Underused Midjourney v5 Prompt Commands :: How to use Text Weight and Image Weight
ฝัง
- เผยแพร่เมื่อ 28 มี.ค. 2023
- Today we're taking a look at two underused Midjourney prompt commands and learning how and why you should be using Text and Image Weights in your Midjourney prompts!
I've noticed some confusion with how Image and Text Weights works in prompts, and I think that might be due to overly complex tutorials, so this is a very simple and easy to understand video explaining the concepts.
Here are links to videos mentioned:
Leonardo.AI Pose to Image: • Leonardo.AI - Prompts ...
Spielberg's Lost “Dark Tower” Film: • Spielberg's Lost “Dark...
AFFILATE LINKS:
Camera: amzn.to/3yXMDY2
Microphone: amzn.to/3K1jIZm
Audio Interface: amzn.to/3lDX9kf
Coffee: amzn.to/3JZuBeq
Follow Me on Twitter: / theomediaai
-------------------------------------------------
Thanks for watching Theoretically Media! I cover a large range of topics here on the Creative AI space, Technology, Tutorials, and Reviews! Please enjoy your time here, and subscribe!
Your comments mean a LOT to me, and I read and try to respond to every one of them, so please do drop any thoughts, suggestions, questions, or topic requests!
. - วิทยาศาสตร์และเทคโนโลยี
photobashing seems like an incredibly powerful technique
It really is MJ's superpower. I should probably do a dedicated video on it next week.
@@TheoreticallyMedia Yes please! 👏
Great tutorial! I've watched it over and over again to really grasp the underlying principles. This must be one of the best videos I've seen on MJ prompting. Please keep them coming :)
Will do!! I'm actually working on all the Dash Dash commands as we speak!
Tim :) Thanks again for another insightful video. I love your style of teaching and your honesty! It's great to hear about your whole process not just the "cherry picked" elements that show case what you're talking about and teaching, that helps me in my expectations and my process.
Thanks so much, Graeme! Yeah, I think it's important to show the flubs. I don't know how many times I've watched a YT tutorial on, well-- really any subject, and my work ended up a mile off.
Weirdly, I took inspiration to be honest about this stuff from the TH-cam guitar community. There are tons of videos of guitar players doing these amazing speed licks, and it can be both inspiring and discouraging to watch them. But for a minute, there was a trend of these players showing how many times it took them to get that "perfect" take.
Of course, that trend ended and now we're back to monster speed runs that look effortless.
...oh, and everyone has amazing hair in them too. haha.
Thanks for your video Tim. Very very and clearly spoken, easy to follow along 👍
Fantastic to hear! Thank you!!
Another excellent video; thank you. Hoping to see the subscriber count rocket up as you're one of the smarter folks doing these videos.
Oh, dude-- thank you so much! Ha, I'm hoping that sub count rises as well!
Very useful techniques. Thought I was the only one who couldn’t get the machine to follow neg prompts. I feel better.👏🌟👍
Glad to hear! And yup, I'm trying to make these tutorials as transparent as possible-- I don't want to give the illusion that everything is 10/10 with Midjourney all the time!
Loved the vid! Thanks for clarifying what works in v5.
Fantastic! Really makes my day to hear that!
Great tips! I’ve been trying to get in the habit of weighting my prompts more often
It can work wonders for honing in on things! I'm just as guilty, but have been adding it in more and more lately. I think it really helps out a lot after you run your first prompt and start playing around with the output to get closer to what you're looking for.
Thanks for the informative way you described everything easily explained
Thanks so much! I really wanted to uncomplicate the whole thing about weighting, glad to hear it came through!
thanks for sharing. Subscribed!!!
Very useful techniques. Thanks.
No problem! Happy to hear!
At about 3:05 you say there is a space after the colon-colon before the weighted number and after the weighted number. But in all of your prompts there aren’t any spaces between the colons and the number, which is how I believe it works. Just thought I’d point it out. Great video, and great channel, very useful. Thanks for all your work.
Oh man, I missed that up! Good catch, I’ll put it in the description! And indeed: no space! Thank you!
Thanks for this, I was puzzling this out.
Great tips, thank you. I just rang the bell and will be looking forward to future tutorials.
Fantastic to hear, Frank! Should have something new tomorrow!
Thanks for the tutorial! Great help!
Happy to hear!!
The cupcake and cup cake images are awesome!!!🥳
Haha, I love how AI can do so much, but still be fooled by something as simple as a Cupcake. Hot Dog is another one. Is it a Dog that is Hot?
You just took me to the next level. Thank you
Excellent to hear!
Thanks for the tricks!!
Glad to help!!
You are super knowledgeable and that’s clear from your videos. As a newcomer to MidJourney, it’s hard to retain much of this information with your teaching style. It’s hard to define, but your approach seems to be presupposing we already know the basics.
To be honest, that’s always the problem with TH-cam tutorial channels, really across any subject. Older viewers don’t want repetitive material, newer viewers need fundamentals.
I’ve long thought that any thing you learn on TH-cam has some knowledge gaps. I ran into this a lot on the Music Theory side of things. That said, I have been meaning to do a series on basics soon, and your comment definitely pushed me into doing it sooner!
I kinda agree. New to it myself and there seems to be no place where a current bible of commands is listed.
@@grandmasoffgridliving @logosgaming9228
That's a good point. I'll put together a starters video for you all next week. Seems overdue.
I'll also add in a command bible, that does seem worth having in print...
Although, I'll say, it might get quickly outdated, as the commands are constantly changing!
@TheoreticallyMedia it's much easier for us to add some notes to an already compiled up to date Bible of commands. You could add cheat sheets to newer ones and if need be a longer update to complex ones in pdf files. We on our end could add them to a folder or if we have good pdf software just add it to y9ur bible.
Thanks for this - as many people come to the channel in different spots, take a moment to explain each step even if you've done it before.
Thanks Jon! And yeah, I know-- it's always the issue with TH-cam videos, it's tough to structure, since you never know which video someone is going to stumble across first. But, you also can't be overly repetitious to returning viewers.
That said, I am putting together a starter guide in the near future. I've been meaning to do that!
@@TheoreticallyMedia I really like the PDF that you put together for the other video it was very stylish and informative.
@@jonmiles8091 Excellent! I'm going to make that a recurring thing for future videos!
Useful information, clarified a couple of issues I was having.
excellent to hear! I gotta dig through your channel as well, I've been thinking about adding Amino Acids into my pre-workout stack. Did an AA supplement years ago and felt like I was getting some good benefits out of it!
@@TheoreticallyMedia I have specialized in amino acid therapy for 35 years. Body builders understand this more than most people.
Fantastic tips bro
Glad you like them!
very useful information!, i would be happy to subscribe
Thank you! Looking forward to seeing you again!
Hey Tim, FWI, --iw does go all the way down to zero if you just want a little hint of the image reference. However once you hit zero it becomes a 1 again (or more probably --iw gets ignored and it goes back to the default of 1) . So --iw .1 is perfectly fine. I believe the docs have been updated to reflect this.
I think the secret sauce in doing these projects is def using image referencing and doing exactly what you did - photoshop the hell out of it. I do this often as well. Did you try prompting the hairstyle with a larger weight to the hairstyle to remove the hat? If the prompt is quite long for the image weight, I try to break up that part of the prompt and weight each bit as well. It's all trial and error, and as soon as I think I figured out the code - it takes a left turn. haha. But this is great content and you break it down nicely. Nice work. Great channel and always fun to see your videos.
So, I tried "Sweaty hair" but no love there. I was going to try something like repeating "hair" through the prompt, which I hear sometimes works...I think maybe it's like buying multiple lottery tickets for the token distribution-- but, I kind of gave up and just moved into photobashing. Which is actually the fun part, anyhow!
@@TheoreticallyMedia haha, yeah - I find the whole process is sometimes more fun than the actual result.
This is a really great video. Love it! Please How do I photobash?
Can you show us how to generate LOAB?
Thank you!
You’re welcome! Glad you enjoyed it!
Thanks for the good videos. I’m on again off again with MJ. It just doesn’t give enough control, even with weights. I think if MJ adds some in/outpainting and controlnet / open pose capabilities, it will reign supreme.
I hear rumors that is in the works. I sort of bounce back and forth between Leonardo and Midjourney, as Leo does offer a lot more in terms of control-- but yeah, I get it. As soon as MJ has some of those more advanced SD features, it's going to be a serious powerhouse.
thank you Tim
Welcome!!
Love the photo bashing. Surprisingly, it works well.
I honestly think it's the real superpower of Midjourney.
True
I've had challenges with combining a custom image and text prompt. I really struggled and wasn't satisfied. I'll give the photo bashing idea a try and see if that helps. Thanks again for another useful tutorial.
Please let me know how Photobashing works out for you! I think I'll do a dedicated tutorial on it next week. And thanks so much for the watch/comment!!
@@TheoreticallyMedia A dedicated tutorial on photobashing would be really helpful. Thanks!
@@TheoreticallyMedia not sure if you need an example for your Photobashing video but I'm having an impossible time getting Midjourney to create what I'm looking for: "A photo of Robert Oppenheimer wearing welding googles with the reflection of the nuclear mushroom cloud reflecting from the glasses." There's an upcoming movie about Oppenheimer so it might be a good crossover topic for a vid. Just an idea. Figured I'd throw it out there. Thanks!
@@Justin_Hikes Oh cool! I'll keep that in mind! I'm hoping to have that done this week (new video, but not on Photobashing tomorrow)-- But, the kids were sick today, so that's thrown a lot of this week's schedule off track.
Hmmm, maybe ChatGPT can take care of them while I work?
You know you can just add "no hat" to the prompt. Always works for me
This video is a little old, and I’ll say that -no has improved a lot since then, I’d still say I’m 50/50 on No. I just ran into a weird one where I did a Zoom Out and MJ suddenly decided to put sunglasses on the character! I -no Sunglasses, but the result came out super googly eyed. Haha, kinda funny though!
@TheoreticallyMedia hahah, yeah it's been doing some wacky things with zoom! I've barely started messing with that but it's exciting
I find it tricky with text weighting when you want specific things in the image, as MJ will add other things you need to try to get rid of. Changing the aspect will then ratio throw it out of whack again.
Reordering, weighting, “no”, all techniques to try and convince MJ to get what you’re looking for. But, in some ways I’ve started to think of Midge as a collaborator, rather than a servant.
If I was (and I have) art directing, I would always give the artist agency with the end result. I mean, they can’t go too outside the box, but allowing for some creative flexibility sometimes gets you amazing results.
But haha. Yeah; that aspect ratio thing, that’s a quick way of screwing EVERYTHING up!
Thank you for a great video.
Is there a way to make it exact. For example if you say you want a picture of a girl sitting in her yard playing with 5 cats.
More often than not I see 6-10 cats in the generated images.
I have trouble getting exact numbers on the images.
Great info, thanks! Question about an important distinction, though… when you add a :: are you only weighting? The one word that it is connected to? Or are you weighting all of the words as a group that follow any :: before it? In other words “rough beard::2“ isn’t actually weighting “rough beard“, it’s only weighting “beard“, correct? So then is “mystical forest::2 rough beard::3“ weighting two phrases? Or only the words, forest and beard? Thanks!
Super good question, and I actually had to digging around for that answer. I've always presumed that when you add the :: MJ would "read" the phrase, at least up until the previous comma.
Documentation has an example as follows:
A twilight cafe:: Two friends talk about coffee
Wherein "A twilight Cafe" is the opener of the prompt, and the :: reads the whole phrase. So, I DO believe that "Rough Beard" would be the same.
But, we can apparently control the overall roughness of the beard, with the following:
An Old Samurai, Holding a Katana, With a Rough Beard, Standing in a Mystical Forest::Rough::2
The whole weighting and multiprompting thing is pretty zany. I'll read through all the documentation this weekend and do a video for you next week. It feels like this is worthy of a deep dive.
What I'm wondering about is if the text weight only applies to the last word or all the words in between the numbers? Like your example: an old samurai::1 rough beard::1 holding a katana::4 in a mystical forest::2 - is the text weight only applying to samurai, beard, katana and forest? I´ve tried different prompts with text weight, and I'm not sure myself. Maybe there should be a system where everything you wrote inside brackets would be affected by the text weight, like [an old samurai]::1 [rough beard]::1 This way you could weigh a whole concept instead of maybe just a keyword.
I believe the colon weight applies to everything that precedes, up to a comma. So technically, old samurai::3 should apply to both Old and Samurai.
You could do Old::3 Samurai::4 holding a katana::2
But I think MJ would ignore Old, since nothing precedes it.
In theory, that’s supposed to be how it works. I do suspect that the weighting has more to do with subjects than adjectives.
Man, your idea on bracketing would be great!
I'm new to your channel. I want to start using a paid subscription of midjourney for my design experiments from magical prompts to risograph etc. Where should I have to start. Thanks
I’ll say, what I need to do is make a “MJ for Starters” video. I’ve been meaning to do that! I’ll see if I can sneak that in this week.
So, as far as prompting advice, you can check out the video I did on Cinematic Images, although the style is based around movie images the prompt logic can apply to anything. So, in the case of your Risograph, you could try something like:
/imagine Risograph Poster, (subject/description), (colors), (texture) (aspect ratio).
Let me know when you get started and I’ll see if I can help dial you in a bit!
Thank you. Love the Tut. One question though. If I want to put more weight in a group of words. Let say for instance "Full Body Shot". How do I place the weight? After each word? Only after "Shot"? But in this case the weight will affect only this word, I guess. What is the correct way? Thank you in advance for a clear explanation.
Only after Shot. It SHOULD take everything from the preceding comma. So if you had: /imagine A Samurai, Full Body Shot::2 In a forest, etc etc...
It should take Full Body Shot as the weighted term referencing back to Samurai. But, it will obviously not weight the forest.
Admittedly, where I find the most success with full body shots is when you work in a 2:3 aspect ratio. What I like to do is generate characters against a white background in 2:3 and then photobash them into backgrounds, then run it all through MJ again. You can check out the technique here: th-cam.com/video/qeRVBKi1QMQ/w-d-xo.html
@@TheoreticallyMedia Thank you so much for taking the time to help. I figured it out after reading some blogs on the subject. I'll check your video, of course, as I am very interested.
Very Good! What can we do with these type of images?
I mean, really-- just about anything you can imagine! I've used them for album cover art, storyboard presentations-- if you're a fan of the film Alien, I made a fictional documentary about John Carpenter making a precursor to it in 1976. All of the images were made in Midjourney! You can check it out here if interested:
th-cam.com/video/BbeQDnBSjD8/w-d-xo.html
I was always pretty proud of that one!
@@TheoreticallyMedia All the images in the video came from MidJourney?
@@downloadmusic4free1 indeed they did!
i want someone to make a handful of images for me. i have a small channel doing fairy tales and with learning video editing finding stories, its a pain! i already spent hours trying to figure these text to image programs out, now my trial is up :(. Also, it was interesting looking at your channel and seeing where your videos kinda took off! like the Midjourney vs Leonardo BOOM lol. i bet that opened your eyes and made you think? Anyway, well done mate (think i might have to ditch fairy tales haha)
If your MJ trial is over, you might want to look a little deeper into Leonardo, since they offer a more generous free tier, and one that resets daily. I'm actually going to do a "getting started with Leo" tutorial next week, and I'd love for you to check it out!
And yeah, haha-- the channel is doing fairly well now! I think I've found a bit of a groove! I think the key is really experimenting. I did private a bunch of older videos recently, but as a quick rundown:
It started as a Music Tutorial/Recording Channel
I focused down on Guitar Recording
I tried Weekly Guitar News
But the real turn happened around November 22, when I did a video on AI Music. I was a little ahead of the curve on it. but I felt VERY strongly that 2023 was going to be the year of AI.
I did some more experimentation utilizing AI tools, but still in music (re: Morgan Freeman Punk Rock Song), and then pivoted into the "Lost Film" series.
MJ vs Leo was totally the one that "blew up" but what's funny about that is it came off a pretty major failure. The Ai-Daption one, which took me over a week. Rough estimate, I'd say I put in 60 hours on that video. It has 564 views. haha.
Even today, I still have videos that do well, and others that bomb. But just keep experimenting with different types of videos. You are learning more skills every time, and they can be applied to future videos!
@@TheoreticallyMedia thanks for the rundown mate. yeah ok, i'll keep an eye out for that one next week. yess this youtube thing can be challenging at times with all the work that goes into a 10 minute video, it really makes you appreciate other youtubers once you had a go yourself 😉anyway, keep up the good work and thanks for your reply and advice. 😊😊
Amazing video. I always get errors in MJ when trying negative prompts. It won't generate anything. Any tips?
I hear ya. The dash dash no can be super hit or mess, depending on how specific you’re trying to get. Sometimes when I’m looking for a “no” the best thing to do is just “show” MJ. Did you happen to catch my video on Photobashing?
th-cam.com/video/qeRVBKi1QMQ/w-d-xo.html
The idea is: crib something up that you want to see, and provide it to MJ as a reference.
That said, there are some things that MJ just had a lot of trouble with, like a Candy Cane with no stripes (as a hypothetical example) and no matter how hard you try, it’s just not going to happen.
@@TheoreticallyMedia Thanks yeah I acutally get a red error. Like it doesn't allow ::-5. It refuses.
Side question, does insight have an option to pay for unlimited generations? The daily limit made me sad.
Hi, can you help me with this problem
So, how do I only take "the outfit" from uploaded image but preserved the every details to another subject and surroundings?
Do you have tips?
I hope you answer my question by give me a prompt example.
Thank You.
To be honest, that's a pretty tough one. MJ randomizes quite a bit, so your best bet would be to upload a reference image, and then in your prompt hit the outfit details hard. Like, let's say it's a woman in a green dress as the ref image. The prompt could be: "Red Haired Woman, in a Green Dress (details about the dress), (The rest of the Prompt)"
It isn't going to be perfect, but it might get you there. I've got a video coming out today that might shed some more light on the subject!
@@TheoreticallyMedia thank you, wait for the video, it will be game changing.
Also, how to do an avatar edit of your photo?
Let's say you took a selfie from your iphone but want to enhance it to be as if captured by hasselblad, I'm trying to do this for hours even with image weights with double collons etc.
Any tips?
So the lower the number the more emphasis that object gets and the higher the number the less it gets, is that right?
Sorry, no: with weighting the higher the number, the more emphasis the bot will put on that word/phrase
I can see concept artist like myself using photo bashing like this in all their work in the future
Yup. I’ve used it to design background images and widgets for pitch decks. There is still a bunch of cleanup work to do, but I’d say on a whole, it saves me hours of time per project,
I also think MJ is a great tool to quickly iterate on ideas.
@@TheoreticallyMedia Thanks for the demonstration! I've been trying out Stable Diffusion and creating my own models too -- which has been another way to create cinematic images using movies I like the look of.
Do you know what is there is any difference between parameter --no ''hat'' and negative value on text prompt? I just learned about this negative value prompting from the video so far I was using --no parameter
According to what I've read, if you do ::0 (colon, colon, zero) that is basically saying "no." I've had a lot of trouble with MJ ignoring "no," and I think the weighted command tends to work a little better. BUT-- as you saw in the hat example...I mean, sometimes there's just no getting around it.
If you weight the negative prompt, like ::-3 (colon, colon, minus three) it is supposed to be telling MJ to actively avoid that keyword. But I tend to find that works better for mundane subjects, like clouds, or trees-- and not so much with very specific things, like characters or stupid hats that characters are wearing, ha!
can you weigh phrases? like (a man sitting in the bottom left in chains)::5?
Just found your channel and you explain really well. Subbed. However, may I kindly advise for you to balance your voiceover and the music a bit more, favouring your voice. The music is a bit louder and it fights your voice level for attention. Otherwise, great work.
Oh that’s great to hear! Yeah, this is an older video. If you check out the more current ones, I’ve gotten the background music sorted out a bit more!
Happy to have you on board!
@@TheoreticallyMedia awesome. Thanks for the heads up. Will definitely check out the other videos. Cheers :)
how come my midjourney doesn't produce photorealistic stuff? I typed in the exact same prompt for the samurai one and it's more of a painting than photo
Hmmm, try “/imagine cinematic still, etc” that usually gets me a photo or film look.
Send over a prompt, and when I have a minute I’ll try it out
This is all v5? Other then the one you siaid was v4. So it doesn't matter if you write a paragraph long prompt? Describing every detail? Because I've noticed a lot of info gets ignored.
Yup! All v5! I’m hoping to have another video up in a day or so that goes into a little more detail on Tokens. I’m starting to realize it is one of those things that isn’t talked about too much (mostly because there isn’t a ton of info on it), and I think it might be worth discussing.
Short answer: longer prompts might actually be worse for your image!
v5 has no style unless its actual man made objects. V4 can add everything just not as photorealistic.👌
MJ is really good but my favourite text to image generator is Blue Willow. I produced some incredible artwork with Blue Willow. Blue Willow recently launched V3 with a brand new model upgrade. Now I got more better-quality images with more aesthetic outputs. More importantly it's completely free
When you gave jay lee a weight of one, you forgot to put the colon before the jay lee as well, that's why you didn't notice any difference.
Ah, good lookin' out!! I need a prompt proofreader!
@@TheoreticallyMedia you can hire me if it's crucial 👌😊
You didn't mention text weight..did I miss something :)
that's basically the Keyword:: part. Although I did mistakenly call for a space there. You do not need a space-- it is (keyword)(no space)(colon colon)(number) (no comma) (space) (next keyword)
What does - - ar 16:9
--ar (no spaces) is telling Midjourney to use an Aspect Ratio. Which are the dimensions of an image. 16:9 is like a widescreen TV. 2:3 would be a vertical photograph. 3:2 is like a horizontal photograph.
You can check out this video, where I go into a lot more detail on it: th-cam.com/video/wqC9---Xmw8/w-d-xo.html
Actors 80 and 90, you have to write young , I do that too.
Haha, yup! No one wants to see 85 year old Clint Eastwood in an action movie! Another good trick I picked up is to name the age, like "Young Clint Eastwood, age 24" since, with an actor that has been around as long as he has, "young" is kind of a lot of years! haha!
holy shit dude u gotta make the soundtrack lower than your voice
Ha! Will do! Actually, if you have a sec, check out the latest one on Photobashing and let me know if the music is at a reasonable level.
I think it must be intentional. Get the user to use as many credits as possible trying to get one simple task the way they want. Really not pleased with this RNG-esque prompt weighting. This service is way too expensive to have to gamble your way to a decent image.
So what about this message form midjourney ... "Custom image weights are not compatible with --version 5" ( ::2 set to 2) How do you do image weight in V5 then ??
Dash dash iw for image weights. The text weights should work, although I’ll say that I have run into some errors here and there , I think while they’re working on some backend stuff. But that usually clears up in an hour,
Can you copy paste one of the prompts you used and I’ll try it?
@@TheoreticallyMedia Thanks you, . after the --iw what ? space and value ? bouble colons ? how exactly should it be ?
And you turned Clint Eastwood into Christian Slater.
Ha! They DO kinda look alike! I never really noticed that!