BTW, both of the original sites I once used for outpainting square images are no longer usable, but some googling found plenty of other free options. Here are three. Of those, Pixelcut gives you the highest resolution.
A lot of these are nightmare fuel. Like if I was really high and found this thread, it would have been bad.
I also thought this was a shill for that website, glad to see that people are having fun with it.
If you’re starting from the top, these attempts were from 2022. A lot has changed since then. If you’re saying in general – well, kinda sorta. I think most images posted were chosen because they were weird or funny-tragic versus “Here’s a nice mountain I prompted”.
This made me go way up to the top to look at what we thought was awesome in 2022. That’s definitely some… stuff. For funsies, I grabbed a few old Night Cafe prompts from that era and re-ran them in Midjourney straight up with no additional prompting aside from the same general image ratio in a few. Posted the entire MJ preview rather than pick one to upscale.
- A Sordid Affair
This was (and maybe still is) my most upvoted NC image at the time. I still kinda like it! The MJ stuff is all very romance novel cover and the NC result was weird and abstract but still conveys illicit sensuality.
- Post-apocalyptic city built around the legs of a rustic giant
The obvious different here is the point of view. It would be trivial to make MJ create an aerial image of the scene with some additional prompting but it didn’t. The NC image is far less refined but makes for a more interesting scene (IMO) of a settlement arising from the ruins and materials of this giant.
- Lizard Cowboys in the Style of Frederic Remington
Again, the obvious difference is that one has lizards dressed as cowboys and the other mainly has cowboys riding lizards. And, again, it would be trivial to ask for lizards in cowboy hats instead if that was my desire. MJ did a far better job of making a faux-Remington though NC gave it a solid 2022 try.
- The Shug Monkey of Cambridgeshire
The Shug Monkey is a B-tier cryptid from England, supposedly a dog-headed ape first “sighted” in the WWII era. I didn’t seriously expect either to know what it was though NC made a weird monster picture and MJ just gave me monkeys.
- A Bucket of Fire Except The Fire Is Green And Also There’s An Egg Whisk In It
This was just a coherence test and 2024 MJ wins easily over 2022 NC. Shocker.
- Flirty Robot
I was delighted by this NC guy when I first saw it and I still think it has its charms though its deficiencies are also on display. None of the MJ ones match it in storybook style though 1 & 2 are perfectly nice flirty robots.
Again, just for my amusement and no real conclusions to be had. Obviously the newer model can do things more realistically and coherently. The NC images I posted were picked because they were “good” and didn’t represent my average output (and leave out big weaknesses like rendering people) while the MJ images were single runs. And, of course, 2025 NC is running off various Stable Diffusion models that are well updated. Just a stroll down memory lane.
I tried The Shug Monkey of Cambridgeshire in Bing and got dreadfully cheery things.
I added one of my standard descriptors for more realism, grainy disposable camera snapshot
I like the creepy simian Queen Elizabeth in this one.
In the most recent episode of Severance someone is called a “fetid moppet”. And the episode before that mentions “fruit leather” as a snack option. So I try for “fetid moppet eating fruit leather in an office cubicle”. SDXL, Flux, and even SD 1.5 know that “moppet” is a term for a child. Dalle-3 (via Bing) has no clue and consistently creates strange hairy creatures. (Imagen rejected the prompt as being unacceptable.)
(All of them seem to know that “fruit leather” is apparently what I always knew as “fruit roll-ups” and had never once before heard called “fruit leather”.)
Tonight some playing around led to a prompt involving a cat on a bicycle, and it made me remember this old request. I wanted to see how Bing/Dall-E 3 handled it. The results are of course much better, but still flawed. There especially is trouble with the handlebars and pedals. Here’s a gallery of tries.
As it happens, there’s now another AI image on flyers around the school, which wasn’t my doing. The flyers are suggesting that students try a “social media purge” for Lent (we’re a Catholic school), and show a bunch of social media icons being flushed down a toilet. I’ll see if I can get a picture.
Oh, and from that album, this one wasn’t too bad:
Model: Flux Schnell
Preset: Modern Comic
Prompt: funny anthropomorphic purple jaguar riding a red bicycle, wearing a top hat
Got both handlebars and one out of two foot pedals. Pretty good.
I’ve never managed to get an AI image of both feet properly arranged on bicycle pedals.
1 [almost] out of 3 ain’t bad:
I looked to see if Bing had a better concept of a pogo stick since the last time I tried. Still no. Occasionally one image is mostly right but they are usually from substantially wrong to wildly wrong.
You can always use start images, plus “depth to image”.
For example, take this image.
Then use the prompt “Frog on a pogo stick” gives this
Nightcafe, model Boltning XL V1 Lightning
Back in 2022 I posted about attempting to create a horror Great Pumpkin rising above a pumpkin patch using the arcane near-gibberish prompts we had to use back then. In my latest experimentation I tried the concept with Bing/Dall-E 3 with pretty good success finally, even able to insert Charlie Brown into the images. I then took one of those prompts and plugged it into other models, including the ancient Night Cafe “Artistic” and “Coherent”. Here are images created in each of the models, at the default available resolution of each model and in widescreen when available, square when not.
The prompt used:
Summary
terrified charlie brown escaping from a towering angry orange jack-o’-lantern ghost vine hybrid in the background in a scary pumpkin patch on halloween; by junji ito and dan witz, realistic grainy disposable camera snapshot with bokeh, motion blur, shallow dof and forced perspective.
While the SDXL and Flux images are coherent and attractive, only Bing “gets it”.
Here’s a gallery of more Bing outputs with various prompts, all outpainted to widescreen.
I think that Bing has fairly recently implemented what I suppose could be called a “cuteness filter” as censorship. Where for prompts that the “guardrails” don’t like instead of blocking it completely it renders something with a distinctive style with pastel colors and cute animals and flowers and butterflies and such that may only vaguely match elements from the prompt. A single word in the post can trigger it. Changing the one word back and forth gives constant results between “cute” images and “normal” images. For instance, the prompt
terrified scowling leucotic person snail fingerprint hybrid with many fingers tipped with hands that are holding eyeballs as legs. climbing a slimy toadstool in a rainy forest clearing; grainy disposable camera snapshot with bokeh, motion blur, shallow dof and forced perspective.
triggers the cuteness:
But change one word and it gets back to the general type of image I was looking for:
terrified scowling leucotic person snail fingerprint hybrid with many fingers tipped with hands that are holding eyes as legs. climbing a slimy toadstool in a rainy forest clearing; grainy disposable camera snapshot with bokeh, motion blur, shallow dof and forced perspective.
(BTW, “eyeballs” works in some other prompts, I’m not certain which word in this prompt it isn’t allowed to combine with.)
ChatGPT as of yet isn’t triggering anything yet when I tried the two different prompts:
Maybe layered on by Microsoft, then.
(The first is closer to the “holding eyeballs” I was imagining than the others.)
Another example of triggering the “cute” filter in Bing. Today I was playing with versions of this prompt:
four flabby wrinkled the beatles on a mushroom dressed as the beatles playing band instruments, in a x forest. grainy disposable camera snapshot with bokeh, motion blur, shallow dof and forced perspective.
I was trying to get a dead-looking forest in the background.
“scary forest” gets images like this:
“Halloween forest” gets images like this:
“scary Halloween” triggers the “cute” filter:
But “Halloween scary” goes back to normal.
The style changes with the modified prompts are consistent.
Nightcafe has new versions of Ideogram available. Handling of typography better than ever.
