🎄 AI ART: Technique to make AI photos more realistic (Patreon)

🐇 Hugh Hefner doesn't want you to see this! 🐇

You'll need to check my other AI ART posts before this if want to get started with Stable Diffusion!

For some very nice insight on prompts I recommend this article https://prompthero.com/stable-diffusion-prompt-guide .

MY PROCESS TO GET NICE ARTISTIC PHOTOS

Normally I take prompt inspiration from example prompts I like from https://lexica.art/ or https://prompthero.com/ . This kind of photorealism is very new and you don't see it there a lot (yet!) but there are many cool prompts nonetheless that work on any AI model. For this example I used a prompt somebody here sent me and I'll walk you through the process of how I'd try to make it a bit better. 

To work with the prompts I turn the Seed random toggle to Off. That way each time you run an image, it will be the same "randomness" and you can test that way how the prompts influence what the AI does. From there you just steer it in the direction you want little by little.

Ignore the noisy images below, I was just stopping it mid-rendering, 2 photos were enough for testing. I didn't use any input image, just the text.

INITIAL PROMPT

Started with a very nice prompt with good results on the HassanBlend v1.4 AI model (https://www.patreon.com/sd_hassan)

1. ADDED "PHOTOGRAPH OF" AT THE BEGINNING

AI cares about the order of words, it gives more importance to what you say to it first. It's also easier for it to build the scene if you go from big details towards small details, just like you would explain it to a person. Instead of adding ",drawing" or ",photograph" at the end of the prompt, you'll get better results if you start with it, like "A drawing of a cat" or "A photograph of a".

In the example the images got a lot crisper just from that alone

2. ADDED "8K"

Adding 4k, 8k to the prompt will make the AI 'think' more of the images it learned as being tagged with 4k,8k and those images are of great quality (usually). Adding it made the images even more detailed.

You can see that the AI will go in different directions based on the prompts. It's not like a filter thing where it stays on the same image. Adding 8k made it think more of those kind of images. For the 2nd image it stayed in the same area, for the 3rd it went to a different pose. That's because there are many more 8k images that have that kind of angle rather than the previous one, so for the AI this looked like a better match.

3. ADDED "SHARP FOCUS"

This made the AI 'think' more of images tagged like that which are very sharp, so it made the results more detailed and sharper

4. CHANGED "PHOTOGRAPH OF" WITH "PHOTO OF"

The AI doesn't know what the words actually mean, it only knows of what kind of images were tagged like that from what it learned. "Photograph" is something you'd see more in regard to artistic shots. "Photo" is more common and I figured the AI will have more things to draw inspiration from for sexy shots

The quality got a little better. Also gooier. I think all the goo is because of the hentai & adult content blending in the AI's 'thoughts'. For those prompts (poses + wet hair + nice ass) it has also seen lots of hentai & shower/bath photos and it draws inspiration from those too. I think.

5. ADDED "WET SKIN" AND "PHOTOREALISM"

"Photorealism" made the quality a bit better. If it's thinking of things that are not real photos, this makes it put more emphasis on things it learned as being tagged like that, which are usally higher quality, and prevents it to draw much inspiration from lower quality images.

"Wet skin" made the skin look a little better. But the AI started thinking even more of goo! This is again probably because the "wet hair" and "wet skin" prompt makes it think also about shower soapy shots

6. ADDED "F/2.8, 1m away" AT THE END

It's a camera focus/aperture thing, the AI will think more of photos that were tagged like that. And those kind of camera settings to my understanding do: better colors, sharper closeup focus and blurry depth-of-field background to make the person pop more

7. ADDED "NUDE, EROTIC"

The nipples were not looking so good and that's because the AI didn't see too many nipples in photos that were tagged as the previous prompts. Adding nude or erotic makes the AI think more of those kinds of photos too.

While it does help with nipples & gens, it draws the quality of the images a bit down as you can see. The colors got less saturated, more "normal", because most of the nude photos the AI has seen were normal shots not artistic vivid-colors shots. It also started doing weird anatomy in the 2nd shot. I think it's because the nude shots have so much skin in so many different positions that the AI gets confused sometimes. Rendering smaller images or running them for longer usually fix some of these issues

8. HELPING THE AI

I noticed the AI was thinking a lot of nature. That's probably from the "realistic photography" + "wet hair" that were in the original prompt, "photography" is something you'd tag artistic shots with. Most artistic photos of women with wet hair you'd probably see in nature, around beaches, lakes, rivers so the AI goes that way.  The AI however is not 100% certain it's a nature shot so it draws insipration from lots of other types of photos, hence the goo/soap which doesn't make sense in this kind of shot. So I added "posing in nature" to let the AI know what's going on in the scene and draw inspiration more from those kind of photos. This made the AI think less of soap. I probably could've just added "soap" or "shower" as negative prompts and fix it that way too.

9. FINALLY ADDED MAKEUP IN WET CIRCUMSTANCES BECAUSE HOT!

I also added "gorgeous makeup, ((lipstick))". The paranthesis if you missed it in my previous post are to increase the AI's attention to that. You can also type in the weight:
"whatever" = "(whatever:1)" - default weight of 1
"(whatever)" = "(whatever:1.1)" - makes the AI think about it more (+10%)
"(((((whatever)))))" = "(whatever:1.5)" - makes the AI think about it a lot more (+50%)

you can also make the AI think about it just a bit like "(whatever:0.5)" which is -50%

The AI changed the 2nd shot to a frontal because photos tagged with makeup are rarely back shots. The new words made the AI think more of frontals and less about back shots, so it started seeing that particular randomness in that direction.

That's kind of the trick to understanding how it works. You might give it something very specific to do like "makeup" but the AI will not just magically add it anywhere. It will just 'think' more of photos of women that were tagged with makeup from what it learned. Faces will get more makeup but there are consequences too:
- most makeup photos are closeups so the AI will pull more towards closeup shots
- most makeup photos are frontals so the AI will avoid other angles a little more
- most makeup photos are not erotic in nature so the AI will go a bit more towards glamour

Every prompt has consequences like that and often copy-pasting too many things might do more harm than good. It's better to keep it simple and try to steer it in the right direction for your particular kind of effect/shots like I did above.

I did the images at 448x640 with only 25 iterations. For the final two I just hit "Upscale" and got these

With only the text prompt this is now an infinite collection of hot wet babes with makeup! But you can also use it as a "filter" for any other photo, like vam screenshots, and the AI will adapt that photo using this new style. A few examples of what it did with a simple vam screenshot as input:

BONUS: NATURE IS BORING!

Title. So I personalized that recipe and went towards more artistic shots. This is what I used for the images at the top of the post. It's a hit & miss, 50% of the results are meh but it does make some very cool photos too in a very specific & unique style! I did only 40 images and picked from those.