Crude beyond primative graphics really distracted. No real redeeming qualities for me; I stopped trying after 8-page turns
2. The Warm Hill - Better characterization. I liked the story, but it was pretty simple graphically and got confusing towards the end turns; the image consistency between characters was pretty poor. I could see the decline in the son, but it wasn't consistent. Speech bubbles jumped around, so it was not clear whose dialog was whose.
3. Small Hours - Visually the most interesting for a graphic novel. You can see those Gemini tokens doing their work. I really liked the sci-fi nature of this story, but found it getting a bit jumbled in its logic the more pages we turned. But this was the one I could genuinely see being adapted to a TV series. There were still some inconsistencies with the characters. The Dock worker morphed several times, yet remained remarkably consistent. His daughter seemed to change into a totally different character between scenes.
All in all, great to see this work.
I wonder if you do most of your work in Polish in the Claude interface, as the syntax and phrasing don't seem constructed by a native English speaker.
Thanks for comment and insights. I have similar feelings. The last one is “best”, but still would need some hours to make it at least “good”. More creative people than me, would probbaly use AI to create something really good!
Effort matters, but it is picky about where it lands.
A kid can spend two hours rewriting prompts and come out with no more taste than when he started. That is real effort in the wrong place. What built your critic was not the hours, it was reading your own comic and admitting it was boring. Judgment against your own output, with nothing riding on it but honesty.
That move is the one nobody assigns. We assign the work and grade the product. The part that made you able to say round two was lifeless never appears on a rubric.
Don't let AI write the dialog.
The cleanest result here is the one that costs nothing to reproduce: rounds one and two ran nearly the same token budget and landed a league apart. Scale was never the knob. Judgment was.
I teach high school chemistry and physics, and your three levers are the exact three my students do not have yet. Notice what round three actually required of you. You had to read your own comic and know, honestly, that it was boring. You had to write a brief that specified a credible antagonist. You had to look at a page and say redraw. Every one of those moves is taste, and taste is what a fifteen year old has not built.
So a student hands the machine a one-line prompt, gets round one, and cannot tell. That is the whole problem in my classroom, and you just ran the controlled experiment for it. You had a critic inside you that told you round two was lifeless. Build that critic in the kid first and the machine amplifies a mind. Skip it and the machine amplifies an absence.
Taste is the scarce part is right, and it is also a curriculum.
100% that. The more “human-input”, direction and creativity you give, the better results are. But TBH - it’s just the universal rule of life. The more effort you do, the better!
Fascinating experiment Pawel. I read all three (or attempted to)
1. Afterlight - Unreadable. Crude beyond primative graphics really distracted. No real redeeming qualities for me; I stopped trying after 8-page turns
2. The Warm Hill - Better characterization. I liked the story, but it was pretty simple graphically and got confusing towards the end turns; the image consistency between characters was pretty poor. I could see the decline in the son, but it wasn't consistent. Speech bubbles jumped around, so it was not clear whose dialog was whose.
3. Small Hours - Visually the most interesting for a graphic novel. You can see those Gemini tokens doing their work. I really liked the sci-fi nature of this story, but found it getting a bit jumbled in its logic the more pages we turned. But this was the one I could genuinely see being adapted to a TV series. There were still some inconsistencies with the characters. The Dock worker morphed several times, yet remained remarkably consistent. His daughter seemed to change into a totally different character between scenes.
All in all, great to see this work.
I wonder if you do most of your work in Polish in the Claude interface, as the syntax and phrasing don't seem constructed by a native English speaker.
Great work
SM
Thanks for comment and insights. I have similar feelings. The last one is “best”, but still would need some hours to make it at least “good”. More creative people than me, would probbaly use AI to create something really good!