Stable Diffusion AI sneezes

[baron_von_gotta]

Just a quick sneeze break on the jog home. It's all these wildflowers, it can't be helped. But two or three sneezes ought to do the trick. Okay, no more than six. Just one more - or perhaps three more. It still tickles! There's no choice, one last sneeze, better make it count! ..... Perhaps just one more?

[PollenFiend]

Hi everyone, I've been away for a long time, but this post has made me come back... To the OP - I have to say THANK YOU for the guide and the inspiration.  I've been playing with the tech a LOT and am starting to get fantastic results.  I can trade tips after I've learned and experimented some more.

I left abruptly years ago to deal with some personal challenges in my life at the time - nearly separating from my partner (we're good now), work troubles, parents... etc.  To those of you who I may have wronged or disappointed, I'm sincerely sorry.

To try to make up for past wrongs and lost time, I'm posting three images I made after iterating a TON with Stable Diffusion.  They are inspired by my past stories!  See if you can identify the scene in each.  I even put in captions.

Again - thank you so, so much for the tech tips.  I'm really digging it and will do more.

https://imgur.com/xJDenpE (captions)

https://imgur.com/iY2d7jU (captions)

https://imgur.com/NNh1zKO (captions)

Edited December 27, 2022 by PollenFiend

[webmeistro]

Awesome!!

[baron_von_gotta]

15 hours ago, PollenFiend said:

Hi everyone, I've been away for a long time, but this post has made me come back... To the OP - I have to say THANK YOU for the guide and the inspiration.  I've been playing with the tech a LOT and am starting to get fantastic results.  I can trade tips after I've learned and experimented some more.

Those are awesome! Looking forward to hearing about tips. What model are you using?

[baron_von_gotta]

Here's a cute one that I made with absolutely 0 inpainting. I'm trying to figure out how to get it to generate sneezing into a tissue/handkerchief. It seems like half the time the AI decides to paint their face with tissue/handkerchief pattern, almost like a weird version of a face mask. 

Also here's a random one that I made but didn't feel proud enough (at the time) to post. I've got a few of these in the can from when I was learning it, so I'll try to include one on my next few posts for a little extra fun:

[PollenFiend]

Here's another one to go along with my previous stories.  I rather like how elegant (and allergic) this one feels.

https://imgur.com/CkcikEa (captions)

 

I will post again later to describe my workflow and a few tips I've learned...  Basically lots of iterations to reinforce the good parts, in-painting, a little image editing to force stuff that the AI can't understand (e.g. sneeze spray), choosing the right keywords in the prompts regarding the style you want (not just describing the action/thing you want to see).  Really having way too much fun with this.

Edited December 27, 2022 by PollenFiend

[PollenFiend]

2 hours ago, baron_von_gotta said:

Also here's a random one that I made but didn't feel proud enough (at the time) to post. I've got a few of these in the can from when I was learning it, so I'll try to include one on my next few posts for a little extra fun:

Ooooo!  I like the cowgirl theme.  Have you been working on improving this one more?

[baron_von_gotta]

2 minutes ago, PollenFiend said:

Here's another one to go along with my previous stories.  I rather like how elegant (and allergic) this one feels.

She's beautiful, I really like the clothes too.

 

1 minute ago, PollenFiend said:

Ooooo!  I like the cowgirl theme.  Have you been working on improving this one more?

Not really, but I might give another crack at it now that I know there's an audience. My least favorite thing about it is the unrealistic proportions. I've been using someone's custom checkpoint to make most of these images, I like the cartoon feel but the body proportions are simply unrealistic. That's why I'm curious about what model/checkpoint you used, is it the base stable diffusion?

[baron_von_gotta]

I set up the novelai leak locally. https://rentry.org/sdg_FAQ

Getting pretty nice results so far (albeit anime-style). I'll post once I have something worth posting.

[baron_von_gotta]

Merry Christmas, from a naughty elf with a christmas tree allergy:

And here's a bonus stocking stuffer that I didn't want to post when I made it for whatever reason:

[Chronosynclastic]

5 hours ago, PollenFiend said:

https://imgur.com/Ejvrml5

I love this one. The sneeze spray is really great.

[baron_von_gotta]

I can't describe exactly what's going on here. I wish I could write a story for this. I have a vague fantasy about some sort of religious ritual that intentionally ends in a giant sneeze. Just trying to hit that vibe.

And if that's too weird for you here's a sneezy librarian:

[baron_von_gotta]

One thing I've recently been employing is using the standard NovelAI negative prompt list with some slight modifications. Seems to help a lot on quality, even if not using the NovelAI model. Here's my list of negative prompts: "lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, black and white, greyscale"

[Via]

i'm pretty leery of AI art as a general rule tbh, but 'ultra niche kink art' is one of the few angles i can see a use for it! and that being said holy shit wow this thread is spectacular

1 hour ago, baron_von_gotta said:

I can't describe exactly what's going on here. I wish I could write a story for this. I have a vague fantasy about some sort of religious ritual that intentionally ends in a giant sneeze. Just trying to hit that vibe.

oh i LOVE this as a scenario idea; i know you said it was tricky to describe but if there were any further details you did have in mind i'd be v curious to know!

[baron_von_gotta]

6 minutes ago, Via said:

oh i LOVE this as a scenario idea; i know you said it was tricky to describe but if there were any further details you did have in mind i'd be v curious to know!

I imagine early dark ages in some remote place, a certain type of flower grows that embodies a pagan goddess of sneezing and allergies. The flowers' pollen is mystically torturous, very tickly but also very difficult to trigger a sneeze. When the sneeze finally does come, it's huge and extremely physically satisfying.

I imagine priestesses trained to handle the flowers, both on a gardening level but also how to handle themselves when dealing with the pollen. How to channel the goddess's energy into massive, satisfying sneezes.

At some point I want to make a CK3 mod about all of this... I have it started, but modding is just so much more work than typing stuff into an AI image generator 😝

[baron_von_gotta]

Some random prompt advice to play around with:

"grotesque nose" - at least for the models I'm using that are engineered to make beautiful women, "grotesque nose" just means anything other than a dainty little button nose. It doesn't produce something truly grotesque, if it does for your model, then maybe try a less extreme adjective.

"angry/irritated scream" - similar to a sneeze, at least based on how the computer sees things

"sneeze/sneezing fetish" - I've stopped putting this in my prompts, but it seems to focus in on the sneeze a little bit more. It's not out of the question that the models were trained on art that was tagged with "sneeze fetish". The reason I've stopped putting it in my prompts is mostly because of the next one:

"ahegao" - The AI understands this quite well in my experience. It's similar to a sneeze face. And anything too weird can be toned down with inpainting. 

"pollen vibes" - typically if I want to show an irritant in the air, I type "<irritant> everywhere, <irritant> vibes". Seems to work best with pollen though.

[PollenFiend]

<<edited below>>

Edited December 26, 2022 by PollenFiend

[PollenFiend]

Hi again.  So here are some tips that have worked for me, described along with sample images from my workflow in creating the picture of the redheaded hostess.

Since I started off with the How-To guide linked at the beginning of this thread, I'm still using the SD 1.4 model though I'll probably upgrade soon.  Checkpoint is SamDoesSexyBlend though I will start experimenting with HassanBlend.

Prompts for the redhead hostess image:

POSITIVE=
sexy elegant freckled redhead, ((about to sneeze)), very bright red nose, suffering from allergies, can't stop sneezing, green eyes are closed, pale skin, wearing a classy red dress and gold jewelry,  yellow pollen dust floating everywhere, in a nightclub surrounded by flowers, pink neon lights, led lights

Amy Adams, soft lighting, beautiful detailed neon lights, reflections, masterpiece, best quality, highest quality, extremely detailed, 8k wallpaper, focused, 4k wallpaper, ultra realistic, photorealistic, sharp focus, absurdres, (HDR:1.5), trending on artstation, high contrast, photograph, insanely detailed and intricate, ((by Daniela Uhlig))

NEGATIVE=
(CG),black and white,border, frame, lowres, jpeg artifacts, low quality, normal quality, worst quality, poorly drawn, error. abstract, asymmetrical, blurry, cropped, disconnected, duplicate, extra, missing, signature, text, username, watermark. amputee, bad anatomy, deformed, disfigured, disproportionate, eye bags, fused, malformed, morbid, mutated, mutation, mutilated, hands, fingers,bad hand,extra limbs,missing limb,extra leg,missing leg,missing hand,extra hand

 

I found it very helpful to put in descriptions of the art itself, not just the thing you want to see.  E.g. high-res, detailed, and following the style of an artist you want the AI to emulate.  I also put in the name of a model/actress/person for the AI to get inspiration from, which I found useful for tricky facial expressions.  The negative prompt is also super important.

 

Then I do a big batch (say 50-100) for rapid generation of many different poses, angles, interpretations to choose from:

Steps: 8, Sampler: DDIM, CFG scale: 15, Size: 512x512

For the hostess picture, that produce this decent shot to serve as a basis for further iteration:
https://imgur.com/QSSAl9V

 

I then do img2img for a few iterations to reinforce the shot overall, usually with a few extra steps per iteration, say 15-30.  After +3 iterations, the AI got to this:
https://imgur.com/eSqypjQ

 

Now on to inpainting.  I wasn't happy with the background, and I wanted more of an underground lounge/nightclub feel so I iterated that a few more times using the same method above but just inpainting the background, changing the prompt to focus solely on the environment and lighting (nightclub, neon lights, pink lights, etc.), and each time it gets closer to what I want, I dial down the denoising:
https://imgur.com/PdYVMt9

 

Did the same method for the jewellery:
https://imgur.com/OcQe5UK

 

And the facial expressions, inpainting MANY iterations focusing solely on the mouth, then nose, then eyes.  Each time changing the prompt to focus on those areas, e.g. "mouth partly open", "bright red nose", "furrowed eyebrows".  Each batch I do something like 20ish, high number of steps (50-80), and choosing the best version to iterate and reinforce further:
https://imgur.com/tZiejdi

 

If the AI has trouble getting the effect I want (usually I notice this for things like a sneeze spray or a really red nose particularly if I am having the AI make a "realistic" picture), I use an image editor to crudely force the effect I want and then have the AI clean it up.  Here is my forced image editing:
https://imgur.com/kte1s3Y

And the AI figuring out I wanted a really red nose:
https://imgur.com/EfetFf1

 

Then it's just more inpainting and image editing for things the AI is having trouble with, in this case, pollen specks where I wanted them:
https://imgur.com/5ryCOHq

 

After a few more iterations and things are looking near final, I'll typically resize to the max my PC can handle (900x900ish):
https://imgur.com/1P85H5w

 

Iterate/reinforce a bit more, then it's done!
https://imgur.com/Ne9NFbs

And I like to add captions, since for now my pictures are scenes from my past stories.
https://imgur.com/xJDenpE

 

 

Edited December 26, 2022 by PollenFiend

[baron_von_gotta]

2 minutes ago, PollenFiend said:

Hi again.  So here are some tips that have worked for me, described along with sample images from my workflow in creating the picture of the redheaded hostess.

 

Woohoo!! Neat tips! Interesting that you use such a high CFG value. I've been using the default. Probably mostly because I'm not sure what I want and want the AI to show me some examples.

I'm going to have to play around with some of the "high res" stuff.

[PollenFiend]

2 minutes ago, baron_von_gotta said:

 

Woohoo!! Neat tips! Interesting that you use such a high CFG value. I've been using the default. Probably mostly because I'm not sure what I want and want the AI to show me some examples.

I'm going to have to play around with some of the "high res" stuff.

Ah right, i forgot about that.  I typically start my CFG at 15 since I'm confident in my prompts now.  And I dial it up to 20-25 as I simultaneously dial down the denoising (0.02-0.05) when it's very close to what I have in mind.  But I've also found that if CFG is too high for too many iterations, the AI brain explodes and turns everything back into noise.

Edited December 26, 2022 by PollenFiend

[aridynamic87]

This is rather fascinating to see. I've always wanted to try this, but I can't say I'm particularly good at setting this all up, haha. The results here are looking fantastic! 😍

[baron_von_gotta]

I made this relatively quickly (i.e., not much obsessive inpainting/iterating) using some of the prompting tips from PollenFiend. Thanks again!! 

I also started looking into the HassanBlend and I found something interesting in his guide:

"I mostly will use Euler A for a fast test, then I switch to DDIM, HEUN or DPM2 Karras". I'm going to start messing around with his checkpoint and also with the Euler A as an initial algorithm. 

Bonus catgirl from when I was still on the NovelAI leak. I think this is my last unposted image from that period, at least that's appropriate to be posted on the non-adult board.

[baron_von_gotta]

This was created using HassanBlend on basically the same mechanic/garage themed prompts, and similar methodology as the one above:

I LOVE that it's managed to create a women of reasonable body proportions instead of a barbie doll. I do like the cartoon style of SamDoesSexy but the waists are disgustingly skinny.

What I don't like about HassanBlend, and every other "realistic sexy" model I've tried, is that it tends towards nudity, even if you explicitly ban that in the prompt. I mean, it makes sense that the NSFW models would be trained on nude images, but I can't post them unless there's clothes lol.

[PollenFiend]

Looking good, dude! Glad my tips were helpful. I'll post more as I learn more as well. Loving this thread and the collaboration.

[baron_von_gotta]

43 minutes ago, PollenFiend said:

Looking good, dude! Glad my tips were helpful. I'll post more as I learn more as well. Loving this thread and the collaboration.

Thank you! Me too!

Here's another HassanBlend image, more obsessively iterated/inpainted. I think I'm going to keep playing with this model, and leave in the "realistic, high quality" prompt fluff. The results are pretty nice.

Note that part of the inpainting for this image involved "grotesque nose" and "unhinged jaw", two things that would be horrifying to see in a real sneeze, but here just allow me to pull the AI away from its natural tendency to make smooth, beautiful features. I feel like a good sneeze face is a disruption in beautiful smoothness, a temporary lapse of poise and grace.