Stable Diffusion AI sneezes

[Demosthenes]

15 hours ago, baron_von_gotta said:

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.

This is a wonderful scenario.  I had an incomplete drabble idea with a very similar concept.  About a priestess or princess who is revered everywhere for the incredible strength and power of her magnificent sneezes.  Every year, thousands make a pilgrimage to witness her yearly ultimate sneeze ceremony, featuring a special sneeze she built up and held back for months in advance.

[Via]

On 12/26/2022 at 12:29 PM, Demosthenes said:

This is a wonderful scenario.  I had an incomplete drabble idea with a very similar concept.  About a priestess or princess who is revered everywhere for the incredible strength and power of her magnificent sneezes.  Every year, thousands make a pilgrimage to witness her yearly ultimate sneeze ceremony, featuring a special sneeze she built up and held back for months in advance.

adore both of these! hoping one of them will come to fruition someday, would love to read (or even collab on) a story in that vein

Edited November 10, 2023 by Via

[baron_von_gotta]

2 hours ago, Via said:

adore both of these! hoping one of them will come to fruition someday, would love to read (or even collab on) a story in that vein

Well, it ain't a story, but here's one more entry in the canon:

[PollenFiend]

So I've been trying out HassanBlend and it's definitely better with more realistic anatomy.  As for closeups, I'm finding it leaning towards a realistic 3D/CGI kind of look but when I try to make it truly photorealistic it takes away a lot of the "sneezy" styling.  Will keep experimenting...

Here is a version that's a bit more CGI-like:
 

 

And when I try making the above more photorealistic, it does look like a real human face but it sacrifices a lot of the sneezy styling:
 

[baron_von_gotta]

16 minutes ago, PollenFiend said:

So I've been trying out HassanBlend and it's definitely better with more realistic anatomy.  As for closeups, I'm finding it leaning towards a realistic 3D/CGI kind of look but when I try to make it truly photorealistic it takes away a lot of the "sneezy" styling.  Will keep experimenting...
  

Personally I think the extra sneeziness of the CGI image is worth the unrealistic effect

[PollenFiend]

12 minutes ago, baron_von_gotta said:

Personally I think the extra sneeziness of the CGI image is worth the unrealistic effect

I agree.  And I like the pic you posted above!  Sorta has a FF-X Yuna vibe going on.  I dig it.

[baron_von_gotta]

Came from a very similar prompt as the priestess one. I've been messing with taking images from reddit and then running img2img a few times. This one had 0 inpainting, just a couple of iterations of img2img.

Positive prompt:

beautiful princess, long blonde ponytail, huge breasts, wide hips, ceremonial floral garb, intricate floral headdress, intricate lacy floral lingerie, about to sneeze, scream sneeze ahegao, terrible allergies, wide open mouth, tongue, unhinged jaw, lips quivering, sneezing ritual, floral altar, flower sneeze ceremony, red nose, eyes slightly open, eyes looking up, faraway stare, face side profile, golden heavenly lights, golden pollen vibes, pollen everywhere in the air
,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

Negative prompt:

(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'm not 100% sure I used the same positive prompt for all of the iterations. Might have fiddled a little.

[MikiSan]

if only I could use img2img 

[baron_von_gotta]

Still playing around with the "floral priestess" vibes. This one feels like it could almost be a real photo:

And on the other side of the coin, bonus bunny lady from when I was on SamDoesSexy still:

[webmeistro]

Nice, what model on the first?

[baron_von_gotta]

7 hours ago, webmeistro said:

Nice, what model on the first?

HassanBlend

[baron_von_gotta]

I found something incredibly useful, it can plot X/Y for your prompts.

I used the standard "high quality/resolution" positive/negative prompts, along with "beautiful woman, riding the bus, about to sneeze, scream sneeze ahegao". Then I chose a seed that I liked, and asked it to do an X/Y plot where it varied "ahegao" to "distressed", "irritated", and "angry" versus step count. The results were surprising to me, ahegao produces kind of a unique result compared to those other words:

You can read about the feature here: https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#xy-plot

I'll be posting some more plots soon because it feels like a pretty useful thing to share.

[PollenFiend]

16 minutes ago, baron_von_gotta said:

I found something incredibly useful, it can plot X/Y for your prompts.

I used the standard "high quality/resolution" positive/negative prompts, along with "beautiful woman, riding the bus, about to sneeze, scream sneeze ahegao". Then I chose a seed that I liked, and asked it to do an X/Y plot where it varied "ahegao" to "distressed", "irritated", and "angry" versus step count. The results were surprising to me, ahegao produces kind of a unique result compared to those other words:

 

Oooohhh nice!  I'm definitely going to experiment with that!  

In the meantime I got better with HassanBlend, trying to do a realistic mid-sneeze face (which seems a lot harder for the AI to manage) vs pre-sneeze (a lot easier, I guess since it's very visually similar to just a normal sexy-face).  

Also trying to get better with sneeze spray (sorry for the mess! This one turned out more like legit allergy snot vs spray)...

So here is another scene following the ancient China theme.  Enjoy.

 

[baron_von_gotta]

Follow up experiments where I varied "scream":

And then fixed it at 20 steps and ran the whole X/Y for both words:

 

What do you think? I think I might start asking it for distressed or irritated yawn sneeze, and leave ahegao out from now on. 

Also here is the last experiment where I removed "about to sneeze" from the prompt, but kept the same variations:

[baron_von_gotta]

1 minute ago, PollenFiend said:

In the meantime I got better with HassanBlend, trying to do a realistic mid-sneeze face (which seems a lot harder for the AI to manage) vs pre-sneeze (a lot easier, I guess since it's very visually similar to just a normal sexy-face).  

That's awesome! Mid-sneeze is really hard I agree. 

[PollenFiend]

4 minutes ago, baron_von_gotta said:

That's awesome! Mid-sneeze is really hard I agree. 

Thanks!  I think we're both gradually getting better at this.  Really amazing work so far.

[SneezeLucario]

On 12/24/2022 at 11:21 PM, PollenFiend said:

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

I recognize this story, this was great. Very good surprise to see you coming back😊.

[PollenFiend]

5 hours ago, baron_von_gotta said:

What do you think? I think I might start asking it for distressed or irritated yawn sneeze, and leave ahegao out from now on. 

Depends on where in the sneeze buildup you want your scene.  The pre-sneeze progression appears to be something like: distressed-allergic -> irritated-scream -> angry-yawn.

Then you can do the mid-sneeze.  For my experiment above the face was accomplished with the a base prompt containing "sneezing harshly, mouth closed, angry" which got the furrowed eyebrows and general appearance.  Then lots of inpainting iterations particularly with the mouth which lucked out with a very simple prompt (to reduce AI confusion) with "sneezing harshly, sneezing, sneezes".

 

[baron_von_gotta]

9 hours ago, PollenFiend said:

Depends on where in the sneeze buildup you want your scene.  The pre-sneeze progression appears to be something like: distressed-allergic -> irritated-scream -> angry-yawn.

Then you can do the mid-sneeze.  For my experiment above the face was accomplished with the a base prompt containing "sneezing harshly, mouth closed, angry" which got the furrowed eyebrows and general appearance.  Then lots of inpainting iterations particularly with the mouth which lucked out with a very simple prompt (to reduce AI confusion) with "sneezing harshly, sneezing, sneezes".

 

How do you add the spray? I know my way around GIMP enough to follow a youtube tutorial, but I couldn't find a good one.

[PollenFiend]

1 hour ago, baron_von_gotta said:

How do you add the spray? I know my way around GIMP enough to follow a youtube tutorial, but I couldn't find a good one.

GIMP is good, especially with it being free and all.  That's what I use too.  I've found that the airbrush produces many small droplets that the AI will interpret as spray or mist, and drawing in short or long lines will produce snotty "ribbons" or "strings".  Color can be white or slightly off-white e.g. a bit pink or blue depending on the scene.  Here are the GIMP edits I performed to produce the spray in three of my pictures.  For the prompt, inpaint using keywords like "water droplets, watery spraying, water shooting from her mouth" etc.

Light spray:
 

Messy spray:
 

Strings:
 

 

After I make the edit, I export as a new PNG file with no compression:

[PollenFiend]

I've figured out a better method for resizing to hi-res that works for me.  I typically start at 512x512 then when I'm 80% happy with the product I'll upscale to 768x768 in img2img or inpainting with the Resize and Fill option.  When I'm 99% happy, I'll upscale again to the max my PC can handle, 900x900 in a small batch to choose the best final version.  That's what I've been posting so far.

For resizing to hi-res, I wasn't happy with ESRGAN alone.  It took some elements from the background and brought them into sharp focus mistakenly:
 

I tried 4x-UltraSharp but it's not as good as ESRGAN for elements that should be sharply in focus, like the face:
 

So what now works for me is:
Resize to: 2160x2160
Upscale: 2.411, visibility: 1.0, model:R-ESRGAN 4x+
Upscale: 2.411, visibility: 1, model:4x-UltraSharp

In the original design the AI for some reason made the teeth kinda yellow so I also whitened the teeth in GIMP using the guide from https://thegimptutorials.com/how-to-whiten-teeth/

Now here is the hi-res, touched-up picture in glorious 2160x2160:
  

[baron_von_gotta]

36 minutes ago, PollenFiend said:

Now here is the hi-res, touched-up picture in glorious 2160x2160:
   

Hell yeah, awesome result. I'll have to try some of that upscaling. And thanks for the tips regarding the spray. So, just to be clear, you draw in the spray yourself and then ask the AI to inpaint over it, in order to make it look more natural? That I am very capable of, it's the "making it look like the rest of the image" part that I suck at 🤣

[PollenFiend]

53 minutes ago, baron_von_gotta said:

Hell yeah, awesome result. I'll have to try some of that upscaling. And thanks for the tips regarding the spray. So, just to be clear, you draw in the spray yourself and then ask the AI to inpaint over it, in order to make it look more natural? That I am very capable of, it's the "making it look like the rest of the image" part that I suck at 🤣

Yes, I very crudely draw in the spray like the above sample pics then inpaint. Typically a large batch with very little denoising, choose the best one, send that back to inpaint and do another iteration maybe dialing down the denoising even more. After 3-4 iterations it usually gets to something decent. If not, I'll start over again drawing in GIMP.

[baron_von_gotta]

24 minutes ago, PollenFiend said:

Yes, I very crudely draw in the spray like the above sample pics then inpaint. Typically a large batch with very little denoising, choose the best one, send that back to inpaint and do another iteration maybe dialing down the denoising even more. After 3-4 iterations it usually gets to something decent. If not, I'll start over again drawing in GIMP.

Nice, thanks! The potential applications of this are pretty interesting. I'm already thinking about adding a feather duster in this manner, because that's another item that the AI doesn't really seem to understand, at least not well.

Here's one that I made with spray, aiming for a mid-sneeze vibe:

And the original, before I added the spray or fixed her glasses, upscaled without edits:

[PollenFiend]

1 hour ago, baron_von_gotta said:

Nice, thanks! The potential applications of this are pretty interesting. I'm already thinking about adding a feather duster in this manner, because that's another item that the AI doesn't really seem to understand, at least not well.

Here's one that I made with spray, aiming for a mid-sneeze vibe:

There ya go! Nice work!

And yes, I've found the crude image editing to be helpful to force the AI into a direction that it can't get to on its own. Eg in my above image I drew some thin curvy black lines as strands of hair stuck to her teary eyes and cheek. Inpainting a few iterations did the rest.