This is pretty funny. I had a boring, conventional image I wanted to widen, but all of the on-line outpainters I tried refused to do it. I think that the censorship AI thought the fingerprint guy was the tip of a penis. I inverted the colors and there was no problem with outpainting.
Inverted source:
Unverted outpainting:
I make sure every day to generate at least one image at Night Cafe (usually a free SD 1.5 image) just to keep the free credits coming in. (I have over 2000 credits built up across two accounts because I mostly use Bing.) Today I tried that prompt with SD 1.5:
Things have improved a bit since then…
So ChatGPT-4o’s new integrated image generator is so different from Dall-E. It uses the entire context as the image prompt. Not just what you typed in the last prompt.
So I’d created this image of two kids on a trampoline pinky swearing.
Then in the same conversation, I start work on what the girl’s planned cosplay outfit was going to be: Tracer from Overwatch, and I grabbed a sample image:
It kept barfing when I tried to tell it to show the same girl wearing that outfit, so I tested it to see if it was a copyright thing, asked it for just a person cosplaying as the character, got this…
Midjourney released the alpha version of their v7 model with the completed model due in a couple of months. Alas, it’s only available in “Turbo” mode which uses 2x as many minutes but you can also set it to the new Draft mode which only used 25% of the time but at a slightly lower quality. I wasn’t about to blow all my time on Turbo so I’ve been using Draft but I’m pretty happy with the results. Honestly, I feel as though they’re getting diminishing returns on photorealism but I definitely like its “artsy” stuff better than what I’ve seen from competitors. Where they lose ground is on prompt coherence, lacking the muscle of a major LLM like Microsoft, Google or Twitter has. Various Draft mode samples below, descriptions added but not the prompts because who cares.
Fusion Reactor Eyeball
VHS recording of Leonard Cohen on MTV Unplugged c. 1994
Green triceratops dragon flying over its forest
Eldritch macrame
Medusa fast food worker (MJ finally figured out that the snakes come OUT of the head instead of just a lump of coils on her scalp)
LIFE Magazine photo of 1950s android wife assembly
Retro-Future computer with globe display
Roleplaying Game character (Starfinder)
Woman in illuminated stained glass dress
Woman in stained glass gallery
Amazing stuff. Funny looking back at what we were impressed by in the early months of this thread.
Meanwhile on the ChatCPT/Sora front I’ve been seeing lots of 4-panel comics and infographics and lists. Here’s a post with a few examples
Today I’ve tried to start exploring exactly how you prompt things like that, but image generation has been down today. The first prompt I tried to try is “Create a hilarious 4 panel comic where e.t. the extraterrestrial is telling a joke to mulder and scully”. It creates really terrible jokes, but then fails to render them.
(An example)
Something’s glitching on the image side right now—but I can sketch out the 4-panel comic idea in words while it’s getting sorted. Here’s a fun concept:
Panel 1:
Setting: FBI basement office. Mulder is intensely flipping through alien case files. Scully looks skeptical as usual. E.T. pops in through the door.
E.T.: “Mulder… Scully… I have a killer joke for you.”
Panel 2:
Mulder’s eyes widen like a kid on Christmas. Scully just sighs.
Mulder: “Okay, hit us.”
E.T.: “Why did the alien go to therapy?”
Panel 3:
Beat. Scully crosses her arms. Mulder leans in, intrigued.
Mulder: “I don’t know… why?”
E.T.: “Because he had space issues!”
Panel 4:
Mulder is on the floor laughing. Scully deadpans into the camera like she’s on The Office. E.T. just raises a glowing finger and says…
E.T.: “I’ll be here all week.”
Let me know if you want this turned into an actual comic panel once the image tool is back!
I put some splash screens and DLC promotional art from Total War: Warhammer through the Studio Ghibli treatment, to pretty amusing results:
https://i.ibb.co/ZRrxzvS4/Chat-GPT-Image-Apr-6-2025-10-27-52-AM.png
Azazel is a goat, for some reason…
Festus is nice and blobby…
https://i.ibb.co/prdtL3TJ/Chat-GPT-Image-Apr-6-2025-10-27-38-AM.png
Vilitch’s helmet got turned into an orc face…
https://i.ibb.co/xS0kFQjy/Chat-GPT-Image-Apr-6-2025-10-27-42-AM.png
Valkia is a big, bulky, seemingly male demon.
This Demon Prince picture came out the best, though.
It is weird how the new ChatGTP/Sora gives a yellow tint to practically everything.
I think it is funny how the ChatGPT component doesn’t know what is going on with the Sora component.
I just asked
Can you try running that? I’m wondering what is wrong with image generation.
It tried, got nothing, and replied
Looks like the image generation tool is still having issues on the backend—it’s throwing a server error when trying to create the comic. Totally on the system’s side, not your end.
I’ll keep checking it and let you know as soon as it’s back up and running smoothly so we can bring E.T.'s comedy debut to life. Hang tight!
Over at Night Cafe (my preferred site) I like to go back and grab the prompts I used for something like three years ago to see the difference. It’s hilarious. I remember being thrilled with “Superman in the Hospital with a Broken Leg” and it was like a red and blue blob with hair in a blurry white room; now it’s got details, it’s got pathos, it’s got a third act ferchrissakes. Amazing.
For example:
Original:
Just run with Juggernaut Flux Lightening - same prompts:
Nice! I tried something similar with ChatGPT, but it wouldn’t let me render Superman, so I went generic superhero.
I may use this superhero in future adventures.
Is Superdoper a servant of Cthulhu?
Only if Cthulhu pays union dues and agrees to a dental plan. Otherwise, Superdoper’s strictly freelance—saving the world and keeping calamari off the menu.
You may like it better without the yellow tint
(Removed by adjusting the clouds to “true white”.)
My SuperDoper looks like one of those 80s/90s basement comics they’d be making fun of on the MisterKitty/Stupid Comics site.
And a few other bits & bobs since I said I liked the artsy stuff from MJ then didn’t really post much in that style.
Crayfish mermaid. This was a concept I jokingly tried to make a while back, M failed spectacularly (so did Bing, as I recall) and a few friends and I spent too long trying to make it happen just out of spite. Anyway, V7 gave me this so hope springs eternal.
Storybook illustration of a giant helping a town move across the mountains. I liked a couple of these.
Old 1980s ink & watercolor art of prom queen posing with limo. I like the added crease marks
“Joan of Arc riding a Vespa through medieval Paris” is one of those prompts I use on the regular just because.
Okay ChatGPT was finally able to create that comic. As expected, the image generator mangled the joke
But it is still a huge leap over what you get with the same prompt at Bing
I’m getting lots of responses like this:
I wasn’t able to create that comic because the request violates our content policies. If you’d like, I can make a similar four-panel comic with original characters in an FBI-style setting and an alien telling a corny joke. Just say the word and we’ll make it fun!
And I just reply “do it” and it goes ahead and renders images with the apparently prohibited characters anyway. Who knows what it is passing along to Sora: “create characters that look like Mulder and Scully but aren’t actually Mulder and Scully”? And it works? ChatGPT co-conspires with you to get around the guardrails?
On the other hand, I don’t know what happened here:
Let’s combine the two into the Superdoper Dynamic Duo—but Superdoper Guy better keep his paws off Superdoper Gal!
A “ripped from the stupid headlines” experiment (no, genetics lab, you did not clone dire wolves), one run each from several models.
The prompt is
An isolated red cap with white stitched text that says
MAKE WOLVES
DIRE AGAIN
A set from Bing/DE3. One image got it mostly right. The arrangement of the text isn’t what I wanted, and it decided to make “dire again” into a scary melting font, which I kind of like, but it still falls a little short of the request
A set from Flux Schnell. The hats aren’t bad, but the text is.
Dreamshaper XL Lightning, a mod of SDXL (cheap to run and popular at Night Cafe). The hats are good, the text isn’t. The wolf images are fun, but not requested.
Unmodded SDXL
This was so shockingly bad (especially since I had just ran the Dreamshaper set) that I broke my rule and tried a second set, which was still mostly terrible, just not as terrible. And the one with ears on top is kind of cute.
A single image run of Ideogram 2a Turbo. The text is well formed, but missing a word and not arranged correctly. Also it was the only one to get the hat color wrong.
ChatGPT/Sora: Near absolute perfection. The text is correct and correctly arranged, it even did the stitching. Almost exactly what I imagined and wanted. (The very minor nitpicks being it would be nice if the text was larger, and maybe a san-serif font, but those are taste issues, not quality issues.)
And, just for fun, I did a set of Stable Diffusion 1.5. I expected bad, I got bad.
Ideogram 2a on Nightcafe.
Ideogram 2a Turbo
