Adventures in Gimp AI

[mtf]

Follow up to my post here : viewtopic.php?f=3&t=1191&p=71078#p71078

The results on that post were me just messing around in a rather haphazard and impulsive manner without taking notes or organizing output files. It was fun, but not particularly conducive to learning . So I’ve went back to the beginning to experiment in a slightly more structured way, and I figure I might as well post some of the results here, both as documentation and as examples.

I’m running ComfyUI locally on Ubuntu Linux 24.04 with an Nvidia RTX 4500 and I followed forum member xs70’s excellent primer posted as the beginning of that above thread.

Starting with this source image - A Gary Roberts sketch :

And working with the “Lineart to Flux” workflow : viewtopic.php?f=3&t=1191

I seem to get better results for this source image using the model from xs70’s other AI thread (“Existing Images” : viewtopic.php?f=3&t=1192) rather than the NSFW model. I’m not sure why - possibly because the subject is clothed, so the NSFW training doesn’t really matter. Neither model seems to be able to reliably generate the gag in her mouth without artifacts, but it seems to be a minor enough part of the original drawing for the AI to ignore if it’s left out of the description.

I settled on the following prompt to play around with. (As noted, leaving the gag out seems to get the best results for me, as does changing the ropes on her wrists to handcuffs.)

Code: Select all

(picture) of an abducted woman in a dungeon.

She is a slender platinum blonde with short hair.  She is wearing white short shorts that are unzipped and open in front and a tightly fitting blue crop top with no bra.  Her wrists are handcuffed together over her head.  A chain connects from the handcuffs up to the ceiling.  She is standing on her toes.  Her ankles are bound together with rope.  She is looking offscreen to her left with a terrified look on her face.

She is in a dungeon setting. There is a rectangular drain in the floor and a small window with bars in the wall.  The edge of a cell door is in the background.  A circular overhead light is at the top left of the picture.

S&M, B&D, bondage

Here’s a few of my favorite outputs from that prompt.

To keep forum resource usage down, I attached reduced size versions of the outputs. Larger versions :
https://hamster.is/images/2025/07/15/Sisters_Orig.png
https://hamster.is/images/2025/07/15/Si ... icture.jpg

[doe.1971]

Nice. Go on.

[mtf]

Playing around with the AI to get different output styles, I changed "(picture)" in the prompts. Here are some samples.

• "(Photograph)" - The results weren't really any different. My guess is the training data uses the two words interchangeably.

  

• "(Photorealistic painting)" - The results here are largely similar, though it does seem to use lighter colors. Interestingly the "painting" part seems to encourage the AI to occasionally put little "signatures" in the output. Presumably because the training data that's labeled "paintings" are often signed.

  

• "(Oil painting)" - Still largely similar to the others, but to my eye this seems to use more redish/orange flesh tones and produce more variation in art styles. Again we sometimes get "signatures" in the outputs.

  

Larger versions here:
https://hamster.is/images/2025/07/15/Si ... ograph.jpg
https://hamster.is/images/2025/07/15/Si ... listic.jpg
https://hamster.is/images/2025/07/15/Sisters_Oil.jpg

[yyy02]

... Playing around with the AI to get different output styles, I changed "(picture)" in the prompts. Here are some samples.
...

Thank you very much for making your experimental images available to us!
But when I try to access the larger images (from your links), I get a "403 Forbidden" message/white screen.
Is there something I could do to alleviate this?

[BobDarkroom]

... Playing around with the AI to get different output styles, I changed "(picture)" in the prompts. Here are some samples.
...

Thank you very much for making your experimantal images available to us!
But when I try to access the larger images (from your links), I get a "403 Forbidden" message/white screen.
Is there something I could do to alleviate this?

Same here.

[doe.1971]

Hamster doesn't allow hotlinking. Use something like https://postimages.org/

[mtf]

Whoops. My apologies folks. Once I cleared my browser cache, I saw the problem with the links. Will post the larger images using another hosting site later.

[mtf]

More output style experiment, changing "(picture)" in the prompts. Here are some samples.

• "(pencil sketch)" -

  

  Interestingly, the AI interprets this as black and white, but will still more often than not pick up the 'blue' in 'blue crop top.'
• "(charcoal drawing)" -

  

  Honestly when I put this in, I expected much more rough sketch/impressionistic style outputs rather than what got produced. Maybe the training data doesn't have much in the way of charcoal drawings, so it tends to focus on "drawing?"
• "(line art)" -

  

  Lots of mixed results here. I was hoping for mostly outputs in the style of the first and third examples, but it just as often produces more filled in artwork.

Will post link to the larger images later.

[mtf]

A few more output style experiments, changing "(picture)" in the prompts. Here are some samples

• "(watercolor painting)" -

  

  Mostly this seems to affect the style of the backgrounds, but I think this prompt yielded my favorites.
• "(anime drawing)" -

  

  The AI mostly does a pretty good job applying an anime style to her face, but not much else.
• "(comic book drawing)" -

  

  Here the AI seems to be pulling from modern comic book and graphic novel styles. I'm a little disappointed that it didn't do any Curt Swan or Jack Kirby inspired styles. Lack of training on those older comic books, maybe?

I'll post all the larger versions in a later post.

[mtf]

Last batch for this set of output style experiments: changing "(picture)" in the prompts. Here are the samples

• "(comic strip drawing)" -

  

  These are sometimes indistinguishable from the "comic book" drawings but sometimes you get a little more of a comic-strip style (eg. samples 2 and 3)
• "(cartoon drawing)" -

  

  This one was interesting - the styles varied from disney-inspired to 80's afternoon cartoon-ish to anime-ish
• "(hanna-barbera cartoon drawing)" -

  

  I really didn't know what to expect with this one, but the AI uniformly chose a brighter color palate, and about half the time it went for a simpler and flatter art style.

I did try some other styles like "surrealist painting," "cubist painting" and a few others, but the results were basically the same as the "picture" and "oil paintings." Obviously I didn't even begin to scratch the surface of the vast varieties of art styles out there ("Woodcut?" "Sidewalk Chalk Art?" "Cave painting?" LOL), but I just wanted to experiment to get a feel for what the AI and model was capable of.


Now, gotta give my graphics card a breather, but I have a few ideas for my next experiment. Back before long.


Larger sizes of all images here:
https://postimg.cc/gallery/RKHWQHm