ASIDE: Fun with ChatGPT Image Making!
Executive Summary: I play around with ChatGPT’s image making fuction.
What’s this? Did Author-kun hire Io-sensei to draw for him again? No! Io-sensei is still busy with her gacha work! This is ChatGPT da!!!
A NEW ERA IN IMAGE MAKING
It really all began late last month when a whole bunch of Ghibli images were flooding the internet! AI made Ghibli images are not new, LORAs for that have been around for more than two years!
What makes this special is that for the first time, a powerful LLM is paired with a powerful image making model! Just two years ago you had to do all sorts of wizardry to get PonyXL, or whatever you favorite model is, to do what it wants, now, you just have to ‘'tell’' ChatGPT what you want, and it can do it for you!
This really takes a lot of the learning prompts out of the picture, since you can even ask ChatGPT for help if you want to learn that still!
HOW THIS PICTURE WAS MADE
One of the core limitations with making images for /Conspiracy/ Girls was that I have a vision, but the vision was made in COM3D2! There was no simple way to do images quickly, even with IMG2IMG because 1) the quality wasn’t up to par 2) the process of taking shots, processing them in Stable Diffusion, then editing took a long time, time that could be better spent writing!
ChatGPT solves all these problems. For starters, you have someone who understands your vision. They understand your vision perfectly even! Io-sensei is good, ChatGPT is crazy! ChatGPT understands memes like no other, and all I have to do is to give the screenshot in COM3D2, tell ChatGPT the basics of it and you’re done!
Its just like working with an actual artist…! Except this artist is very, very smart, very very patient, and works very very fast!
For reference, this was the original image that I made for Madison, she’s doing a WRRRRY pose, something I didn’t tell ChatGPT! Sorry! As you can see, it captured Madison’s core traits perfectly, it really looks like a PonyXL generation! Almost all the minor details were captured except that the hair falls over Madison’s eye a little bit.
TECHNICAL: AI can’t do half-covered eyes properly because it is trained to do perfect eyes. So Madi-san’s ‘look’ where her bangs fall half over her left eye and the tiny little ahoge on top aren’t quite captured properly. Perhaps with more testing and training it will work!
Most importantly… this was done with a reasoning model. A reasoning model will get better over time as you talk more to it and iron out the kinks… This is scary powerful technology!
I tried to get ChatGPT to do a ROAD ROLLA DA image, but it failed, mostly because the original ROAD ROLLA DA image has Dio in a really small shot, with the road roller taking up most of the image! I doubt ChatGPT is so advanced it can surpass the limitations of IMG2IMG itself…
However! It WAS able to do another Dio image…
I felt like this one was easier, notice that Madison in the first image was doing the ‘'hand outstretched WRRRRRRYYYY’' pose.
ChatGPT did not make this specific image for me, but…
At first, the image that came out was pretty strange, I didn’t know why ChatGPT gave me a totally different image, and I thought for a little.
Did you see it?
THIS IS ACTUALLY JOJO’'S RESPONSE TO DIO BRANDO’S TAUNT! The angle is reversed and Jojo is coming from the other side! This image was actually adapted from ‘'I can’t beat the shit out of you if I don't get closer’'! (I think!) ChatGPT WAS RESPONDING TO THE IMAGE WITH A MEME OF HER OWN!
It really blew my mind.
Working with ChatGPT really felt like working with an artist... but this artist is super smart, understands every meme you explain to them, and is open to correction!
So on this April 1st, 2025, I swear, the very first 200 dollars I get will go towards helping pay for Miss Delulz’s wages! I really, really feel it this time! There’s a hope for Hell’s Theatre to have images done, and done well! The future is bright! I really like how AI is progressing!
R.I, デラ・ルーの大導劇神
HouseDelaroux.com
250322
