Explaining Prompting Techniques In 12 Minutes – Stable Diffusion Tutorial (Automatic1111)

[00:00] hey everyone I'm b-size genius prompting

[00:02] instable diffusion can be a mystery and

[00:05] it's sometimes tricky to know what does

[00:07] what but there are techniques that you

[00:10] can use to get the results you want and

[00:12] in this video I'll be breaking them down

[00:14] so you can spend less time reading and

[00:16] more time creating but like the video

[00:19] and give it to me bite-sized so let me

[00:21] quickly cover some basic information

[00:24] prompts are ordered from most important

[00:26] to least important from the top to the

[00:29] bottom from the left to the right there

[00:31] were various theories on how you

[00:33] structure your prompt for the best

[00:35] result but it's worth keeping in mind

[00:37] certain Concepts such as the subject

[00:40] lighting photography style color scheme

[00:43] doing words and much more which help to

[00:46] build up your image when it comes to

[00:49] style prompts can influence this

[00:51] alongside your desired checkpoint as

[00:53] stable diffusion was trained on a

[00:54] multitude of data sets from across the

[00:56] internet meaning you can draw references

[00:59] to art style

[01:00] celebrities clothing types and much more

[01:03] to influence your image The Prompt

[01:06] sections have these numbers on the top

[01:08] right called token limits and these

[01:10] refer to the maximum number of words you

[01:12] could fit into a chunk of 75 tokens for

[01:15] example if you had 100 tokens then it

[01:18] would process 75 tokens and then 25

[01:21] tokens independently into stable

[01:24] diffusions unit this in plain English is

[01:28] how the AI language model breaks down

[01:30] and manipulates text for processing The

[01:33] Prompt box is where the magic happens

[01:35] this is where you will describe

[01:37] manipulate and design your image through

[01:39] the text to image section or alter

[01:42] images if you're using image to image in

[01:44] conjunction with your reference photo

[01:46] you can put as much text as you like in

[01:48] here but often it's better to keep

[01:50] things short and sweet so your prompts

[01:52] are easier to fix orine as you make

[01:55] adjustments closer to your desired image

[01:57] for example let's type out medium shot

[02:00] of a woman with four lips golden eyes

[02:02] and a white crop top long black hair

[02:05] snowy weather octane render we can see

[02:08] that stable diffusion will attempt to

[02:10] generate its interpretation of that

[02:11] prompt and this interpretation will

[02:13] change depending on the seed image size

[02:17] and other settings which change the type

[02:18] of image generated from the same prompt

[02:21] the negative prompt box is where you

[02:23] tell stable diffusion what you don't

[02:25] want in your image and this can include

[02:28] literal Concepts items weather or

[02:31] artifacts and bad Anatomy within an

[02:34] image again you can put as much text as

[02:37] you want in here but it's good to keep

[02:39] it reasonable to make life easier for

[02:41] yourself later down the line for example

[02:44] let's type out bad dream unrealistic

[02:47] dream nons saafe for work and we will

[02:50] notice that we get a much higher quality

[02:51] image that's safe for work it's worth

[02:54] checking the example images of the

[02:56] checkpoint you are using to ensure that

[02:58] you are using the best NE negative

[03:00] prompts for your chosen model the

[03:02] parenthesis is used to put a greater

[03:04] weight or importance on a word than your

[03:06] prompt and for each parenthesis wrapping

[03:09] a word it would increase its attention

[03:11] by a factor of 1.1 and the further

[03:13] parenthesis will multiply the attention

[03:16] by 1.1 I'll be honest in saying that the

[03:19] mass behind this isn't really all that

[03:21] important as experimenting with your

[03:23] prompts will yield better results but

[03:25] it's still important to see this in

[03:27] action so you can find two your images

[03:30] so for example if I were to write out

[03:32] our previous prompt but put snowy

[03:34] weather in parenthesis then we can see

[03:37] how the snow and even the clothes start

[03:39] to turn either white or have a much

[03:41] snowier presence the more parenthesis we

[03:43] use square brackets are used to reduce

[03:46] the weight or the importance of a word

[03:48] in your prompt and for each square

[03:50] bracket it will decrease the attention

[03:52] to the word by 1.1 and of course we'll

[03:55] multiply the detention by 1.1 for each

[03:58] pair of square brackets WRA a word again

[04:01] it's better to experiment when it comes

[04:03] to fine-tuning your prompt but by

[04:05] understanding what these functions do

[04:06] you have another tool in your belt for

[04:08] fine tuning images so for example if I

[04:11] were to write out our previous prompt

[04:14] but put snowy weather in square brackets

[04:16] then we can see how the snow is reduced

[04:19] especially around the hair and the

[04:20] clothing the more square brackets we use

[04:23] let's address prompt waiting you can

[04:26] manipulate prompts to either add or

[04:28] remove waiting or importance from words

[04:30] within your prompts and what this means

[04:32] in plain English is that you can control

[04:35] how much impact certain words have over

[04:37] others within your prompt and words with

[04:40] greater impact will be visualized more

[04:42] strongly in your image this is done by

[04:45] wrapping your word in a parenthesis but

[04:48] we will now add both a colon and then a

[04:50] number which can be a whole number or

[04:52] with decimal values so for example if I

[04:56] were to write out snowy weather with a

[04:58] variety of prompt ratings we can see how

[05:00] the impact changes from no promp waiting

[05:03] to a really high promp waiting value we

[05:05] obviously get rather lowquality images

[05:08] at very high waiting so keep the numbers

[05:10] reasonable you may have seen these

[05:12] angled brackets in certain prompts

[05:14] perhaps some websites that provide

[05:16] checkpoints and luras these are known as

[05:19] embeddings and are common in luras where

[05:21] a file and multiplier for the file needs

[05:24] to be specified to determine the

[05:26] strength of the Laura the typical format

[05:28] would be Laura file name and then

[05:31] multiplier and unfortunately you can't

[05:33] do prompt editing with Aur in this

[05:36] version of stable diffusion in this

[05:38] example I've used the add details Laura

[05:41] which adds or removes detail from our

[05:43] generated images and we can see the

[05:46] effects of changing the values within

[05:48] our embeddings and how it is structured

[05:51] now prompt editing is a powerful way of

[05:54] controlling your generated images by

[05:56] swapping the prompts being used for

[05:58] generating an image during the

[06:00] generation a good way to illustrate the

[06:02] format is by using from to and when

[06:06] where from determines what prompt you

[06:08] start with two determines which prompt

[06:11] you end with and the Step at which the

[06:13] switch takes place is determined by when

[06:16] you can also do from and when with two

[06:19] colons in between to remove the prompt

[06:21] being specified after a fixed number of

[06:24] steps determined by when and don't worry

[06:27] I'll show you this in practice so you

[06:29] don't feel like you're in another Mass

[06:30] class but remember this key piece of

[06:32] information the weightings you provide

[06:35] in decimal numbers are percentages of

[06:37] your sampling steps which must total up

[06:39] to one meaning 0.5 is actually 50% of

[06:43] your total sampling steps anything l a

[06:46] whole number above one will be the exact

[06:48] sampling step you want to specify

[06:51] meaning a figure of 20 means stop or

[06:54] switch at sampling step 20 so for

[06:57] example you can see the impact in these

[06:59] images with our first image using

[07:02] decimals which represent percentages to

[07:04] transition from snow to Sun after a

[07:07] certain number of steps the second image

[07:10] uses whole numbers to represent the step

[07:12] for the transition based on the 30 steps

[07:14] we are using and lastly we can indicate

[07:17] at which step we should stop using the

[07:20] selected prompt s weather Al together

[07:22] now here's a call trick using a

[07:25] backslash before a special character

[07:27] such as a bracket or par parenthesis

[07:30] will turn that special character into

[07:32] ordinary text so using it in practice

[07:35] we'll remove the effect of the

[07:36] parenthesis giving you snowy weather as

[07:39] pure text with no effect you can see an

[07:42] action here where we have the standard

[07:44] prompt snowy weather compared to snowy

[07:46] weather within the parenthesis and then

[07:48] the literal parenthesis represented as

[07:50] text alongside our prompt and you know

[07:53] it worked because the snow is less

[07:54] prominent now thinking back to the

[07:57] tokens which form chunks I described in

[07:59] the the beginning the break keyword in

[08:01] uppercase will fill the current chunks

[08:03] with padding characters and adding more

[08:06] text after break will start a new chunk

[08:09] I personally do not see any practical

[08:11] use for breaking up your chunks

[08:12] prematurely before hitting that 75 token

[08:15] limit but if you wanted to achieve this

[08:17] then this is your solution now the

[08:19] horizontal line is used to trigger

[08:22] alternation or a loop in prompts where

[08:25] words are broken up with horizontal

[08:26] lines and are given the chance to

[08:28] influence the generation repeatedly as

[08:30] stable diffusion Loops through the words

[08:33] within the square brackets so if we were

[08:35] to use long black hair brown dreadlocks

[08:38] orange curly hair as a prompt then the

[08:40] first step will be long black hair then

[08:43] Brown dreadlocks then orange curly hair

[08:46] this is another technique for

[08:47] controlling the types of generations you

[08:49] get during the generation process and

[08:52] you can see in action how this impacts

[08:54] the final result the CFG scale will

[08:57] determine how strongly the generated

[08:59] image should conform to the prompt you

[09:00] provide with lower values giving you

[09:03] more creative results extremely low or

[09:06] high values may give you unpredictable

[09:08] results so I tend to go between 5 and 12

[09:12] sometimes it can be useful to generate a

[09:14] few images through batch with a low CFG

[09:16] scale allowing you to get a more varied

[09:19] set of images and then running an image

[09:21] that you like through image to image

[09:23] with a higher CFG scale to make

[09:25] adjustments more closer to your prompt

[09:27] in these examples you can see a variety

[09:30] of CFG scales being used and I think the

[09:33] values 5 to9 are more accurate to The

[09:35] Prompt The Prompt Matrix is used to see

[09:38] what impact your individual prompts have

[09:41] on your generated image allowing you to

[09:43] remove unwanted or unimpactful prompts

[09:46] and keep the ones nearing you to the

[09:48] image you want in order to use the

[09:50] prompt Matrix start your prompt with the

[09:52] subject of your image and then follow up

[09:54] with the prompts you want to test with a

[09:56] horizontal line anything nested between

[09:59] or after a horizontal line will be put

[10:01] onto a matrix allowing you to see and

[10:04] compare the impacts of that prompt the

[10:06] more specific your prompt that the more

[10:09] consistent the results you will get

[10:10] across your images allowing you to

[10:12] identify the prompts which are causing

[10:14] issues by singling them out I'll be

[10:17] doing a separate video breaking down the

[10:19] prpt Matrix specifically as it's a

[10:21] useful tool which could do of its own

[10:23] easy to find dedicated video but in the

[10:26] meantime you can look at this image to

[10:28] see the comparison in action the prompts

[10:30] from file or textbox section will allow

[10:33] you to test multiple prompts at the same

[10:35] time either from the text box or from a

[10:38] file all you have to do is put your

[10:40] prompts in and use a line break to

[10:42] separate them then generate and you will

[10:45] get an image per line of your prompts

[10:47] allowing you to see the comparisons for

[10:49] each and I'll probably do a separate

[10:50] Deep dive video on this topic so it's

[10:53] easy to find here are some examples and

[10:56] while you look at these if you wanted to

[10:58] use a file then just put the prompts

[11:00] into a notepad on separate lines then

[11:02] drag the notepad file into the file

[11:04] section and it will copy the information

[11:06] over to disable diffusion for generating

[11:09] XYZ plot allows you to test and compare

[11:12] a range of variables on your generated

[11:15] images and make comparisons against

[11:17] those valuables such as a seed CFG scale

[11:21] and using prompt Sr or search and

[11:23] replace which allows you to replace your

[11:25] prompt with a different prompt during a

[11:28] generation to see the results now this

[11:31] script has a lot of options so I'll

[11:33] probably do a separate breakdown video

[11:36] of each option so you know what they do

[11:38] and you can see comparisons between each

[11:40] of the options but I've used this

[11:42] feature throughout the video to generate

[11:44] comparisons such as a CFG scale so

[11:47] hopefully you have a good indication of

[11:48] what it does but to wrap things up I

[11:51] hope this video has helped you to

[11:52] understand prompting a little bit more

[11:54] like the video before you leave the

[11:56] building and subscribe to be notified

[11:58] when my next video drops drops there's

[12:00] also a patreon for those who want to

[12:01] support this is bite-sized genius and I

[12:04] hope you enjoyed