Changing a Flat

[BBB]

"Changing a Flat"

https://www.imagebam.com/view/MEMUW9I

I'm glad the forum is back up so I can finally post this: my latest attempt at using an AI image generator to show process. Breast expansion is the main attraction, with a healthy dose of bimboification thrown in, too. Each frame uses the last frame as input, with a low denoising value to cause gradual changes. It may take a minute before it animates smoothly.

[Prinz Eugen]

very nice !

[jwyattguy]

Conspiracy theorists: "The end is nigh! AI will take over and kill us all.

BE lovers: "Haha, grow big booba."

Anywho, jokes aside... The waist got a biiit too narrow for my liking but, otherwise, awesome work!

[pingu1]

I haven’t found a good AI generator that produces expansion. Can you recommend one please? Or offer advice on the commands to put in haha

[BBB]

Originally Posted by Prinz Eugen

very nice !

Thank you!

Originally Posted by jwyattguy

Conspiracy theorists: "The end is nigh! AI will take over and kill us all.

BE lovers: "Haha, grow big booba."

Anywho, jokes aside... The waist got a biiit too narrow for my liking but, otherwise, awesome work!

Ha, thank you!

Originally Posted by pingu1

I haven’t found a good AI generator that produces expansion. Can you recommend one please? Or offer advice on the commands to put in haha

I'm really hooked on this stuff. I feel like I could talk about it for hours.

I'm using pornworks.ai. The watermark they add is supposed to say that, but for some reason it gets cut off when I use it. Pornworks has several advantages: there's no signup needed, you can input whatever prompt you want, and you can use images as input.

The prompt and negative prompt are what matter most. You can find an image in pornworks' public gallery that's kind of like what you want, and use it as your starting template to get a feel for the syntax. To avoid redundancy, I put in extra line breaks to organize the prompt and negative prompt tags into groups including:
what the subject looks like (size, shape, clothes, pose)
what the background look like
what kind of image is it (realistic color photo, 3-point lighting, highly detailed, etc.)

In order to get process, I separate out the things that are going to change from those that aren't. You need to describe a baseline in order to generate an image that isn't too different from the original. Pornworks has a handy link to a "prompt predictor" that tells you how AI would tag your starting image. For this project, that part of the prompt that was always the same was, "a young naked woman with downcast eyes in front of a bright green car,
platinum blonde, tanned skin, tanlines, fully nude, blurred background, open mouth,". You should do some trial runs to see if your endpoint (last frame) will get to where you want it to be.

I was hoping to basically have one prompt specifying huge breasts, and rely on the low denoising value to make each step increment a bit towards that endpoint. But that tends not to work so well. So I ended up writing six different prompts, and using each one to get a couple iterations. The first prompt had tags like very small breasts, skinny, and thigh gap. The last prompt had tags like G cup breasts, thicc, and Barbie. And some of the original prompt tags got put in the negative prompt, to e.g. specifically not do a thigh gap by the end.

The way you get an animation out of this is to take one output and feed it in (using the "use as reference" button) as the input for another generation. When you do, the denoising value specifies how much the new image will deviate from the input. I put denoising on 2/10, so that the change between frames is gradual. If denoising is too high, it'll skip to the end without giving you many intermediate frames. If denoising is too low, it'll take too many renders, and more things will morph in out of nothing.

Notice how the earrings appeared out of nowhere. I never specified earrings - random noise just turned into earrings. I rolled with it, but it can start doing things you don't want, like adding a sixth finger to a hand. So you have to be vigilant and check each frame you render for emerging problems, and re-render the frame if something crops up, before it becomes more prominent.

That's probably way more than you wanted to read, so I'll stop now.

[pingu1]

Thanks so much. You really know your stuff!

[BBB]

I call this, "One Filled-Out Filly".

https://www.imagebam.com/view/MENTMK9

(Once again it's slow to load for me, but I can't tell if it's because my computer is old and slow, or because GIFs aren't supposed to be this big.)

Breast expansion is the primary process, though there's a lot of weight gain too.

I'm trying to get more efficient at these, but there's still a lot of time spent on every frame. Both in the AI image generator - staying vigilant for compounding errors and insuring the important parts actually change in the right direction every step - and in Photoshop. I described the details of the model's face, hair, and makeup in more detail this time. Alas, I fear her face still changes rapidly, making her not look like the same woman in the end, if one were to not see the intermediate steps.

On this one I was fighting the AI in the second half. Despite a negative prompt telling it to not do monochrome, black & white, or a simple background, it was determined to desaturate the images more and more, and morph the background into something muddy and unrecognizable. It's like making a photocopy of a photocopy. It's not clear why "Changing a Flat", which has the same number of steps, didn't have this problem.

So I had to roll with the image degradation, adding the zoom in and glow/fade to white. It wasn't desirable, but it's how I could salvage all those generated images. Of course, it would be easy to render a single final image and cross fade from the before to the after. But that would be barely better than a 'poof' transformation. It's seeing the movement, the continuous change in size over time, that makes believable process.