Theo’s AI Quicksandbox

[Theo]

I wouldn't rule out the possibility that, if you use a striking feature, like red hair, the AI will draw on the available images that exhibit that feature in it's generation process. Consequently, similarities to people we are familiar with are not that surprising, particularly, where the scenario fits as well. In your recent excellent images I can see hints of Ludella, Emily Marilyn (dark haired version), a neck deep Crystal, etc; I'm not convinced these similarities are totally coincidental. Anyway, great images, really enjoyed them.

Thank you cerberus! I don't think I've heard of Emily Marilyn, but thanks for introducing me to her And who is Crystal?

So you're saying that because someone like Emily Marilyn is a fetish model, or Ludella who actually does quicksand videos, that the AI draws on them as an example, even though I'm not using their names in the prompt?

[cerberus]

Yes, basically that's it, AI has to get it's source material from somewhere and that's most likely going to be the web. So people who are popular and have lots of images on the web will have more influence on the results. It's not necessarily fetish/porn that dominates but e.g. if you go to Google Images and search for women in quicksand you'll find a lot of images that originate in this community, which, I think, will bias the output.

This is Crystal:

https://quicksandvisuals.umd.net/downlo ... solo-scene

[Theo]

Yes, basically that's it, AI has to get it's source material from somewhere and that's most likely going to be the web. So people who are popular and have lots of images on the web will have more influence on the results. It's not necessarily fetish/porn that dominates but e.g. if you go to Google Images and search for women in quicksand you'll find a lot of images that originate in this community, which, I think, will bias the output.

This is Crystal:

https://quicksandvisuals.umd.net/downlo ... solo-scene

Interesting. I kinda understood that it would use photos from the web, but I assumed it'd only be of the person named in the prompt, unless it's just a generic "blonde woman sinking in quicksand" or something, than maybe it would more broad.

Oh you mean "Krystal." I'm very familiar with her

[cerberus]

Interesting. I kinda understood that it would use photos from the web, but I assumed it'd only be of the person named in the prompt, unless it's just a generic "blonde woman sinking in quicksand" or something, than maybe it would more broad.

Oh you mean "Krystal." I'm very familiar with her

Really, I'm just guessing, but you can add your own prompt tags to SD using Textual Inversion, where you supply a series of pictures highlighting the feature you're after. I'd expect the main checkpoint models to be built in a similar manner. I've noticed that the various checkpoint models I use each generate a different selection of faces and I sometimes switch between them if I want a different "look" for the subject. I think it would be reasonable to guess that this is down to the images fed the AI when building the models.

Yes, I meant Krystal, my implicit pedanticism tends to auto-correct these type of things!

[Theo]

For this week, here’s one of my most ambitious AI projects yet. I ended up with too many pictures so I decided to switch to a different format (good old PDF). However due to file extension limitations here I had to post it in stories, so simply follow the link -> Fond Memories

Please enjoy, and let me know whatcha think!

[cerberus]

I did indeed enjoy. Getting that many consistent pictures must have been quite a challenge and a delightful story as well. Well done, I'm impressed.

[Theo]

I did indeed enjoy. Getting that many consistent pictures must have been quite a challenge and a delightful story as well. Well done, I'm impressed.

Thank you Cerberus! (takes a little bow ). Like I’ve said before, when you generate 400+ pictures with the same character prompt it’s not too hard (just time consuming). But it took just as much time to sort through the ones I saved and pick out the ones I liked.

Interestingly, Perchance seemed to bounce between two slightly different characters and styles with the exact same prompt. I made some side-by-side examples for comparison (I have plenty of extra pictures to pick from lol). The first one on the left was more Caucasian looking and animated-ish in style I thought, and the second one was maybe little more Asian and more lifelike in appearance. I really liked both; almost went with the first one, but in the end, I thought the second one look so much more lifelike. So that became the chosen one.

Extra 01.jpg

Extra 02.jpg

Extra 03.jpg

Extra 04.jpg

[cerberus]

Like I’ve said before, when you generate 400+ pictures with the same character prompt it’s not too hard (just time consuming). But it took just as much time to sort through the ones I saved and pick out the ones I liked.

I know what you mean. If you are trying to do a sequence it is a good idea to keep the clothing simple (or missing) as you've done here, otherwise you get a different outfit with each shot as well.

Interestingly, Perchance seemed to bounce between two slightly different characters and styles with the exact same prompt. I made some side-by-side examples for comparison (I have plenty of extra pictures to pick from lol). The first one on the left was more Caucasian looking and animated-ish in style I thought, and the second one was maybe little more Asian and more lifelike in appearance. I really liked both; almost went with the first one, but in the end, I thought the second one look so much more lifelike. So that became the chosen one.

I noticed that when I used Perchance, I think it is depends on the images used to train the checkpoint models and which model/s is/are being used. With SD, if I use different models I get different faces.

Oh, and I think you made the right choice, the images on the right are more lifelike.

[Theo]

I know what you mean. If you are trying to do a sequence it is a good idea to keep the clothing simple (or missing) as you've done here, otherwise you get a different outfit with each shot as well.

I noticed that when I used Perchance, I think it is depends on the images used to train the checkpoint models and which model/s is/are being used. With SD, if I use different models I get different faces.

Oh, and I think you made the right choice, the images on the right are more lifelike.

Thanks! Funny thing about the clothing, I never even specified the color of the bikini, and 99% of the time it turned out the same. Probably because I specified "maroon lipstick" because if I say "red lipstick" it can be a little too bright fire engine red, and the AI just decided to apply the same color to the bikini, which was fine by me cause I liked the way it looked anyways.

I haven't tried costumes in a while as like you said it's too easy to go nuts trying to get the right progression and continuity, but I wish I could do more in that department. Be nice if you could somehow create a specific character and save it as "JaneDoe001" or whatever, then all you have to say in the prompt is "JaneDoe001." But that would be way too easy

[Theo]

We know pirates and explorers get into a lot of trouble with quicksand, but what about cowgirls?

Cowgirl01.jpeg

Cowgirl02.jpeg

Cowgirl03.jpeg

Cowgirl04.jpeg