Stable Diffusion AI sneezes

[baron_von_gotta]

It turns out the Stable Diffusion AI is pretty good at generating sneezing-related content. With some limitations which I'll get into, but first here's my favorites so far:

Anime:

Not Anime:

I don't even know:

 

Compared to midjourney and other services, I think stable diffusion gives more opportunities for tweaking your output. The huge upside/downside is that you run it locally, which means privacy, but it also means it's limited by your GPU.

If you want to try to install it, I followed this guide and then made a few adjustments afterwards: https://www.howtogeek.com/832491/how-to-run-stable-diffusion-locally-with-a-gui-on-windows/

Another big upside of stable diffusion is the ability to "inpaint" which is where you erase part of an existing image and get the AI to fill in something there. I used that in making all of the images above. I started with some prompt, sometimes I would ask it for a woman about to sneeze and add details to that, other times I'd start with the woman and then try to get it to add a sneezy face. It's surprisingly good at understanding what "about to sneeze" means, which is the main thing I've been putting in.

 

Current limitations/considerations:

It's a little bad at hands/mouths. 

It's definitely really bad at trying to show someone pinching their nose or touching it in any way. I've stopped trying

It doesn't seem to really understand what tickling a nose with an object would look like. The closest I've been able to come is holding flowers.

It does understand female celebrities, but it's not any good at changing their face to look sneezy. When I use inpainting to add a sneezy face, the AI just kinda plops a new face on there, which is fine for an anime woman but not really what you're looking for from the celebrity photo.

It can generate nudity but I've obscured any with clothes to make it appropriate for this forum, the AI is quite good at that.

I generally put in "about to sneeze" because that's what I like, but also because to the extent that the AI understands "sneezing", it doesn't really produce something too different from "about to sneeze". Although it did just make this directly from the prompt "beautiful anime woman sneezing":

But it also made this from the same prompt:

 

So far, I've only asked for women sneezing because that's what I like. That said, I'm open to any (forum appropriate) requests, it doesn't take me much effort to put in someone else's direct prompt and I'm curious what people come up with. For men sneezing, or other things which don't pique my interest, I probably won't spend a whole lot of time tweaking it, especially because I wouldn't know what to tweak. But I'll happily type what you give me into the program. And who knows, maybe I'll get inspired

[webmeistro]

If you want to try Stable Diffusion without installing anything, go here: https://huggingface.co/spaces/stabilityai/stable-diffusion

If you want to install it locally, that guide is a bit out of date. Your best bet is the Stable Diffusion Reddit.

[baron_von_gotta]

Nice, good call on the reddit, I already learned a lot of great tips from https://www.reddit.com/r/StableDiffusion/comments/x41n87/how_to_get_images_that_dont_suck_a/

 

[baron_von_gotta]

My favorite so far. I did prompts like "african priestess intricate ceremonial floral garb, about to sneeze" to generate most parts of it

 

[baron_von_gotta]

Playing around with adding text to the images afterwards. This is my first attempt:

And second attempt:

Imgur album containing these images, the versions without text added, and also two random ones that are pretty good but not quite "add-text-worthy"

https://imgur.com/a/HbjhrhS

[Maximusace9]

What are the full versions of your usual prompts?

[MikiSan]

7 minutes ago, Maximusace9 said:

What are the full versions of your usual prompts?

I need to know as well

[baron_von_gotta]

1 hour ago, Maximusace9 said:

What are the full versions of your usual prompts?

1 hour ago, MikiSan said:

I need to know as well

My strategy changed massively after reading that reddit guide I posted. And a lot of the below is straight up lifted from there, so give it a read for sure. Also, I use the latest version of "SamDoesSexyBlend" as my model instead of the base Stable Diffusion.

The basic strategy is that I run 16 batch size 8 step DDIM with some particular seed, and then if I see something I like, I do the same prompt, same seed, 1 batch size 50 step LMS. Then I repeat the same process with inpainting several times to get particular details. Since it's repetition and zooming in, I don't use one universal prompt to get the image. 

To give an example of what I mean, I'll post all of the steps along the way of generating the redhead I posted earlier. 

1. "closeup of a beautiful woman's face, curly red hair, tall open mouth, yawning, screaming, sneezing, surprised" (16 batch size 8 step DDIM). That produced this (along with 15 other results that I didn't like as much to proceed with them):

2. "closeup of a beautiful woman's face, curly red hair, tall open mouth, yawning, screaming, sneezing, surprised" (1 batch size 50 step LMS):B

3. That produced something close to the final product, but I didn't like some of the details, so I fixed them with inpainting. First I selected her nose (which means that it will leave all of the image except for her nose alone. Then I changed my prompt to: "closeup of a beautiful woman's face, curly red hair, red nose, flaring nose, huge nose, wide nose, fat nose, gotta sneeze, terrible allergies" (16 batch size 8 step DDIM). Along with 15 results I didn't choose, that produced:

4. Then I tried the LMS version, didn't like it, so I decided to leave her nose alone. Next I wanted to fix the weirdness in her moth, so I selected her mouth, and changed my prompt to "closeup of a beautiful woman's face, curly red hair, sneeze scream yawn surprise, open mouth". For some reason I didn't do DDIM 8 this time, just decided to roll the dice with LMS 50, and I got:

5. Then I selected her eyes and changed my prompt to "closeup of a beautiful woman's face, curly red hair, scream sneeze yawn surprise, about to sneeze, closed eyes, squinting eyes, fluttering eyelashes" (16 DDIM 8 step) My favorite was: 

6. Then just LMS 50 with the same prompt, and upscale it to 4x using R-ESRGAN produced the final image.

Hopefully that helps! Let me know if you have any other questions. I'd love to see what other people are able to do with this.

[MikiSan]

39 minutes ago, baron_von_gotta said:

My strategy changed massively after reading that reddit guide I posted. And a lot of the below is straight up lifted from there, so give it a read for sure. Also, I use the latest version of "SamDoesSexyBlend" as my model instead of the base Stable Diffusion.

The basic strategy is that I run 16 batch size 8 step DDIM with some particular seed, and then if I see something I like, I do the same prompt, same seed, 1 batch size 50 step LMS. Then I repeat the same process with inpainting several times to get particular details. Since it's repetition and zooming in, I don't use one universal prompt to get the image. 

To give an example of what I mean, I'll post all of the steps along the way of generating the redhead I posted earlier. 

1. "closeup of a beautiful woman's face, curly red hair, tall open mouth, yawning, screaming, sneezing, surprised" (16 batch size 8 step DDIM). That produced this (along with 15 other results that I didn't like as much to proceed with them):

2. "closeup of a beautiful woman's face, curly red hair, tall open mouth, yawning, screaming, sneezing, surprised" (1 batch size 50 step LMS):B

3. That produced something close to the final product, but I didn't like some of the details, so I fixed them with inpainting. First I selected her nose (which means that it will leave all of the image except for her nose alone. Then I changed my prompt to: "closeup of a beautiful woman's face, curly red hair, red nose, flaring nose, huge nose, wide nose, fat nose, gotta sneeze, terrible allergies" (16 batch size 8 step DDIM). Along with 15 results I didn't choose, that produced:

4. Then I tried the LMS version, didn't like it, so I decided to leave her nose alone. Next I wanted to fix the weirdness in her moth, so I selected her mouth, and changed my prompt to "closeup of a beautiful woman's face, curly red hair, sneeze scream yawn surprise, open mouth". For some reason I didn't do DDIM 8 this time, just decided to roll the dice with LMS 50, and I got:

5. Then I selected her eyes and changed my prompt to "closeup of a beautiful woman's face, curly red hair, scream sneeze yawn surprise, about to sneeze, closed eyes, squinting eyes, fluttering eyelashes" (16 DDIM 8 step) My favorite was: 

6. Then just LMS 50 with the same prompt, and upscale it to 4x using R-ESRGAN produced the final image.

Hopefully that helps! Let me know if you have any other questions. I'd love to see what other people are able to do with this.

oooo im getting better results now thanks for telling us 

[baron_von_gotta]

This one's a little rough around the edges and I couldn't improve it, but it's still too good not to post

[MikiSan]

though i dont know how to use inpaint though...

[webmeistro]

2 hours ago, MikiSan said:

though i dont know how to use inpaint though...

Assuming you're using the AUTOMATIC WebGUI:

Use the 1.5 inpainting model

Load a picture into the img2img tab

Use the drawing tool to select the area you want inpainted

Enter your prompt and go!

[PapaBear70]

2 hours ago, baron_von_gotta said:

This one's a little rough around the edges and I couldn't improve it, but it's still too good not to post

Rough or not, this appears to be a massive sneeze in the making!

[Chiguy]

Love these!

Really nice work. 

[MikiSan]

1 hour ago, webmeistro said:

Assuming you're using the AUTOMATIC WebGUI:

Use the 1.5 inpainting model

Load a picture into the img2img tab

Use the drawing tool to select the area you want inpainted

Enter your prompt and go!

I've been using the web version since i dont know how to use git 

[webmeistro]

Yeeeeeeeah, it's not easy.

Anyway, no inpainting on the web version, sorry.

[MikiSan]

3 minutes ago, webmeistro said:

Yeeeeeeeah, it's not easy.

Anyway, no inpainting on the web version, sorry.

awww

[webmeistro]

[MikiSan]

really want to play with inpaint 

[baron_von_gotta]

 

2 minutes ago, MikiSan said:

really want to play with inpaint 

You can inpaint with the free version of DALLE 2 https://openai.com/dall-e-2/

Sadly you have a limited number of credits/month and I can't vouch for how good it is at making sneezing

[webmeistro]

You can do awesome stuff without inpaint. Like that one there. I only use it once in a while for specific stuff.

[baron_von_gotta]

10 minutes ago, webmeistro said:

Kudos on being the first one to post images here besides me, I love it. Makes me want to try some prompts with "medieval peasant girl" or maybe "tavern barkeep"

[MikiSan]

can someone help me install it i cant get it to run

[baron_von_gotta]

Be sure not to walk by any bright windows while you're dusting the house, you just might lose control and throw out your back. Not to mention dropping your feather duster, that can't help the situation.

Version without text:

[baron_von_gotta]

3 minutes ago, MikiSan said:

can someone help me install it i cant get it to run

What's happening when you try?