AI-generated images thread

[Jopsy Pendragon]

Woman’s iPhone photo of son rejected from Sydney competition after judges ruled it could be AI

Probably Not AI: Hands / reflection
Possible post-processing: odd halo around his hair & shoulders
Maybe AI: there's something about his left eye that seems unmatched with his right in that sorta way that AI struggles with.

*shrug* who knows. Also, why is there a gucci exhibit in a bathroom?!

 

[Noodles]

The real tell here, is that AI doesn't really properly understand perspective and thus doesn't properly understand mirrors.

 

[Argent Stonecutter]

Getting better.

 

[Argent Stonecutter]

Studio Ghibli version.

 

[WolfEyes]

Doesn't understand perspective and reflection.

Studio Ghibli version.

Ah. Dixon Hill.

 

[Chalice Yao]

And this is what I get from spelling Isambard correctly. The fox *is* Isambard!

Hey. Hey, pssst. Hey.

The secret lies in inpainting. Essentially I did an image that was simply "a humanoid fox painting a painting", then I switched to inpainting, I marked the area of the painting (it was showing something different) and made the prompt "Isambard Brunel's portrait, realistic" (The 'realistic' just made sure it would not be a rather weird painting) and after a couple of tries it turned out like that.

Here is the original, for reference:

And you can repeat the inpainting with other parts of the image. Seriously, that stuff is what takes AI image gen way beyond "let me try a prompt set over and over again, and cross fingers"


Sadly there is no inpainting tutorial for AI Runner, but maybe w4ffl35 can provide one!

.

 

[Argent Stonecutter]

I'm not trying to produce a specific picture, I'm trying to explore how much of the prompt is actually parsed and how phrases actually control the generation. So far it seems to be "not very much". Individual keywords make it likely that there will be specific elements in the picture, but those keywords are phrases of not more than three or four words and will be collapsed to single words if it's even vaguely inconvenient.

That is, I'm more interested if there's any meaning to the generator in "a humanoid fox painting a picture of thing that thing things", and how many of those qualifiers make it through the process.

Until recently the two word phrase "humanoid animal-name" was interpreted as "animal-name" and "animal-name doing thing" more often than not produced a picture of "a human doing thing" with an animal in there somewhere.

So when "a picture of a humanoid ferret wearing a fedora and a trenchcoat, leaning against a lamppost while staking out a gin joint" gets the humanoid ferret and kind of the right clothes, and a lamp-post... it's still not very good but it's way better than it has been.

 

[Bartholomew Gallacher]

Looky looky at what I got, no inpainting. It is possible to get it without, but I had to generate several random seeds until one fitted my expectation. Also I had to tweak the prompt a lot to get the AI understand what I want, which was a good exercise. It could need a little bit of tweaking to be amazing, but so far I am quite pleased with the outcome.

Here are the parameters:
Model: Rev Animated
Sampler: DPM++ 2 Karras
CFG Scale: 10
CLIP skip: 2
Sampling steps: 30
Seed: 871817926
Prompt: A nice looking fox girl wearing no hat is standing in her atelier. An easel is mounted on a staffage. With her trusty paint brush in her hand the fox girl is painting an illustration of Isambard Brunel on the easel, which she is looking at. Fantasy style, D&D, 4k, masterpiece

Result:

 

[Kamilah Hauptmann]

His hands are fused and one of his fingers is a penis.

 

[Bartholomew Gallacher]

Well, that's the problem with SD 1.5 based models at the moment, but right now I don't feel like editing that. For me the overall image composition and getting the vibe was the main challenge. It was really not easy to make the painter this fox girl without a hat while Isambard wears one and also to make her paint Isambard. It seems Isambard is so heavily connected to his hat, when pulling him into a picture alone suddenly all other peoples are wearing that cylinder as well.

I used for many tries kitsune as description to get a convincing fox girl, since these are in Japanese mythology nine tailed foxes, which can shape shift into human/anthropomorphic forms and back into an animal. But this didn't work out so well.

 

[Argent Stonecutter]

Very impressive, but I can't unsee Isembard's finger now. Thank you Kamilah.

Also her tail is growing out of her left sleeve.

 

[CronoCloud Creeggan]

Studio Ghibli version.

That's one handsome ferret.

 

[Erich Templar]

Studio Ghibli version.

Best one so far.

 

[Free]

 

[Noodles]

Hands, how DO THEY WORK????

 

[Noodles]

Prompt: "Stable Diffusion as a Waifu"

It actually made several pretty consistent versions for random seeds. Is this how Stable diffusion sees itself?

 

[Argent Stonecutter]

A fox sitting at a kitchen table trimming its claws. Various attempts at describing things to convince stable diffusion that a fox can hold a pair of nail clippers.

... so it goes.

 

[Noodles]

That last one sort of got it. Though that first one is kind of an interesting though of, how would tools be designed for creatures without thumbs, but they are still smart enough to design things.

Just stick that paw in and mash it!

 

[Argent Stonecutter]

I kind of liked the telekinetic fox, but the last one appears to be trying to trim its claws while wearing gloves.

Edit: I think this customer needs an orthopedist, not a manicure...

 

[Yure4u Sosa]

eku and i made this, enjoy( like and subscribe or eku will smack you)