[Hiram]

Fabulous.

[MaxBigfoot]

 

[MaxBigfoot]

 

[MaxBigfoot]

A properly themed post.   

[Nimrod]

I decided to study a bit more of the AI Art Generation process. Just wanted to see what other things could be accomplished. So, after some research into how the image is generated I've determined the core limitations of the AI image generation:
1. It is only as good as the reference library used to act as the feeding material which in turn is tied to...
2. The description (seed text) that more closely follows that library which still...
3. Can't generate a novel item or content.

That is why, so long as you use terms and reference a "style" library that has something similar to the terms you can create some of these rather realistic if amazingly proportioned women.

I experimented and found if one deliberately uses a reference (training content) that does not contain elements of the seed text you just can't generate anything relatable to your intent. That is why so many of these augmented and exaggerated women have a similar style or how two renderings will look more like variations of the same pose. That started me thinking.

What would happen if you start with an image and then describe BE onto it?
The problem is that all these AI generators are great at matching image to text, but not good at coming up with something not already within a seed image, regardless of the seed text in one step.

So, what happens when you use AI to create a fantasy woman with a generic description, that is an image created from pulling from the reference library and feed that image back into the AI, this time with a BE description. That way the original is built from the same stock of reference as the subsequent BE efforts.

Repeating the process with the result every time and looping many times with the description not directly describing the result but rather only a larger chest from the previous and setting the "strength" or permanence of the source image (just low enough to allow for change) one can generate a BE animation via AI art generation.

This is a fine line. Any stronger and the previous image locks in the major attributes and prevents growth, a little too low, and the image generated is too radically different to make a viable animation. It took a great deal of work finding not only that balance, but also the seed prompt to use across all frames that still allowed for BE and didn't imply some arbitrary maximum after which growth stopped.

So, here is a slightly different use of AI art generation - BE animation.
Essentially 64 consecutively inter-dependent images based of an original AI base image looped each time with a growth descriptive seed text all referencing a conducive library.
Hope people like.

[Nimrod]

Building on what I mentioned above.

Here is the effect of having a slightly TOO TIGHT permanence to the base image while still doing the feed the previous AI generated image back into itself. This also provided insight into the reference library used to train this particular style. Two things should be noticeable:

1. The apparent race of the subject slowly shifts. Apparently there are no very busty mongoloid women in this library. The text kept building upon the chest of the previous and at some point she grew beyond the reference women of Asian persuasion.

2. Because the base starter image in a bit too controlling, the BE is relative. If one watches to overall area occupied by the chest, it is rather fixed. It is the rest of her that shrinks to get a similar effect. The breasts don't really cover more image area. Again, AI doesn't like to create what is not there, it simply shifts.

At the very end I wanted to see if just adding "Caucasian" to the seed text would effect anything. The final image is essentially the previous with a different skin tone. Interesting. I tried to make her Mongoloid and every time her chest reduced significantly along this image sequence. This AI tied bust size to race only in that degree. It appears to be a strong association within that training library.

Enjoy.

[MaxBigfoot]

 

[T_Man]

Nimrod, thank you for the interesting insights.

This is from AnotherAIPrompt at DA

[Shadowmuse Blown]

^Great stuff guys, thanks for posting!
I love this one from Pikasoian on DA

[CarlTL]

 

[MaxBigfoot]

 

[Nimrod]

What happens when you balance both base image permanence (rigid adherence vs. fluid interpretation) and tap into a supportive reference library with a pure open-ended BE description and loop the images again and again ?

You get the following sequence of 176 frames.

(It eventually stabilizes as can be seen in the larger sizes)
This demonstrates that as the size increases the image migrates toward an obvious base image until the seed text pushes it beyond a size threshold and she shifts toward another reference image. It ends up at a maximum because there is no other viable image in the library to get any larger per the original base image.

Also - I deliberately tried "make" her Asian and the AI tried as best it could. But, just like one of the earlier tests, it eventually shifts to "bigger" races which I still find silly.

[Charade Recherche]

^Thats fascinating, Nimrod.  I had no idea.  Thanks for talking us through it and giving us the demonstration.  Really cool.

[Shadowmuse Blown]

Big turtles

by rodesca on DA

[SamV]

I decided to study a bit more of the AI Art Generation process. Just wanted to see what other things could be accomplished. So, after some research into how the image is generated I've determined the core limitations of the AI image generation:
1. It is only as good as the reference library used to act as the feeding material which in turn is tied to...
2. The description (seed text) that more closely follows that library which still...
3. Can't generate a novel item or content.

That is why, so long as you use terms and reference a "style" library that has something similar to the terms you can create some of these rather realistic if amazingly proportioned women.

I experimented and found if one deliberately uses a reference (training content) that does not contain elements of the seed text you just can't generate anything relatable to your intent. That is why so many of these augmented and exaggerated women have a similar style or how two renderings will look more like variations of the same pose. That started me thinking.

What would happen if you start with an image and then describe BE onto it?
The problem is that all these AI generators are great at matching image to text, but not good at coming up with something not already within a seed image, regardless of the seed text in one step.

So, what happens when you use AI to create a fantasy woman with a generic description, that is an image created from pulling from the reference library and feed that image back into the AI, this time with a BE description. That way the original is built from the same stock of reference as the subsequent BE efforts.

Repeating the process with the result every time and looping many times with the description not directly describing the result but rather only a larger chest from the previous and setting the "strength" or permanence of the source image (just low enough to allow for change) one can generate a BE animation via AI art generation.

This is a fine line. Any stronger and the previous image locks in the major attributes and prevents growth, a little too low, and the image generated is too radically different to make a viable animation. It took a great deal of work finding not only that balance, but also the seed prompt to use across all frames that still allowed for BE and didn't imply some arbitrary maximum after which growth stopped.

So, here is a slightly different use of AI art generation - BE animation.
Essentially 64 consecutively inter-dependent images based of an original AI base image looped each time with a growth descriptive seed text all referencing a conducive library.
Hope people like.

Like others I have posted AI generated pics in this thread that I found interesting. I have also posted some comments from a few AI artists, about what it takes to produce an AI image that is what they had intended, whither real, or more fancifully, without some manual processing. Nimrod's comments here only prove that the state of AI, while a powerful tool, still has far to go until someone like me with little artistic skill can even get close to creating something that remotely looks real.