oppaipriest
patreon
OppAIs - Part II: Prompting
🕑 Added 2023-01-24 02:31:17 +0000 UTCIn the previous tutorial we could see that, in order to obtain an image, it is necessary to write a prompt. The final result will depend largely on the prompt we write. My work methodology consists of writing enough to obtain a result that is as close as possible to what I am looking for, but without going into too much detail.
Now, how do I structure my prompts?
First of all, there is a couple of "tags" that I keep on using: masterpiece and best quality. All my generations start with those two tags. They help a lot and don't take up much space.
Then I write down the general characteristics of the image such as: drawing style, lighting, perspective and anything else that comes to mind.
The models I use are usually trained with data from Danbooru, so using the tags from that web site is very helpful: https://danbooru.donmai.us/wiki_pages/tag_group%3Aimage_composition
In this example, I want to work with a design that has: Korean RPG style, cinematic lighting and from above view.
Our current prompt looks as follows:
masterpiece, best quality, korean RPG style, cinematic lighting, from above
NOTE: Each attribute should be comma separated.
After them, I like to specify everything about the character in the following order:
- Gender
- Hair
- Eyes
- Physique
- Clothing
- Posing
As a fan of oppais myself, I will work with one of them. In this case, I want it to be a female with long black hair. Her eye color will be blue. As for her physique, I will only be in charge of increasing her breast size. Finally, I want her to be wearing a white silk dress and to be seated.
In that case, the prompt you would write would look like this:
female, long hair, black hair, blue eyes, huge breasts, wearing a white silk dress, sitting
If we add everything to our prompt, it would look like this:
masterpiece, best quality, korean RPG style, cinematic lighting, from above, female, long hair, black hair, blue eyes, huge breasts, wearing a white silk dress, sitting
I then take care of specifying the character's location and any other extra details I may have left in mind. For example, I would like her to be in a medieval style room.
masterpiece, best quality, korean RPG style, cinematic lighting, from above, female, long hair, black hair, blue eyes, huge breasts, wearing a white silk dress, sitting in a medieval-style bedroom
As you may notice, I usually add the character's pose along with the setting. Is it necessary? No, but I feel I get better results that way. It's not something I've spent time verifying, just a hunch.
Finally, I need to add negative prompts. These are everything I don't want to see in my images. I tend to have a number of prompts that I use no matter what I am looking for but, as with prompts, it all depends on what we are looking for.
These are the negative prompts I always use in my designs:
bad quality, worst quality, close up, elf ears, pointy ears, blurry, (dim colors), bland, watermarks, chromatic aberration, film grain
My models tend to generate a lot of elves when I include words like castle, medieval and fantasy. That is why I include pointy ears and elf ears to the negative prompts.
Anyway, you have seen several elves created by me. In those cases, I remove both tags from my negative prompts.
After trying all of the above, I get:
It worked! I got what I wanted. Now, I am 100% sure that your results will not have an anime girl in them. This is because I use a different model. I'll explain where and how to get models like mine to create specific content. In the meantime, let me teach you something else.
You can increase the weight of a tag by wrapping it between parenthesis and adding (not necessary, a decimal number).
By increasing the weight, you are basically telling SD that you really want what you are writing. For example, I want all images to have cleavage. So I will edit my prompt to have that new tag with increased weight:
masterpiece, best quality, korean RPG style, cinematic lighting, from above, female, long hair, black hair, blue eyes, huge breasts, wearing a white silk dress, (cleavage:1.4), sitting in a medieval-style bedroom
You will notice that I only added 1.4, SD is very sensitive to the number you put in. Try testing with a 10 and you'll see that your image will go to hell. Between 1 and 2 is more than enough (the latter being an extreme case). I recommend you to try it on your own.
This is the result I got after adding that new tag:
They all have cleavage now! Awesome.
Now is your time to experiment; you must not cling tightly to my prompting structure. Everyone has a different style of arranging things and the important thing is that you are satisfied with the result. What you must remember is:
- You must separate things using commas.
- Do not increase the weight by more than 2.
- Less is more, especially with negative prompts.
- There is no such thing as a perfect result, especially not the first try.
Comments
Charrise
Thank you!
Oppai Priest
If you refer to the model I'm using for the tutorial images, it is OrangeMix Hard from WarriorMama777, you can find it here: https://huggingface.co/WarriorMama777/OrangeMixs
Charrise
What stable diffusion are you using?
