AI tips and tricks thread

Tyranomaster

Guy who writes stuff
Joined
Oct 5, 2022
Messages
335
Points
108
While I find myself in the airport waiting to board my plane, I thought to pass the time by passing on some tips and tricks to generating AI art.


I'm going to separate this out into two sections, website generation, and own pc generation. Own pc generation assumes you already have things set up and are here for generation tips, not setting up ai generators on your own pc. I won't be posting it now, but when I get home and have access to my pc I will.


Websites

As many of you are aware, using websites for generation usually comes with some restrictions. Either limited generation, long queues, or a pricetag. So these tips are here to help you get the most from each generation!

The Basics

1. Formatting - You'd be surprised at just how much of a difference your prompt format makes. A good prompt breaks things into small information packets. A bad prompt is a run on sentence. Most important info first, separated by commas.

Good prompt example: masterpiece, 8k, highres, best quality, 1girl, ornate dress, aqua hair, ponytail, sitting, fancy table

Bad prompt example: a girl with an ornate dress and aqua hair in a ponytail sitting at a fancy table.

2. Quality modifiers - most websites add a few of these in at the start of any prompt. They perform best at the start as well. Things like masterpiece, 8k, etc. There are some hidden gems out there that help prompt quality when you add them, digging around on the internet will help.

3. Danbooru tagging - I'm on the fence myself on this one. They're a good starting point. Know that Danbooru is 18+, but the tags from there make good comma separated prompt items. There are lots of things that aren't on danbooru you can use as well, so experiment!

Intermediate Techniques

1. Check png meta data from websites to see what they use in positive and negative prompting. This is important to know for not duplicating those quality tags like masterpiece.

2. While there is no hard limit on the size of a prompt, 75 tokens is an advised limit not to go over. One comma separated tag represents between 1 and 5 tokens. When you start to cross 75 tokens, the quality usually tanks in my experience. (Many websites don't show you tokens.)

3. Parentheses - [Edit: Parentheses add emphasis on what is inside] () some people like to stack them (((blue eyes))). It's better to get in the habit instead of (blue eyes:1.3) which is approximately the same. One parentheses set is identical to (example:1.1). They can be nested in non linear ways ((blue) eyes:1.2). ((Blue eyes)) is (blue eyes:1.21) (1.1x1.1).

4. You can group items within a single comma, and the ai tends to get the idea.
Blue, sky isn't the same as blue sky.

5. Punk - you're familiar with this. Cyberpunk, steampunk, etc. But did you know that the ai basically generalized this? Punk still means punk (spiked hair, spikey clothes, black), but it also works as a modifier meaning genre. Things like tentaclepunk, firepunk, goldpunk. These are all different than using either word on their own.

6. Complex words - avoid them. IF the ai was trained with them, the sample is small, and will not help much. Words like Exasperated, or Differentiable

Advanced Techniques

1. Alternating prompt - [cat|dog] this format makes the ai alternate between cat and dog on every step. These can be nested and used in a lot of ways. Think of it as half and half.

2. Change partway - [cat:dog:5], on the fifth step, switch from cat to dog. This is much more useful. Wish you had a car shaped like a fish? Stop saying fish car, instead turn a fish into a car. You can also say [cat:dog:0.5] which will do the change halfway. (0.25 is a quarter etc). Know that different things manifest in different steps. Hair color is generally set in the first few steps.

3. Trimming - Have a prompt almost perfect, but just shy? Use the same seed, and modify the attention slightly. Maybe (boxing gloves:1.2) can be (boxing gloves:1.17). It can easily help perfect a close images.

(Edit) 4. Models - different models have different inbuilt importance on things. In one model, blue eyes might be fine, in another, blue eyes might try to hard bake blonde hair in as well. When you switch models, you might need to tinker with weights for your go-to prompts.

I'm sure I forgot some things, but my flight is here, so until then, enjoy!
 
Last edited:

melchi

What is a custom title?
Joined
May 2, 2021
Messages
1,937
Points
153
Combing prompts is a nice trick as well. Suppose your character has a blue eye and a green eye.

Blue eyes, Green eyes

Doesn't usually work that well

Heterochromia works, however with just that tag the eye color tends to be whatever it around it.

Grouping those three tags can give a bit better results.
EX: {heterochromia, green eyes, blue eyes}

Blue eye, green eye Doesn't work too well, I think that the machine learning works better with eyes plural.
 

Tyranomaster

Guy who writes stuff
Joined
Oct 5, 2022
Messages
335
Points
108
Combing prompts is a nice trick as well. Suppose your character has a blue eye and a green eye.

Blue eyes, Green eyes

Doesn't usually work that well

Heterochromia works, however with just that tag the eye color tends to be whatever it around it.

Grouping those three tags can give a bit better results.
EX: {heterochromia, green eyes, blue eyes}

Blue eye, green eye Doesn't work too well, I think that the machine learning works better with eyes plural.
Might try [blue eyes|green eyes], (heterochromia). I haven't tried it, but it sounds like it might work.
 

SirDogeTheFirst

Lord Of The Potatoes
Joined
Feb 11, 2021
Messages
307
Points
103
Good prompt example: masterpiece, 8k, highres, best quality, 1girl, ornate dress, aqua hair, ponytail, sitting, fancy table
So I tried this, and holy ***. It makes a lot of difference. I am currently testing this on several AIs to see how much difference I will get in each one compared to my old prompt.
 

Tyranomaster

Guy who writes stuff
Joined
Oct 5, 2022
Messages
335
Points
108
Some ai art Gens let's you use negative prompts.
Negatives can be useful to remove a conflicting idea.

Example: A prompt keeps overriding a haircolor with a different color from the prompt.
Pos: Blue eyes, black hair
Negative: blonde hair, blue hair

Other than that, negatives are generally a long, generic list, that most websites all use the same ones. Things like mutant, or extra arms.
 

melchi

What is a custom title?
Joined
May 2, 2021
Messages
1,937
Points
153
You can deemphasize a prompt too. Some words overpower everything and need to be told to calm down.
 

Civilian

Active member
Joined
Apr 18, 2023
Messages
51
Points
33
Thanks, I downloaded Stable Diffusion a few days ago, but everything turned out like The Thing from 1982 so I kinda gave up on it. Will have to practice a bit with these.
 

Tyranomaster

Guy who writes stuff
Joined
Oct 5, 2022
Messages
335
Points
108
A Brief History of Stable Diffusion Models, for Prompters

Stable Diffusion goes back quite a ways, and I'm not going to get into the entire history of who did what, but I'm going to start where home generation really took off.

October 2022

In October of 2022, NovelAI's checkpoint leaked onto the internet. There was a lot of drama at that time around this, and man oh man did things take of with generation and prompt engineering. Why? Because the internet is for p0rn, that's why. NovelAI had spent a few hundred thousand dollars training on top of the Stable Diffusion 1.4 model on anime images from Danbooru, keeping their tags for textual inversion. 4chan and the "Automatic1111" interface quickly figured out the default settings on NovelAI, and were able to 1to1 replicate images on their own pcs. Various settings were messed with, and some basic addons were developed. The most popular at that time was wildcard.

People who could get in on this were gamers who already had high end cards. Hobbyist generation machines were theorycrafted. You can probably build a PC for generation of AI Art for about $500 now. You basically buy a bunch of cheap parts, and an RTX 3060

Stable Diffusion 1.5 also released at the end of the month, improving heavily on Stable Diffusion 1.4

The short and dirty: Almost all models for Waifu generation have danbooru tags baked in to some degree or another. Other things still work though.

November 2022

4chaners en mass had been prompting at an astounding rate. New addons cropped up, and the use of Prompt Search and Replace became common on X/Y grids. November was probably more groundbreaking than October though. Why? Model Mixing and Home Training.

The first of the serious model mixes started coming out in November. The big one at that time was Berrymix. Berrymix was made by mixing a few models designed for NSFW materials, but ended up improving other basic aspects of generation, leading to fewer body horror images.

A few weeks after Berrymix, following in its philosophy, I did a massive mix matrix, spending a few days mixing and testing images using some of the new models that dropped, like SD 1.5, and Z222(now defunct and almost impossible to find). Thus blueberrymix was formed. You can't find it anymore anywhere (civit has something called blueberry, but it isn't og blueberry). Others used the mix ratio to make similar mixes using other models and versions to merge.

Then came Anything. Anything was basically a mixed chinese model that heavily used NovelAI. At this point, we're pretty sure no extra training went into it. Almost every Anime Model you find today, whether Abyss Orange, or Counterfeit, likely has its origin in mixes based on Anything, which is based on NovelAI. Know that the more you mix, the more the original danbooru tags get lost in the weeds.

Prompt S/R (Search and Replace) allowed us to check variations of words and strengths by replacing single or multiple words in a prompt, to see exactly how it affects outputs in a grid format. Ever wonder how well anime studios work as prompts? Put each in an SR. I can attest that the only one that seems to work well is a-1 pictures.

Home training also took off. People before were attempting this stuff, but figuring out how to use Lora and Hypernetworks to add on top of a model rather than modifying the entire model was a breakthrough for hobbyists.

The short and dirty: Modern Waifu models are all descendants of NovelAI and use Danbooru tags. Addons exist in home run machines to allow massive prompt check and testing for ease of discovery.

Everything Else

The rate of hobbyist development really slowed down from here, but there were still breakthroughs. Prompt X/Y became Prompt X/Y/Z, and we got tons of addons that are useful. DAAM script and GIF generation are examples. New model mixes came out, and new versions of old mixes as well. CivitAI now allows easy and safe distribution of Models.
 

Cipiteca396

🐉🪽🍂🌑🍀🪶🌺
Joined
Jun 6, 2021
Messages
2,200
Points
153
Check png meta data from websites to see what they use in positive and negative prompting. This is important to know for not duplicating those quality tags like masterpiece.
How do you do this? Or is it something that you need the actual program to do?
 

Tyranomaster

Guy who writes stuff
Joined
Oct 5, 2022
Messages
335
Points
108
Sometimes you have to back up a little bit with the camera to detect the QR, but this generally works. I've been tinkering with making QR codes with AI. Simple QR codes work really well. Links like this... not so much.

00015-3235498034-8k, highres, masterpiece, ultra-detailed, (high detail_1.1), (library_1.2), b...png
 

dummycake

Already daydreamed about this interaction
Joined
Apr 4, 2023
Messages
1,168
Points
128
I can't find a perfect equivalent of doing this blood background

girl, transparen 912b8.jpg

with blue instead of red

the best I found was this one with "drippy blue paint background"

girl with brown  c8336.jpg

it's pretty good but I think it could be better???

what do you think?
 

dummycake

Already daydreamed about this interaction
Joined
Apr 4, 2023
Messages
1,168
Points
128
It might not actually be that hard. It's pretty straight lined. I mean I'm at work, looking at this through my phone, so maybe it looks different.
I meant relatively hard, it's harder than trying a bunch of combination and generating again
 
Top