Help With AI Prompting

[Theo]

I've been trying for weeks to get Perchance to sink a character to the depth shown in the picture below, and the deepest I can get using a "submerged face deep in mud/quicksand" prompt, is about shoulder/neck deep. I was able do this a few months ago for some other projects, but now, I can't for the life of me get the character that deep. It Does anyone have some prompting advice/tips for accomplish this?

FaceDeepInQS.jpg

Oops.jpg

[cerberus]

Hi Theo,

I think this is more a Stable Diffusion issue than with Perchance, though it may be contributing. Stable Diffusion, like most AIs, has certain safeguards/censorship built in to prevent violent or hateful images; I believe this is, at least, part of the reason that you can't seem to get images of faces framed in mud/quicksand. I've encountered the same problem, both with Perchance and using my local install of SD and haven't found a way around it yet. Generally, sinking seems to stop at about neck depth and it doesn't matter what you put in the prompt, you can't get it deeper. That said, occasionally, SD will generate an image that is deeper off a "normal" prompt, the deepest I've seen so far is up to the nose, but with the head upright. It would seem then that SD applies the rules rigidly to what you ask for but allows a bit of latitude on what it ends up with. The best thing I can suggest is to attempt to give it as much opportunity to bend the rules as possible!

To this end I'd suggest you try framing your images in an appropriate way and leaning on the weighting options. my suggestions of things to try are:

Depth of subject is affected by aspect ratio, so choose square or landscape.
Attempt to fill the image: use "extreme close-up shot", "close-up shot", "long shot" and sometimes "wide-angle shot" works.
Specify perspective and view - "low angle view from side", "high angle view looking down", and other variations (note: with "extreme close-up" you usually seem to get shots from the front).
Use weightings, this can apply to anything in the prompt from the framing instruction to the sinking instructions etc. and experiment with the numbers, some things are affected more that others. Generally things get bizarre if you go higher than 1.9 though.

You'll probably have to generate several images to get near what you want, SD doesn't always respond to these things immediately, but it is meant to be an AI so it is supposed to be capable of learning and, from my experience, the more you do the better it seems to get. Well, up to a point anyway.

You may have noticed I haven't suggested things like adding "neck deep", "up to her ears" to the prompt. I assume you've tried all those already, I know I did and they just get ignored!

Hope this helps.

[cerberus]

I've been experimenting a bit, still haven't found what you want. I've been playing with this prompt:

close-up shot (high angle view:1.2) , attractive woman aged 27, (looking up:1.2), (sinks in mud quicksand until only her face is visible:1.5) on mudflats

Some results are pretty much head only but still not "face framed by the mud" and you get a few ghouls! This is one example I got:

00922-2013704165.png

[bogbud]

AI may have drawn some inspiraton from mud-spas or the like.
Very much the same pose as if she is taking a mudbath in a tub.

[cerberus]

AI may have drawn some inspiraton from mud-spas or the like.
Very much the same pose as if she is taking a mudbath in a tub.

Could be, there's no telling where the source images the checkpoint models were trained on came from. You can't predict what the AI will give you for a specific prompt, derivations can be tenuous prompt terms may be ignored.

[cerberus]

Found this in my backlog, probably near as close as I've managed. The prompt used is below the image. You may have to generate a number of images to get anything like this. Good luck.

00472-1880491047.png

wide-angle shot from below front view looking ahead, beautiful curls blonde haired woman aged 27 in page boy outfit (sinking in quicksand:1.3) in a beach, twisting, turning, pulling, reaching, Horrified, Wide-open eyes, a furrowed brow, and an open mouth, magenta, winter, sunset, (hurricane:1.3), (nipples:0.7)

[cerberus]

This prompt resulted in the attached images (amongst a lot of other "not quite")

red sonja strode forward purposefully onto the wet muddy plain, sinking deeper and deeper in the muddy ground, (quicksand:1.2) mud, she sank (deeper:1.5) and (deeper:1.5), distant gathering strormclouds, partial profile, extreme close-up

00813-4245976998.png

00827-4245977012.png

00874-3158513452.png

[cerberus]

There's this one as well. This is, I think, as deep as I've managed to get anyone. The prompt used:

Catwoman fled across the remote mudflats on a moonlit night falling into quicksand and (((sinking quickly))), profile, averted gaze, muddy, mud spattered, extreme close-up view

00117-2692304214.png

00118-2692304215.png

00127-2692304224.png

As you can tell from the prompts I've used, it is mostly a matter of luck (well, model plus prompt plus format plus seed) and if you generate enough images you'll get some that are like this.

If you've seen any my AI images that you'd be interested to know what I did to get them, please feel free to ask. I'm not at all protective of what I'm doing and would be quite happy to share prompts and settings. I embarked on this as a learning experience and have no issues with sharing my experience and findings with anyone who is interested.

[Theo]

AI may have drawn some inspiraton from mud-spas or the like.
Very much the same pose as if she is taking a mudbath in a tub.

I have tried that before and it does sometimes helps get them neck deep, but not much deeper.

I think this is more a Stable Diffusion issue than with Perchance, though it may be contributing. Stable Diffusion, like most AIs, has certain safeguards/censorship built in to prevent violent or hateful images; I believe this is, at least, part of the reason that you can't seem to get images of faces framed in mud/quicksand. I've encountered the same problem, both with Perchance and using my local install of SD and haven't found a way around it yet. Generally, sinking seems to stop at about neck depth and it doesn't matter what you put in the prompt, you can't get it deeper. That said, occasionally, SD will generate an image that is deeper off a "normal" prompt, the deepest I've seen so far is up to the nose, but with the head upright. It would seem then that SD applies the rules rigidly to what you ask for but allows a bit of latitude on what it ends up with. The best thing I can suggest is to attempt to give it as much opportunity to bend the rules as possible!

I was kinda afraid of that, though I know it will do what I want as I've been able to do it in the past, but it just refuses to cooperate on a set I've been working on. I did try some of your suggestions from yesterday, and right out of the gate I got one that was almost perfect, but for some reason, it was black and white. I added "color photo" after that, and dozens more generations later, I never got anywhere near that again

As you can tell from the prompts I've used, it is mostly a matter of luck (well, model plus prompt plus format plus seed) and if you generate enough images you'll get some that are like this.

Oh yes, trial and error is synonymous with making AI pics. Nevertheless, thanks for your help, I'll keep on trying your suggestions and see if it helps. Maybe one day I'll get lucky

[cerberus]

I was kinda afraid of that, though I know it will do what I want as I've been able to do it in the past, but it just refuses to cooperate on a set I've been working on. I did try some of your suggestions from yesterday, and right out of the gate I got one that was almost perfect, but for some reason, it was black and white. I added "color photo" after that, and dozens more generations later, I never got anywhere near that again

I understand the frustration, I've experienced similar. Adding "looking up" to the framing part of the prompt, will often result in the head being tilted back. Then it's a case of getting the depth covered, I've had some success with having something along the lines of "sinks in quicksand until only the face is visible". Combining those two ideas may give a better chance of hitting what you want.

Oh yes, trial and error is synonymous with making AI pics. Nevertheless, thanks for your help, I'll keep on trying your suggestions and see if it helps. Maybe one day I'll get lucky

It's not really trial and error, it isn't that you are doing anything wrong. SD uses a seed number when generating images (you don't get to see this with Perchance), this is a large number (ten digits or so) and seems to be used to derive the noise pattern from which the resulting image is generated. Normally, a new seed is used for each generate, so you get a different result every time. This means that getting what you want is about pushing the image generation in the direction you want then generating until a good seed gets picked! The closer you can push the image generation prompt the more scope you have for getting a sympathetic seed.