ENGINEER Your Midjourney Prompts with SEEDS

Summary

本视频通过分离变量的科学方法,利用种子号测试 Midjourney 中各个词语对图像生成的影响,从而指导用户如何有效地设置提示词。它强调种子号的重要性,解释了如何利用种子号隔离并测试词语,找出哪些词语真正影响结果。视频还讲解了测试类似词组的更快方法,并提到种子号不能保证一致性。总的来说,它向用户展示了如何有效地设置提示词来产生理想的 AI 图像。

Takeaways

• 🔍 通过改变种子编号（seed number）来测试单个词语对生成结果的影响，保持其他变量不变。
• 🌱 中旅程（Mid Journey）有超过40亿个种子编号，每个提示可能产生超过40亿种结果。
• ✂️ 通过去除测试词语，并使用相同的种子编号重新生成，可以观察该词语对结果的影响。
• 🔬 发现某些词语（如'gai group'）对生成结果影响不大，而其他词语（如'graffiti inspired portraiture'）则有明显影响。
• 🎨 测试艺术风格或艺术家名称对生成结果的影响，使用种子编号确保测试的准确性。
• 👁️‍🗨️ 种子编号保持生成结果的形状相似性，而不是确保完全一致。
• 🔄 种子编号只对当前版本有效，不同版本使用相同的种子编号会产生完全不同的结果。
• 🎭 使用排列（permutations）测试相似词语，加速测试流程。
• ⚙️ 种子编号对于测试个别词语的重要性，使用户能够精确控制生成内容。
• 📐 种子编号不保证跨提示的一致性，相同种子编号的不同提示可能产生不同的人物或场景。

Q & A

种子号的主要功能是什么?

种子号的主要功能是隔离变量以测试提示词语的影响。通过保持种子号相同而改变词语,可以看到哪些词语对生成的图像有重大影响。

如何找到一个图像的种子号?

要找到一个图像的种子号,可以对图像作出反应并使用信封表情符号。Mid Journey 然后会回复种子号。

为什么种子号无法保证图像之间的一致性?

因为种子号只能保证同一版本和同一提示下的类似形状。它们无法跨版本或提示保证一致性。

使用排列组合测试多个词语的效果方法是什么?

使用波浪括号括起要测试的词语,然后在同一个种子号下生成所有排列组合。这可以更快速地比较不同词语的效果。

测试艺术家名字的重要性的方法是什么?

保持种子号和提示相同,仅更改是否包含艺术家的名字。如果结果差异很小,则说明艺术家的名字对结果影响不大。

种子号如何帮助优化提示?

通过测试不同的词语和条件,种子号可以帮助确定提示中最重要的部分,从而创建出更优秀的图像。

为什么需要测试提示词语?

因为提示中使用的每个词语都会影响结果。测试可以找出哪些词语真正重要,哪些可有可无。

如何快速生成多个提示的图像?

使用排列组合和同一个种子号可以同时生成多个相关提示的图像,加快测试过程。

不同 Mid Journey 版本的种子号效果一样吗?

不一样。同一个种子号在不同 Mid Journey 版本中会生成完全不同的图像。

获取种子号有什么技巧吗?

确保打开设置允许陌生人发私信,使用纯信封表情而不是其他表情,并确保是对图像作出反应而不是回复。

Outlines

00:00

🔬如何通过变量隔离法探索生成艺术中的关键词汇

这一段介绍了如何使用种子号来测试生成艺术提示中单个词汇的效果。通过保持种子号不变，改变提示中的其他词汇，可以观察到这些变化对生成结果的具体影响。示例中提到了'gai group'这一词汇对结果影响微小，而'graffiti inspired portraiture'这一短语对生成的面孔有明显影响，显示了不同词汇在艺术生成过程中的重要性。此外，还提到了通过排列组合测试不同词汇组合的方法，以及如何使用种子号来加速测试过程。

05:01

🎨艺术风格与表达的精确测试

第二段深入探讨了如何通过改变提示中的关键词汇来测试生成图片的差异。通过对比使用‘speeding’与‘rushing’等词汇生成的图像，说明了不同词汇可能导致的细微差别。此外，还展示了如何通过改变情绪词汇（如‘excited’与‘disappointed’）来改变图像中人物的表情和氛围。段落末尾提到了艺术家名称对生成结果的影响，并通过实验验证了某些情况下艺术家名称对结果影响不大，推荐在不确定时可以省略艺术家名称。

10:03

🌟种子号的局限性与高级提示技巧

这一部分讨论了种子号在保持生成结果一致性方面的局限性，并指出种子号对于不同版本的生成算法不通用。通过实例展示了即使在相同的种子号下，不同动作或场景的添加会产生完全不同的人物形象，强调了种子号不能保证完全的一致性。此外，还包括了一些关于如何有效使用种子号和排列组合来测试不同的提示词汇，以及如何通过设置来确保能收到种子号的技巧。最后，给出了一些建议，帮助观众更有效地利用种子号进行艺术生成测试。

Keywords

💡种子数字

视频中多次提到“种子数字”。它是生成每张图片的随机数,范围在1到40亿之间。使用相同的种子数字可以重现相似的图片。所以它是测试prompt中每个词的效果的关键。例如视频中去掉“急速”这个词后,大部分图片和保留这个词时很相似,说明它对结果影响不大。

💡测试

本视频的主要目的是介绍如何测试prompt中每个词的效果和重要性。方法就是保持种子数字不变,改变某个词,然后比较结果,看是否有明显不同。

💡prompt

prompt就是给AI的文字描述。视频详细解释了如何测试prompt中每个词的效果,找出对结果最重要的词。

💡排列组合

如果要测试多个类似的词,可以使用排列组合的方法,把要测试的词放在括号里,然后自动生成包含这些词的所有prompt组合。这样可以快速看到不同词的效果。

💡相似性

种子数字可以保证prompt中相同词语生成的图片有一定的相似性和连贯性,但不能保证完全一致。版本和模式的不同也会影响结果。

💡艺术风格

视频还演示了如何测试不同艺术家或艺术风格的名字是否对结果有影响。方法与测试单个词相同。

💡连贯性

种子数字只能保证prompt中相同词语有一定连贯性,不能保证完全一致。因为版本和模式的不同也会影响结果。

💡版本

同样的prompt和种子数字,在不同版本的Mid Journey中效果也会有差异。所以不能保证完全一致。

💡分步描述

视频详细步骤演示了如何进行关键词效果测试。第一步选prompt和种子数字,第二步去掉要测试的词,第三步比较结果。重复这三步可以找出对结果最重要的词语。

💡用例

视频通过多个用例演示了测试方法,如急速和其他词的比较,艺术风格中名字的必要性测试等。这些例子生动地解释了理论,增强了理解。

Highlights

第一重要内容摘要

第二关键内容摘要

第三主要内容摘要

第四突出内容摘要

第五值得注意内容摘要

第六意义深远内容摘要

第七引人注目内容摘要

第八重大内容摘要

第九不同寻常内容摘要

第十独特内容摘要

第十一关键内容摘要

第十二重要内容摘要

第十三值得一提内容摘要

第十四引人入胜内容摘要

第十五精华内容摘要

Transcripts

take a look at these pictures and this

prompt purple and yellow abstract

collage in the style of graffiti

inspired portraiture light cyan and

black mixed patterns gai group mundane

materials Serene faces close up I got

this prompt from the describe feature

but how are we supposed to know which of

these words actually matter to the bot

well it's actually very easy to find out

and it involves the most timeless

technique in science isolating variables

we are going to keep the strongest

variable the same while changing another

to see the effect and that strong

variable is the seed number you see mid

Journey has over 4 billion seeds there

are over 4 billion outcomes for each

prompt you write and the system chooses

randomly for you every time you generate

you can react to any of your images with

the envelope emoji and the bot will send

you a seed number to that image you

created this will help you recreate a

set of images if you run the exact same

prompt with the exact same seed number

on the exact same version now that's

great and all but the true value of

seeds is to help us test out the effect

of individual words and we can do that

by specifying a seed number ourselves by

adding dash dash seed space and then a

number between 1 and 4 billion at the

end of our prompt personally based on

this prompt I wanted to know what gai

group really meant to the generation so

I'm going to run that exact same prompt

again but I'm also going to choose a

seed number to run this test on it does

not matter what seed number you choose I

literally hit random numbers on the

keyboard so these images are what this

prompt creates on seed number 686 320

pretty cool pictures but that's not what

we're here for now we are going to run

this exact prompt minus the word we're

testing which is Guai group along with

the seed number we just used this is

what it will look like every word the

same except for that word we're testing

we keep the same seed number and these

are the results I'll flip back and forth

between them but keep your eye on number

four what do you notice it's the same

right I mean as close to similar as

possible I'm talking specifically about

the lines on the skin that appears in

both of the images now what that tells

me is that gai group really had no

effect on the end result of the prompt

the results of these Generations are

practically the same granted they are

not completely the same they're not

exactly twins but when you're testing

out certain words to see if there's a

difference what we want to see is a

difference and personally I don't think

I see that here at all however using

this prompt I can show you a group of

words that do make a difference we're

running the entire original prompt here

on seed number 942 424 and this is what

it creates a bunch of beautiful faces

now what I was really wondering about

was this part of the prompt in the style

of graffiti inspired portrait how much

did that phrase affect the generation so

we're going to run that test again by

excluding that part of the prompt but we

are going to keep the same seed number

and these are the results what do you

notice I think in my opinion there is a

big difference we are no longer getting

faces in 1 2 and four sure there are

some in number three but I think the

differences between the other parts of

the grid are big enough to say that that

phrase makes a difference in fact this

is a good example of what seeds actually

do if anything they can keep a similar

shape between prompts so take a look at

number one there are some Silhouettes

here as we move from the yellow to the

purple to the black to the purple to the

yellow over to the blue now let's look

at what happens in the original prompt

with the portraiture included we have

the same color setup but faces appear

here again that tells me that something

like portraiture really matters to the

bot it sort of brings out facial

features and elements like that isn't

that interesting isn't that fascinating

that's what prompt engineering is really

about in mid Journey testing individual

words while isolating a key variable

that variable is the seed number if you

have fast hours to spare and want to run

these tests at a quicker Pace you can

write your prompts using permutations

and I think this could be good for

testing similar batches of words to see

if what you choose really matters like a

miniature superhero speeding rushing

racing hurrying down a busy Street full

of angry dinosaurs then I chose a

completely random SE number we are going

to enclose the words we want to test

within those squiggly brackets and when

we hit enter it's going to ask us if we

want to imagine all of those prompts we

can hit show prompts to see if we made

any mistake while writing it everything

looks good here we can click yes and it

will generate all of them for us again

that really just helps us speed up the

flow of these tests now do you think

there's going to be a difference between

speeding and racing down a street I

don't know but this is how you test it

here's the miniature superhero speeding

this is sort of going to be our base

canvas and we're going to compare the

other words to this generation don't

worry about tiny small details the

pictures are always going to change when

you change the prompt we're more focused

on the goal the end result what vibe did

we get from using this word and what

vibe are we going to get from using the

other word I think if we notice one of

these pictures in the grid showing up

again there really is not much of a

difference so from speeding we'll go to

rushing is there a difference what do

you notice one and two are similar four

is slightly different but three is like

the exact same I'll do a quick little

before and after Da Da Da Da like look

at that when that happens you'll know

that there is no difference between

speeding and rushing it means the same

thing to Mid journey and that's how

you're going to engineer your prompts

effectively here's the word racing and

maybe you could argue we get something

slightly different I think number one is

cool number four has its own Vibe number

two is sort of similar to the other

number threes when it comes to something

like this it's going to be a personal

decision don't let this stop you from

using different adjectives in your

prompts but if you ever want to be sure

of what you did this is how you would

test it and finally here is hurrying I

think you could argue that these are

different enough to place them in their

own category I like number two a lot now

that's what the process will look like

when you're testing similar words but I

think this whole thing works even better

when you test a bigger dichotomy like

stock photography of an excited old man

when he sees who is at the door we lock

in the seed number and we get these

pictures pretty cool same seed number

same prompts except we're going to be

using the word disappointed instead of

excited and these are what we get same

similar groupings of the old man but the

feeling the expression has completely

changed and this is a fun way to see how

mid journei interprets the adjectives

you use I also recommend testing art

styles or artists in the same way

because there are over 4 billion seeds

the only way to test a prompt accurately

is to keep the seed the same if you

don't any differences you notice between

prompts could be entirely random like

one in four billion random and some of

you might not be comfortable using

artists in your prompt so this is how

you would test whether you really need

their name or not like here I found this

prompt and described a man and his

grandson standing by the side of a

street in the style of lazette model

intuitive gestures 1920s Yankee core

candid photojournalism childlike

innocence photo montage but it appears

as though LTE model is a real person an

Austrian born American photographer so

I'm going to run this prompt using her

name and a specific seed number this is

what we get running the same prompt

without her name on the same seed number

and this is what we get I think this

proves that her name does not really

matter that much to the rest of the

prompt and therefore you might be more

comfortable leaving her name out of it

that's the power of seeds that's how we

can really know what we're doing

permutations and Seed testing works

great when you're trying to prompt for a

person like here a Fierce Female warrior

prepares for a battle on a beach the

weather is overcast I want to see The

Prompt just like that but I also want to

see what the prompt looks like when I

adjust her hairstyle so I start the

permutation with a squiggly bracket I

use a comma and I'm going to say the

female Warrior has then I'm going to

place another open squiggly bracket and

I'm going to list my hairstyles french

braided hair pigtail hair pink and blue

hair red hair green hair burning flaming

sparkling fire y hair and finally

cartoon anime hair I'm going to close

that with a squiggly bracket and I'm

going to close the whole thing with

another squiggly bracket that's a little

Advanced way of using permutations don't

worry about that too much for now just

remember that what you're testing goes

inside of squiggly brackets and then

we're going to test it on seed

364-7974 here's the first foundational

prompt Fierce Female warrior weather is

overcast cool pictures sure why not but

now we get into the tests the Female

warrior has French braided hair I think

these look pretty amazing I love number

two a bunch it's that specific hairstyle

same promp same seed number except we

have the warrior has pink and blue hair

these are pretty cool here's pigtail

hair and I don't think I worded that

quite correctly the results weren't

great it's not what I was looking for

here it is with green hair I like these

a lot as well red hair looks great

cartoon anime hair didn't seem to do

anything I'll make a note of that now I

know that's not the right word words to

use whatever I was describing and then

we have burning flaming sparkling fiery

hair and these are great number two

might be a little too much I like the

idea though right now there are a few

more quick notes about seeds in case

you're wondering cinematic still of a

dog wizard casting a spell through its

mouth seed 935 893 pretty cool pictures

what I want to point out is that

rrolling this prompt will give you an

entirely new seed number even if you

specify a seed number up here you you

see the seed number didn't show up in

the prompt even though that it's showing

we made it from the previous seed number

now we're going to react to this image

with the envelope emoji and if you've

done this enough times you can actually

just rightclick on an image and the

envelope Emoji should be there for you

so we see that that grid was created

with this new seed number however what

you need to know is that the seed of an

upscale is the same seed as the grid

meaning that four different images will

have the same seed number and that's why

seeds are not the answer to consistency

across prompts portrait of a female

crime scene investigator with platinum

blonde hair and a dark green lab coat

seed

72497 these are the women it generates

if we keep that heart of the prompt and

add a new action with the same seed

number we are not going to get these

same people same start of the prompt but

then I added the female is trying to

plead her innocence with a robot police

officer same seed number and now these

are different women seeds do not

guarantee consistency they barely

guarantee the same shape of what you're

going for don't let anyone tell you that

seeds are the answer to consistency I

just want to be clear on that seeds are

also only applicable to their current

version same prompt same seeds in

different versions will generate

completely different images the

degeneration of society version 6 seed

2854 239 cool right great here's the

same prompt in 5.2 5.0 4 and three this

seed number does not matter when you're

changing versions of mid Journey seeds

also act a little funny when you're

using multi- prompting but that's a

little complicated for this video so

we'll save that for another time if

you're having trouble with any of this I

have some quick troubleshooting advice

for you if you want to find your seed

numbers make sure you go into your

settings and allow DMS from strangers so

the bot can message you furthermore make

sure you are reacting to an image not

replying to it this is reacting it'll

bring up this list replying will bring

you down here to the prompt box we do

not want that and the last thing make

sure you are using the plain envelope

Emoji these other ones will not work

running these tests are great but if you

want to know the best Cadence for your

prompt like how long should it be and

where should you add details I think you

should watch this video right here I

hope you're doing well take care and

I'll see you next time peace