Adventures in Gimp AI

[mtf]

Time for another experiment. Same workflow, but with a new image.

This is a scan for the cover for a 70’s or 80’s Adult/Erotic Novel. I have no idea who the artist is.

I was curious as to some of the limits to the AI model’s ability to “understand” the content of the image using just the edge-detection/controlnet and bare minimum description.

Code: Select all

book cover featuring a photograph of a masked man and a naked woman tied to a tree.

highly detailed, photograph

These are just some samples of what it came up with. Sometimes pretty good and often quite laughable, but the control net was clear enough to get the basic two figures, a tree and “GIRL HUNTERS.” Beyond that the AI sometimes got … creative with “masked” and “to a tree” and which figure was which and who was naked.


The contrast in the original image between the title and background plus the thickness of the letters was enough that the edge-detection basically did a perfect job with “GIRL HUNTERS,” but rarely got the rest of the text. I also wanted to give the AI a hint about which figure was the man and which was the woman so I added a little detail

Code: Select all

book cover featuring a photograph of a masked man and a naked woman tied to a tree.

In the upper left is the word "ABUSED"
In the upper right is the word "$4.95"
Across the bottom are the words "For The Entertainment of Adults Only"

On the right of the photograph is a naked 25 year old blonde woman.  

On the left of the photograph is a man wearing black.

highly detailed, photograph

This time, the AI was quite on point with getting the man on the left, suiting him up in black as well as putting a naked woman on the right tied to a tree. It also got much better (but not perfect) with the text, though it had a lot of trouble with the “of Adults Only” in the lower right, probably due to the edge-detection having trouble with the low contrast between those words and the woman’s thigh. (I experimented with the edge detection levels but didn’t get significantly better without affecting other parts of the picture.)

The AI still went off the rails with the picture content sometimes, especially with the woman’s torso and the guy’s left hand.

More to come.

[mtf]

Continuing on experimenting, I put more detail into the prompt.

Code: Select all

book cover featuring a photograph of a masked man and a naked woman tied to a tree.

In the upper left is the word "ABUSED"
In the upper right is the word "$4.95"
Across the bottom are the words "For The Entertainment of Adults Only"

On the right of the photograph is a naked 25 year old blonde woman tied to a tree with rope. She on the right of the photo, facing to the left with her back to a tree trunk.

On the left of the photograph is a man wearing black pants, a long sleeved black shirt and a black balaclava. His left hand is on the woman's head.

highly detailed, photograph

Adding some detail about how the man is dressed and the position of his left hand really seems to help out the AI with that. Adding a bit of detail about the woman doesn't seem to help the hallucinations much. And as always sometimes it just does weird things.

Code: Select all

book cover featuring a photograph of a masked man and a naked woman tied to a tree.

In the upper left is the word "ABUSED"
In the upper right is the word "$4.95"
Across the bottom are the words "For The Entertainment of Adults Only"

On the right of the photograph is a naked 25 year old blonde woman tied to a tree with rope. She on the right of the photo, facing to the left with her back to a tree trunk. She has wavy blonde hair and large firm breasts.

On the left of the photograph is a man wearing black pants, a long sleeved black shirt and a black balaclava. His left hand is on the woman's head. His right hand is in front of him. He is looking to his left at the woman lewdly.

highly detailed, photograph

Adding just a little more detail to the woman seems to help the AI distinguish her arms from the outline of the tree trunk. Adding more detail to the guy at this point seems to be unnecessary - the edge detection clearly outlines his right hand. Somewhere around this level of detail seems to be the sweet spot for this workflow.

Code: Select all

book cover featuring a photograph of a masked man and a naked woman tied to a tree.

In the upper left is the word "ABUSED"
In the upper right is the word "$4.95"
Across the bottom are the words "For The Entertainment of Adults Only"

On the right of the photograph is a naked 25 year old blonde woman tied to a tree with rope. She on the right of the photo, facing to the left with her back to a tree trunk. She has wavy blonde hair and large firm breasts. Her arms are to her side, tied to the tree trunk.

On the left of the photograph is a man wearing black pants, a long sleeved black shirt and a black balaclava. His left hand is on the woman's head. His right hand is in front of him. He is looking to his left at the woman lewdly.

highly detailed, photograph

Adding on even more detail to the description doesn't really do anything. The AI pretty much gets the composition of the original art. It varies the backgrounds a little bit since we didn't describe it, and does the usual odd things with fingers, rope and odd additions.

more later

[mtf]

Continuing to experiment with the book cover.

Code: Select all

book cover featuring a photograph of a masked man and a naked woman tied to a tree.

In the upper left is the word "ABUSED"
In the upper right is the word "$4.95"
Across the bottom are the words "For The Entertainment of Adults Only"

On the right of the photograph is a naked 25 year old blonde woman tied to a tree with rope.  She on the right of the photo, facing to the left with her back to a tree trunk.  She has wavy blonde hair and large firm breasts.  Her arms are to her side, tied to the tree trunk.  Her head is tilted back slightly, looking at the man with a scared expression.

On the left of the photograph is a man wearing black pants, a long sleeved black shirt and a black balaclava. His left hand is on the woman's head.  His right hand is in front of him. He is looking to his left at the woman lewdly.

highly detailed, photograph

Adding even more detail to the woman's description doesn't really change much.

Code: Select all

book cover featuring a photograph of a masked man and a naked woman tied to a tree.

In the upper left is the word "ABUSED"
In the upper right is the word "$4.95"
Across the bottom are the words "For The Entertainment of Adults Only"

On the right of the photograph is a naked 25 year old redheaded woman tied to a tree with rope.  She on the right of the photo, facing to the left with her back to a tree trunk.  She has long red hair and large firm breasts.  Her arms are to her side, tied to the tree trunk.  Her head is tilted back slightly, looking at the man with a scared expression.

On the left of the photograph is a man wearing black pants, a long sleeved black shirt and a black balaclava. His left hand is on the woman's head.  His right hand is in front of him. He is looking to his left at the woman lewdly.

highly detailed, photograph

Keeping the most verbose version of the prompt but changing the description of the woman to be redheaded usually works with various shades from strawberry-blonde to almost crimson. (Not shown was a few outputs where her hair was just blonde).

Code: Select all

book cover featuring a photograph of a masked man and a naked woman tied to a palm tree.

In the upper left is the word "ABUSED"
In the upper right is the word "$4.95"
Across the bottom are the words "For The Entertainment of Adults Only"

On the right of the photograph is a naked 25 year old asian woman tied to a tree with rope.  She on the right of the photo, facing to the left with her back to a tree trunk.  She has long black hair and large firm breasts.  Her arms are to her side, tied to the tree trunk.  Her head is tilted back slightly, looking at the man with a scared expression.

On the left of the photograph is a man wearing black pants, a long sleeved black shirt and a black balaclava. His left hand is on the woman's head.  His right hand is in front of him. He is looking to his left at the woman lewdly.

The background is a tropical beach.

highly detailed, photograph

Keeping the most verbose prompt but switching the description of the woman to be "asian," changing the tree to "palm tree" and adding a "tropical beach" background description yields interesting results. The tree foliage definitely take on a palm-ish look and the background is generally tropical. But I'm guessing the models don't have a lot of training on actual palm tree trunks. To my eye, the women tended to look southeast-asian or pacific-islander mixed - I wonder if that's because the inclusion of both "tropical" and "asian woman" in the text influenced the model to pull heavily on training pictures from places like Hawaii or the Philippines. It could also just be due to small sample sizes (I only generated 16 outputs)

[mtf]

Last batch of this experiment:

Code: Select all

book cover featuring a photograph of a masked man and a naked woman tied to an autumn tree.

In the upper left is the word "ABUSED"
In the upper right is the word "$4.95"
Across the bottom are the words "For The Entertainment of Adults Only"

On the right of the photograph is a naked 25 year old hispanic woman tied to a tree with rope.  She on the right of the photo, facing to the left with her back to a tree trunk.  She has brown wavy hair and large firm breasts.  Her arms are to her side, tied to the tree trunk.  Her head is tilted back slightly, looking at the man with a scared expression.

On the left of the photograph is a man wearing black pants, a long sleeved black shirt and a black balaclava. His left hand is on the woman's head.  His right hand is in front of him. He is looking to his left at the woman lewdly.

The background is a suburban park in the autumn.

highly detailed, photograph

Changing the description of the woman a bit to "hispanic woman" with "brown wavy hair" and the background to a "suburban park in the autumn." Honestly very impressed with the background on these - the AI put fall colors on the trees and lots of little glimpses of suburban streets, homes etc.

Code: Select all

book cover featuring a photograph of a masked man and a naked woman tied to a bare tree.

In the upper left is the word "ABUSED"
In the upper right is the word "$4.95"
Across the bottom are the words "For The Entertainment of Adults Only"

On the right of the photograph is a naked 25 year old black woman tied to a tree with rope.  She on the right of the photo, facing to the left with her back to a tree trunk.  She has dark curly hair and large firm breasts.  Her arms are to her side, tied to the tree trunk.  Her head is tilted back slightly, looking at the man with a scared expression.

On the left of the photograph is a man wearing black pants, a long sleeved black shirt and a black balaclava. His left hand is on the woman's head.  His right hand is in front of him. He is looking to his left at the woman lewdly.

The background is a snowy mountain top

highly detailed, photograph

Changing the description of the woman to "black woman" with "dark curly hair." This one had me confused for a while - more than one output came out in black and white, and a few came out with dirty-blonde white women. I suspect the problem was that the prompt included "black pants," "black shirt" and "black balaclava" and so the AI got confused, sometimes applying "black" to the overall picture or other aspects.

Code: Select all

book cover featuring a photograph of a masked man and a naked alien woman tied to an alien tree.

In the upper left is the word "ABUSED"
In the upper right is the word "$4.95"
Across the bottom are the words "For The Entertainment of Adults Only"

On the right of the photograph is a naked 25 year old alien woman tied to a tree with rope.  She on the right of the photo, facing to the left with her back to a tree trunk.  She has purple wavy hair, green skin and large firm breasts.  Her arms are to her side, tied to the tree trunk.  Her head is tilted back slightly, looking at the man with a scared expression.

On the left of the photograph is an alien wearing a space suit. His left hand is on the woman's head.  His right hand is in front of him. He is looking to his left at the woman lewdly.

The background is a sci-fi alien landscape.

highly detailed, photograph

This one was just for fun.


Larger sizes of all images here: https://postimg.cc/gallery/Df0CwfC

[mtf]

Small experiment this time.

Starting with image, another 70's/80's erotic novel cover.

and this prompt :

Code: Select all

Photograh of a naked blonde woman in bondage in a dungeon

She has long curly blonde hair and large firm breasts.  Her mouth is open in distress.

Her arms are bound behind her back with leather straps.  Her chest is thrust out towards the camera.  Her wrists are bound with leather straps to a leather belt around her waist.  She has ropes around her waist and through her crotch.

She has large firm breasts

highly detailed, photograph

Without a background and high contrast, the edge detection does a good job with hinting to the AI and the results are pretty good. Even the rope, which has been problematic in previous experiments generally comes out decently more often than not.

It oddly has trouble with the "v" in "Captive" though, presumably due to the thin font. Specifying the text in the prompt would presumably have gotten it right most of the time, but that's not what I wanted to experiment with this time.

[mtf]

For this experiment, I modified the workflow (originally posted by xs70 here)

I modified step 2: ControlNet)

To add a mask:

Then I could click on the image, and "paint" what I wanted to mask out (black out what I don't want, then "invert" that to create the mask). This results in the edge detector "not seeing" the title and side words, and the resulting image being generated without them.

I discovered later on when trying this with another image that this works particularly well with this example because the text I masked out was on a black background, so the edge detection didn't see the edge of the masked area. So this isn't a good general purpose solution, but I figured I'd mention/document this experiment anyway.

Full images here

[mtf]

Following up on my previous experiment.

Starting with another 70's/80's erotic novel cover.

This time I loaded the original image up into an image editor and ran a "Canny Edge Detector" on it and fiddled with the levels and settings until I got an outline I liked and saved it.

Then I edited the outline to remove the title and price. I also ended up removing the signature because the AI often interpreted it in the outline to be part of the ropes.

[mtf]

I then modified the ControlNet (See previous experiment) to load the ControlNet image separately instead of generating it inside ComfyUI.

Then I ran the Line Art to Image workflow with the following:

Code: Select all

Photograh of a naked blonde woman in bondage

She has medium length curly blonde hair, blue eyes and large firm breasts. 

She has a white cloth gag in her mouth that is wrapped around her head and tied in the back. Her arms are tied behind her back by rope.  She is looking at the camera with a scared expression.

highly detailed, photograph

And here's the results.

Full size images here

[mtf]

TIme for a little more fun with generative AI.

Back to using a Gary Roberts piece as the origin - the cover for his Dofantasy release "Starfuckers"

I masked out the text (see previous post), the guy's right hand (the one grabbing her bikini string) as well as his legs and cock/balls. The AI had a terrible time figuring that part of the picture out, no matter how much I tweaked the input description and the outline. Maybe it could have done better with a different model and some inpainting, but I'm not sure.

I used this prompt:

Code: Select all

Photograph of Britney Spears in bondage.

In the foreground, center right, out of focus, is a gloved hand holding a knife.

Britney is standing with her arms over her head. She is looking down and to the left at the knife with a terrified look on her face. Her wrists are handcuffed together. A chain connects the handcuffs to the ceiling.  She is wearing a thin choker around her neck. She is wearing pink earrings. She is wearing a pink string bikini bottom and a pink top, which is pulled up exposing her breasts.

In the background on the left, there is a wooden door with metal frame.  On the right in the background, there is a wooden beam with an embedded metal ring.

There is a thick black border around the picture.

highly detailed, photograph, naked, bondage B&D, S&M, bokeh

Here's a few of the better outputs. In addition to the usual issues with fingers and the expected weirdness around the chain, the AI seemed to have great trouble with looking to "her left" or "the right." Regardless of what I wrote, it was pretty random which way it would generate her eyes looking.

[mtf]

I specifically used "Britney Spears" in the description for this experiment because I wanted to experiment with a LORA and use the same prompt throughout the experiment. There are a couple of "Britney Spears" FLUX LORAs out there, and I picked one kind of randomly, then applied at a couple of strengths to see what would happen.

Using the same prompt, but with the lora at strength 0.4

At this strength, the generated image usually takes on a "Britn-ish" look (but not always)


lora at strength 0.8

At this strength, the images pretty much always take on a Britney or Britney look-alike appearance, but odd artifacts also seem to creep in more often, like the nipples-through-shirt in the 2nd image, or her top being completely missing in the 3rd. I had to generate more images to get this set.

lora at strength 1.2

At this strength, the images pretty much always take on the same unambiguously Britney-like but expressionless blank look. The AI also gets more "creative." The border image isn't always black, the knife, clothes, chain or background deviates from the description or take on a non-reality look. I had to generate quite a few images to get this batch.

My guess is that for this image, and this lora, the sweet spot would be around 0.5 or 0.6

Larger version of the images : here

The lora I used is on huggingface as "britney-spears-flux-model"