Taming The Beast - Making AI Art With Your Computer!

Just want to say thanks for sharing.

Not so much your hunky, muscular anime dudes, but definitely for starting the thread and sharing your knowledge.

Looks like I have a new hobby of tinkering and learning AI, what it can/can't do, and how to speak the "prompt language" most efficiently.

However, having a locally installed instance that takes ~7-15 seconds to generate an image helps speed things up.

Now on to learning LoRAs and other engines.

Lots to learn, that's for sure...

How is this even possible?

I have a few more tricks up my sleeve when it comes to speaking the "prompt language". Check back to this thread, and learn from the master!


Seriously though, there are a few thing that aren't immediately obvious that I can teach you.

First, this AI is very far from GPT-4, but it does have a vague and rudimentary understanding of human language. When you write a sentence it will pick up on every word separately, but it will also pick up on the meaning of the sentence, at least to some degree.

This means it's best to write the words in a natural order that flows like normal language.

But don't overdo it! It's much better to write short "staccato" sentences than long descriptive sentences that use a lot of meaningless glue words. I already warned about using commas, because they're essentially adding trash to the prompt.

Now I extend this warning to little words like "the" and stuff like that.

It's also important to understand that the first word gets more weight. Make sure you put the most important words first, even if you must write a somewhat awkward sentence to do it. And don't use any more words than you need to convey your intent, extra words just dilute the power of the important words!

The AI doesn't care if your prompt reads like poetry, it just need the words to come in an order that helps it to understand.

There's more to it, I'll get back with more tips later, I don't have time to tell the whole story right now.

A good question.

The principle behind it is so ridiculously simple and so outrageously silly I would never have thought it could work.

You just add a little noise to pictures and train an AI to remove it. Then you give it a picture consisting of nothing but pure noise, and tell it to remove the noise.

Then it somehow removes everything except a picture that it digs out of the noise. A picture you never put there in the first place!

This absolutely brain dead idea works much better than it has any right to do.

Remember the old joke about how easy it is to make a marble sculpture? You just take a block of marble and remove everything that isn't a naked girl!

That's exactly what this AI does.

Here's a twitter thread that's complementary to the theme of this thread, just archiving it here

[link to twitter.com]

https://twitter.com/_/status/1640783207259074562

Last Edited by Vision Thing on 03/28/2023 04:21 PM

So far I've been very sloppy, just to quickly show what the AI can do. In this session I will be much more careful and explain in much more detail how to prompt effectively. Most advice you'll find on the internet is REALLY BAD, so pay attention here!

The big secret to get good pictures is to say as little as possible, to "inspire" the AI, rather than beating it over the head with detailed commands. The more explicitly you tell the AI what to do, the more you nail down exactly what has to be in the picture, and the less room it has to be creative. The result is a picture that technically contains all you asked for, but that is stiff and unimaginative.

You should start with a word you think does something interesting right off the bat, so you get a flying start. This time I'm going to use "coiffure", which is French for hairstyle.

pos: coiffure

https://imgur.com/iOUAwXx

This word already does a lot of work, and sets up both an interesting setting and an artstyle. There are some things I don't like about it though. One is that it almost always gives you two girls, and one is that it leans heavily towards black and white.

You should always generate a batch of ten or so images so you get a feeling for what a word does in general, it can do well on one seed and poorly on another, and some words are highly focused and know what they want, whereas other words are all over the place. You need to get to know your words!

As I've already said, I've discovered that you get an excellent starting point by using cancellation, that is, by using the same word in both the positive and negative prompts. The effects of the word don't disappear completely, but it gets pushed down into the noise. Then other words you add will pull out the aspects of it they resonate with, and you get synergy effects.

The main effect of using cancellation this way, is that the most generic and boring aspects of the word get suppressed, but the more creative and unique aspects tend to get pulled out as you keep working with it.

So let's try to cancel out coiffure and see what happens.

pos: coiffure
neg: coiffure

https://imgur.com/EGP2cfD

We still get "coiffure", but a much more broken picture. A lot of trash is generated because of the hard cancellation. This is not a big problem, because most of the trash will naturally go away when we add more words that aren't canceled out.

But now I'll show you a way to get the benefits of cancellation with less of the drawbacks. We simply do the cancellation with a different word that means the same thing!

In this case, the natural choice is "hairstyle".

pos: coiffure
neg: hairstyle

https://imgur.com/ZQaGlrp

The overall quality of the picture improves a lot, and we even get some color on the girls' cheeks!

Now I want to target the two main problems I had with coiffure in the first place, there's always two girls, and the picture lacks color. As I said, it's better the less you tell the AI exactly what to do, so it would be great to counter both effects at the same time with just one word. And since the girls look a bit stiff and dull it would be great if this word also infuses some positive energy while it's at it.

So we're looking for a vague and non-descriptive word that has negative energy, and that usually produces several people, mostly in black and white. If we put such a word in the negative prompt we will gently nudge the AI in the opposite direction, and the nudge will be so subtle it won't hammer it over the head and kill the creativity.

A word that does exactly what we want, is "conflict".

pos: conflict

https://imgur.com/MZm3sXP

Since we're going to add this word to the negative prompt it's a good idea to also check what it does alone when used that way, because we're subtly adding that to the picture, by removing the opposite!

neg: conflict

https://imgur.com/0ra68KD

We can see that using conflict in the negative prompt will accomplish what we wanted. We mostly get a single girl, we get color, and we get a positive energy. But the word isn't perfect, it comes with some drawbacks. It produces some nonsensical trash, and it pushes the artstyle towards dull generic anime.

All these effects will be added to our picture whether we want them or not. We can't pick and choose, we will simply have to counteract any bad effects a word introduces later.

For now, this is what we get when we subtract "conflict".

pos: coiffure
neg: hairstyle conflict

https://imgur.com/pekF7Mu

It's getting closer to a good picture, but we can see that we do have a problem with random trash. We need to find a word that cleans up the picture, preferably without screwing it up too much in the process.

One hallmark of a good cleaning word is that it produces pure trash when used in the positive prompt.

One such word is "measured".

pos: measured

https://imgur.com/PuFCUcZ

A good cleaning word should of course also produce a nice clean picture when used in the negative prompt. Does "measured" fulfill this demand?

I'll show you two different pictures produced this way, to illustrate something important.

neg: measured

https://imgur.com/P2cXLah

https://imgur.com/NCUlb1g

One picture is very clean indeed, but the other looks like a hot mess! What's going on here?!

The answer is that we aren't telling the AI what to draw, we're simply telling it to not draw "measured". So it draws absolutely anything it finds in the noise on random, and then "measured" steps in and removes what it doesn't like.

"Measured" has no opinions about the actual content of the picture, it just removes what it sees as "measured". And that happens to be what humans would call "ugly trash".

If you look again at the "hot mess" picture, you can perhaps agree that the mess is a good clean mess! Content-wise it's trash, but it's high quality trash!

All of this means "measured" is a word that tends to clean up poor quality trash without interfering too much with the actual content of the picture. If you stumble on a word like this, write it down, it's going to be very useful!

Let's see what "measured" can do for us!

pos: coiffure
neg: hairstyle conflict measured

https://imgur.com/WO475K7

A big improvement I would say!

We are getting a pretty good picture at this point, but the most important thing is that we have gotten there without saying much at all! We have only used four words, and only one of them is in the positive prompt where we say what we want. And most of that word was even canceled out!

This means we still have a lot of power left to mold the picture to where we want it.

You see, there is a "power budget" in play here. Every word you add draws from the budget and reduces how much power you have left, and when you already have twenty words in the positive prompt you won't get much mileage out of word #21.

The main reason you will get BAD ADVICE elsewhere, is because most people don't realize this. They write very long positive prompts, peppered with weak words that do little good, and of course plenty of commas that just soak up your descriptive power and introduce trash.

And then they write very long negative prompts that list all bad things they don't want to see, desperately hoping this will convince the AI to not draw them.

These are bad practices.

I will keep working on this prompt later and see what we can get out of it!

TO BE CONTINUED

Thank you for saving me hours days of research and trial and error!

And thank you for listening!

It will still take you days weeks months years of searching to find good words that do useful stuff though.

I feel like I could do anything with this if I only knew what all the words actually do!

I was going to continue my "tutorial", but since interest is waning and very few are still reading at this point I think I'll just drop it. I've basically shared what I have to share anyway, what remains is just to know your words!

GLP is mainly an American forum, and I live in Sweden. This means I go to bed right when GLP comes alive. I can never babysit my threads and interact with people during the hours that matter, so I don't make a lot of threads here, they just die anyway.

I keep checking this thread and I have shared it, I am not doing anything with AI right now but I know people who are, and you are sharing important information that someone out there needs.

So I will keep sharing where appropriate because someone needs this for sure.

Thanks for your generosity in making these careful posts.

If you guys are interested in the fast moving development of the personal use of AI, please follow this twitter account, Brian Rommele, he is developing techniques to generate prompts and different interfaces etc.

[link to twitter.com (secure)]

That's the link to his profile

Last Edited by Vision Thing on 04/02/2023 04:13 PM

I'll check it out!

Thanks for your contributions, and for trying to share the thread! I had hoped to gather a handful of people who are playing around with this, so we could discuss our findings and show our results. But it didn't happen. I think I'll add some more this evening anyway (evening in Sweden, that is). If at least one person reads it I guess it's worth the effort.

Last Edited by Lazy Monk on 04/03/2023 12:30 AM

If you are looking for the latest up to the minute on AI, there are other places on the internet. A simple search for stable diffusion will take you there.

There is usually at least one major update a week now in the fields of image or chat ai. I have been in tech since punchcards but have never seen a field in tech move this fast. You sit out a month and you will miss a LOT. Almost as if the AI itself is driving it's own development.

I would go LONG in hardware and SHORT in software unless it is AI only.

Yes, things are moving crazy fast. I think every word I write here will be obsolete in three years. Probably less.

As it is, the image generating AI has a very rudimentary understanding of human language and logic. It's safe to say this will rapidly improve to GPT-4 levels...

Today "prompting" is about teasing a machine that doesn't quite understand you into doing what you want anyway. Tomorrow it's going to understand you better than you understand yourself. It's going to be a completely different beast.

I'm going to post some more pictures, I'll just BUMP us to page 4 so it's easier to find.

GLP won't have it. I was bumping too fast, got IP banned, so I'm changing IP and just posting it...

Sigh. It would have been so nice to put the update right at the top of page four.

Last Edited by Lazy Monk on 04/03/2023 12:29 PM

GIRLS AT WAR!

So far I've used what I call "near cancellation" to get an interesting starting point with some creativity right off the bat. This works very well, but is not strictly necessary. I'll use a more standard prompting strategy today.

Let's make a pretty girl stuck in the middle of war! We will start by mentioning the major points of interest in a human face, so the AI knows where to focus its effort.

pos: eyes lips

https://imgur.com/eBsW9DH

The AI gives us what we asked for, but it's doing it in a very generic and boring way. There is a little trick we can use that helps a bit here. Instead of separating the words with whitespace we can use the builtin operator AND (and no whitespace!).

pos: eyesANDlips

https://imgur.com/YepgoA4

The difference is that in the first case the AI tries to read the two words as a unified sentence, whereas in the second case it tries to read them separately, and then attempts to produce a picture that contains both items. When it looks up "eyes" and "lips" separately it accesses more beauty and fashion related data, so the result becomes slightly less "generic anime" in nature.

It is good to use AND instead of whitespace when you list stuff you want in the picture, such as body parts. It also helps quite a lot with getting the logic right! If you ask for a burning cottage and a dinosaur (separated by whitespace or punctuation) you might actually get a cottage and a dinosaur on fire, but by separating them with AND the AI will understand which thing is supposed to be burning.

The first thing to note at this point, is that we often get a split picture! As usual, the way to fix problems isn't to nag at the AI by telling it to not do this or that. Instead we find some word that consistently produces split pictures, preferably of very low visual quality, and then we put this word in the negative prompt. One such word is "compared".

pos: compared

https://imgur.com/JcPviFc

Writing "compared" in the negative prompt will remove most of the tendency for split screen, but at a cost! The art style takes a step in the wrong direction and gets a bit flatter. Everything you do has a cost, every word you add is at best two steps forward and one step back, you just need to make sure there's a net gain and keep compensating for what went wrong.

pos: eyesANDlips
neg: compared

https://imgur.com/w6cXgAT

We lost a little art style quality, but we can regain some of it by listing more body parts that should be in the picture. I want to zoom out a bit and make it just a tiny bit sexy, so I'll add "ANDmidriff". I could have added "ANDnips" to make it even more interesting, but this is GLP, so we're keeping it very tame!

pos: eyesANDlipsANDmidriff
neg: compared

https://imgur.com/WB172qV

The AI seems to be going in the direction of leather attire with a slight futuristic flair at this point. We could easily convince it to reconsider, but one of the keys to getting good pictures is to develop them in a "dialog" with the AI. You tell it something to inspire it, it gives you a picture to inspire you, and then you keep building on that. So I'm going to work with what the AI does.

I happen to know that "interesting" sends the AI into "cool futuristic leather space", so now I'll use the BREAK operator to mix in a second prompt.

pos: (eyesANDlipsANDmidriff:1)BREAK(interesting girl:1)
neg: compared

https://imgur.com/y0Opao1

At this point we often get some kind of blurry grey stuff in the background that could be city buildings or some form of military gear. I'll make up the AI's mind and tell it to go in the second direction, and also to push the futuristic angle with "scifi".

pos: (eyesANDlipsANDmidriff:1)BREAK(interesting girl:1)BREAK(scifi military installation)
neg: compared

https://imgur.com/F2rWJL9

As we add more prompts with BREAK they each get less power. We can see that the art style is steadily drifting back towards generic anime, which is where the model I'm using wants to go if you let it. I'll add some extra pressure to try to counteract this a bit, and also to make sure the AI doesn't forget about the boobs. Boobs are very important!

pos: (eyesANDlipsANDmidriff:1)BREAK(interesting girl:1)BREAK(eyesANDbreastsANDbare midriffANDscifi military installation)
neg: compared

https://imgur.com/fhpxYf3

All military girls will be dressing like this in the future, God willing! But the environment could use some more detail, and hopefully make a little more sense.

pos: (eyesANDlipsANDmidriff:1)BREAK(interesting girl on carrier deck:1)BREAK(eyesANDbreastsANDbare midriffANDscifi military installation fighter spaceship)
neg: compared

https://imgur.com/gJxe2ww

The amount of skin we get varies, but it's generally a bit too much. So I think we can afford to lessen the "girl pressure" by mixing in yet another prompt, this time focusing on mood. I want a much darker and grittier atmosphere.

I also discovered I forgot to write the ":1" at the end. Ooops! You get ":1.1" by default when you write something in parentheses, which means the picture gets "pushed" with ten percent more "attention" than default. I generally don't like the effect of this, since it leads to overcooked pictures, so I'll reduce the parameter to 1 where it should be.

pos: (eyesANDlipsANDmidriff:1)BREAK(interesting girl on carrier deck:1)BREAK(eyesANDbreastsANDbare midriffANDscifi military installation fighter spaceship:1)BREAK(dark gritty dystopian:1)
neg: compared

https://imgur.com/ZFbldOQ

Continued in a minute...

Last Edited by Lazy Monk on 04/03/2023 12:30 PM

It's still not dark enough for my taste, because we have mixed four prompts, and only one of them is gritty. I'll try to make the "interesting girl" a little darker too. We also get red eyes for some reason, which I don't like, I'll try to remove some "demonic" to see if it helps.

pos: (eyesANDlipsANDmidriff:1)BREAK(interesting deadly serious girl on carrier deck:1)BREAK(eyesANDbreastsANDbare midriffANDscifi military installation fighter spaceship:1)BREAK(dark gritty dystopian:1)
neg: compared demonic

https://imgur.com/sMo7nDm

It's starting to get where I want it, but I'd like to see some more mayhem in the background! I'll just throw in a bunch of chaotic stuff in the prompts and see where it goes.

pos: (eyesANDlipsANDmidriff:1)BREAK(interesting deadly serious girl on bombed carrier deck at war:1)BREAK(eyesANDbreastsANDbare midriffANDscifi military installation fighter spaceship explosion smoke:1)BREAK(dark gritty dystopian burning wreckage:1)
neg: compared demonic

https://imgur.com/dSLm0BS

Yes, that is the standard issue military uniform for girls in 2050! Don't ask.

I'm more concerned about the fact we're drifting too far towards generic anime space for the grittiness I'm after. The word "photorealistic" is pretty shitty and doesn't do as much good as you'd expect, but it can be of some use of you use it with care.

pos: (photorealistic eyesANDlipsANDmidriff:1)BREAK(interesting deadly serious girl on bombed carrier deck at war:1)BREAK(eyesANDbreastsANDbare midriffANDscifi military installation fighter spaceship explosion smoke:1)BREAK(dark gritty dystopian burning wreckage:1)
neg: compared demonic

https://imgur.com/UH2xeL9

What's this? Is it war, or a modeling shoot? The AI doesn't seem to be quite sure any more, so we'll set it straight by removing "modeling".

Some of the pictures we get are pretty good at this point, so I'll show a couple.

pos: (photorealistic eyesANDlipsANDmidriff:1)BREAK(interesting deadly serious girl on bombed carrier deck at war:1)BREAK(eyesANDbreastsANDbare midriffANDscifi military installation fighter spaceship explosion smoke:1)BREAK(dark gritty dystopian burning wreckage:1)
neg: compared demonic modeling

https://imgur.com/hTVGAGK

https://imgur.com/2ROtEIY

https://imgur.com/zkWe4yv

It's getting better, but we still need some more grittiness, and less sexy modeling. I want a sort of "accidental" sexiness, not seductive posing! I also push some more "photorealistic" to push away from the generic anime aesthetics. This is done with a model that is heavily skewed towards anime, you should use another model if you really don't like the style.

pos: (photorealistic eyesANDphotorealistic lipsANDmidriff:1)BREAK(interesting deadly serious girl on bombed carrier deck at warANDdusk fire mayhem:1)BREAK(photorealistic eyesANDbreastsANDbare midriffANDscifi military installation fighter spaceship explosion smoke:1)BREAK(dark gritty dystopian burning wreckage:1)
neg: compared demonic modeling posing doll

https://imgur.com/0XXO0BT

https://imgur.com/KcwXdaJ

https://imgur.com/IcolALO

https://imgur.com/VD4zZe2

And that's it for today! A lot of people tune their prompt with the same random seed until they get a perfect result, but I prefer to make a slightly chaotic prompt that gives more random results and cherry pick some I like. I like the surprise of fucking around and finding out what I get!

If I was serious about getting a good final result that would be used for something I would pick one seed and fine tune the prompts until it's perfect. But I only do this for fun, I delete all the pictures after I'm done!

Last Edited by Lazy Monk on 04/03/2023 12:37 PM

Thank you! FWIW, I enjoy reading your updates.

I've now downloaded ~200gb (+/- some) of LORAs, Textual Inversions, and poses.

Also attempting to learn how to teach it myself for certain objects/people.

Definitely lots to learn.

[link to www.facebook.com (secure)]

Your interest is worth a lot, actually. It makes no sense to keep posting if not at least one person cares!

I haven't even looked into LORAs and stuff like that yet... So far I've just been trying to get a good grasp on the basics, which is how to prompt.

This technology develops so fast it's almost impossible to keep up! New addons are added every week. Some that I really need to start looking into:

DreamArtist - teach the AI to replicate art styles and picture content from ONE training image.

openOutpaint extension - expand existing images with more similar content to make them bigger.

ControlNet - makes it possible to get more control over the generations, such as controlling posing. Extremely powerful tool with a lot of possibilities.

MultiDiffusion with Tiled VAE - an advanced upscaling tool that adds more detail as it upscales.

Aaaand of course LORAs!

Since you're getting into LORAs, why don't you show me what they're capable of and post some of your results! Don't be afraid of imgur, it's super easy to use, and you don't even need an account. You just drag and drop a bunch of images to imgur, and you get a page with links to copy!

honest question for you no hate..

are you a j?

Honest answer, no.

This is the third time I'm getting this question. I do wonder what it is about me that gives this impression. I do have a rather impressive nose, but you can't see that!

Interesting thread. The BREAK command has been extremely useful. Any general tips for when it's appropriate to use BREAK versus when it's appropriate to instead use AND.

I'm curious if smaller search terms with more BREAKS will yield better results than longer sentences using AND

Also, what is your favorite sampling method, what is your resolution, and what are the number of sampling steps you are using?

Honest answer, no.

This is the third time I'm getting this question. I do wonder what it is about me that gives this impression.

well at least you are honest.

i'll leave this programming stuff to those better at it.

can you make the girls into older girls or younger women?

For a more mature look, try using "woman" instead of "girl". The "mature" and "adult" tags may help as well.

Always put (adorable cute dwarf:1) in the negative, as OP suggested, to remove childish traits.

Sure! I'm normally not a big talker since stuff disappears, but your thread gave me the breath for the second chance I needed to take a stab at this stuff. I'm not an expert and I have NO idea wtf I'm doing. I learn a lot by example, trial and error, testing boundaries and seeing what I can do. This means I copy code a lot, then break it down into pieces, then run things separately to see what they do. After I have enough of that under my belt, I start taking the pieces of multiple jigsaw puzzles and put them together to form my own picture/puzzle/program. Sometimes it works, sometimes it doesn't, and it makes me learn more.

That said, here's what I'm doing in my current "copy and learn" phase:

In AUTOMATIC1111 Stable Diffusion WebUI, I found a button under Generate that's labelled "Show/hide extra networks" (it's magenta) and it popped down some extra tabs like "Textual Inversion, Hypernetworks, Checkpoints, Lora". I've heard once that "Lora's are awesome" so I looked into it. A "LoRA" is a "Low-Rank Adaption" (But wtf cares about all that techno-mumbo-jumbo you can read about with a search).

Basically a LoRA is a file that is "specifically trained" for a particular person, place, or thing.
Quick example: A LoRA of Scarlett Johansson can be used to tell SD to create an image of "black widow wearing an ironman outfit"

Let's do this real quick.

I found "civitai.c om" and they have TONS of stuff for the hobbyist there. Most shared pics have an (i) in the corner and if you click it, you will see the positive & negative prompts, the engine used, the steps, seed, etc. If you copy and paste all the info into the appropriate places, you can generate the same image on your PC.

So, we go get a pic of RDJ as Ironman with this info:

Note: The red "keyword" is pulled out of the red lora tag. The Tag is placed in your prompt when you click it in the menu

Prompt: extremely detailed photo a man wearing iron man armor, lorarobertdj, stylish short beard, model photoshoot, 8k UHD, ultra detailed, RAW photography <lora:robertDowneyJrLora_10:1>
Negative prompt: monochrome, ugly, boring, cartoon, anime, disfigured, multiple heads, extra fingers, fused limbs,
Size: 512x768,
Seed: 1351610750,
Model: realisticVisionV13_v13, (Note: I'm using V20_v20)
Steps: 50,
Sampler: Euler a,
CFG scale: 7

So I put this info in my A1111 SD and get this:

https://imgur.com/nBtJ4LK

But, wait -- That's not the red Ironman suit we're used to. I'm guessing this is because I'm using V20 of the engine and not V13, so I update just the prompt:

Prompt: extremely detailed photo a man wearing red ironman armor, lorarobertdj, stylish short beard, model photoshoot, 8k UHD, ultra detailed, RAW photography <lora:robertDowneyJrLora_10:1>

And now, I get:

https://imgur.com/ZMV6lFb

Much better! Now I have a working "Ironman with RDJ" I can work with.

Now, getting back to my quick example, I go download the Scarlett Johansson LORA, then just slightly modify my prompt:

Prompt: extremely detailed photo a woman wearing red ironman armor, scarlett, model photoshoot, 8k UHD, ultra detailed, RAW photography <lora:scarlettJohanssonLora_v10:1>

And we get:

https://imgur.com/qFFj1iX

To me, this doesn't look like ScarJo -- AT ALL. Maybe it's the LoRA model? Maybe it's the engine I'm using? I don't know, but I'm going to bump the "Seed" from 1351610750 to 1351610751 then 1351610752 and check:

https://imgur.com/XyciwVr

https://imgur.com/4i5281k

Eh - The last one kind of looks like it, but lets go grab a couple more "heads" and just replace the Lora and Codeword:

https://imgur.com/65Lfk3H

https://imgur.com/DoXKYFQ

https://imgur.com/uHxFTUr

Eh, it starts to get a little crazy, then you have to tweak the prompts a little bit to fix things here and there.

Also, if you made it this far, you'll now get to learn that "Textual Inversions" and "LoRAs" are basically the same thing. LoRAs, I believe are the "new TIs". Textual Inversion is just a fancy way to say word-substitution.

When you say "photo of a man" -- It will give you a man.
If you provide a LoRA and say "Photo of a man hughjackman <lora:hughJackman_1:1>" it will give you a photo of a man, but to see what that man should look like, you're providing a model to work from. The ":1" tells it "100% of this!"

If you only want it to be "50%" of that, then you would use ":0.5".

Back to fun...

BREAK divides up chunks of tokens to send to CLIP, which is the part of the AI that does the reading comprehension, such as it is. The understanding of text is very rudimentary, but it does have some understanding of how words form sentences.

This is for good and bad. Sometimes you want to take advantage of the added reading comprehension that comes from not using BREAK. But sometimes it just gets in the way of getting your intent across.

Why?

Because the longer a sentence gets, the less impact a word has. When the impact gets to low and words don't do enough, insert a BREAK, and you effectively start all over with a second prompt. You get more impact again, but you push back the entire original prompt, and the more extra prompts you BREAK in, the less impact each of them gets.

I think the best way is to write perhaps five words, then BREAK, and keep doing this over and over until you get too little impact by adding more BREAKs. Then you switch to inserting more words after the five words in the individual prompts. Now you get some more power back, and can keep nudging the result until it's just where you want it.

In other words, BREAK is a tool to manage your "power budget".

Another good "power budget" managing technique, is to insert about the same amount of stuff in the positive and negative prompts. This is because you will naturally move from making big changes to small final nudges. And if you have very little in the negative prompt when you're reaching "end game", then you can't add anything there to nudge, because anything you add has too much power, and will ruin the fine balance you have struggled with.

If you have say four prompts separated with BREAK both in the positive and negative side, and each of them has about five words, then you can add a final nudging word in eight different places, for different effects. That's a lot of nudging power!

You don't need to be religious about this, but it's good to know how to keep your power budget in the black. It sucks to have a almost ready picture going, but you have run out of power and the AI won't respond to any more instructions.

What about AND then?

AND is inserted in the token stream, and the intent is that it shall mean "do this AND also do this". This actually works to some extent! You can separate things that don't belong together with AND, and the AI will understand to not mix them. Well, for the most part.

As I said before, "burning cottage and dinosaur" might give you a cottage and a burning dinosaur, but "burning cottageANDdinosaur" might actually give you what you expect! The "and" is just a word, whereas "AND" is a special operator.

One thing to know, is that statements separated by BREAK are order independent. That is

(beautiful girl:1)BREAK(bare breasts:1)

is the exact same thing as

(bare breasts:1)BREAK(beautiful girl:1)

Everything else you do is very order sensitive!

For a while I wasted a lot of time doing permutation testing to find the optimal ordering of words. This approach scales poorly to say the least. Three words have 6 permutations, four words have 24 permutations, five words have 120 permutations, and it only gets worse from there!

But eventually I realized that there is really only one place to add new words: at the end of the sentence. You can insert it in the middle if it's for instance an adjective that is meant to be read with a substantive, like this:

beautiful girl
-->
beautiful gorgeous girl

But for the most part the right place to insert a new word is at the end of the phrase. Especially when it doesn't really produce a readable sentence wherever you put it.

This is because you are developing your prompt in a dialog with the AI. You add a word BECAUSE the picture looks like it does right now. If you insert it in the middle of a prompt you change the way the AI reads and understands the existing words, which fucks up the balance you have accomplished so far.