PUBLIC SERVICE ANNOUNCEMENT (just for context) Some (very few) people seem to be upset by the way I've set up this experiment and they've identified errors in my interpretations. And that's fine! Please correct me whenever I am wrong. I'm just learning myself 🙂It doesn't matter how I set things up. There's always going to be an expert that will find something that I've done wrong. I'm sorry, but I just don't worry about the small stuff in life. In many of these cases, they are honing in on very specific statements in the video as if they entire video was only about those specifics. And that's simply not the case. They are absolutely right in their assessment of those particular examples, however, they are choosing to ignore how the majority among you are actually writing their prompts. Yes, it might be mostly wrong. But so what! Yes, if you have advanced knowledge about those these AI models work, then you would probably not write your prompts with just a few non-sensical keywords. Or maybe you would. Who knows? It's all about trial and error, isn't it? I've gotten awesome images with just a few words. I don't see the problem in that. 🤷♂ The main message of this video is that the use of commas (!!) is not as relevant or powerful as many believe. This statement is absolutely true, no matter how you write your prompts. To claim that I am spreading fallacies isn't a fair representation of what this video does. Watch it and make up your own mind. Do your own tests. Verify or falsify it. Go crazy! 😂 My objective is to provide helpful insights to the 80-90% of you out there, who are currently struggling to create the images you want or need. It is NOT my objective to tell you how you MUST write your prompts or that you MUST absolutely conform to some sort of a convention that a small group of uber-experts believe to be the only true way. There's a reason that even at the very top echelons of science, people hold radically different views and routinely get upset about positions held by their peers. I don't really have the energy for that, to be honest. 😜 Let's not nitpick about inconsequential details. I'm a pragmatist. I'm a fan of the pareto principle. 20% of the effort is going to get you 80% of th way. The last 20% is going to cost you 80% of effort. I don't really see the point in that. You are here to get results, not to debate the eternal granularity of a natural language processing AI. ✌🤗❤
I certainly don't want to go all poo-poo on your methodologies for sure, but I also think taking them as gospel is wrong as well. There is so much that goes on BEHIND the scene with MJ that we just don't know about. It isn't like Stable Diffusion where the model gets exactly what you typed, and then you get the exact output. With MJ what you enter is further processed by MJ before even being sent to the model, and then the output of the model is further processed before it is presented to you. Furthermore we have all see how MJ tends to "lock in" features in subsequent generations as if there is bleed through from one to the other, regardless of the prompt. Now all that said, I thought I would be able to prove how this methodology fell short by doing essentially the same tests in the OPPOSITE order, IE starting with a prompt with 0 spaces, just one huge word, and expanding out from there. But I'll admit when I am wrong and by golly I was wrong because right from the very first smashed together prompt it actually generated roughly the same thing. I don't have a clue HOW but it did.
That last bit made me chuckle 😂 Listen, I have no intentions of portraying anything as gospel. Quite the contrary. I'm showing in public what I've done and how I perceive the results. Everyone should prompt whichever way they want and what gives them results. Let's s face it. No one knows what's really going on aynway. 🙂
Hi Christian, I found your content fascinating. Love that it opens up a focused discussion on the language learning curve of Midjourney AI or rather any creative AI. I commented separately on my take of the breakdown of syntax interaction in the video and would be interested in what you think. I think it would be great to see an even more in depth discovery sesh on this topic as I feel it is the key to unlocking potentially more control over the results produced.
Yeah the problem is more in how your title and thumbnail of your video conveyed a certain authority to the matter. It's nice that you tested some stuff out, but you still don't know what affects what and so there is no actual strength in this video. There are a lot of experts that can teach you through some videos so maybe do a little more research, then come back and kill this argument; or be butt hurt that people pointed out your general lack of knowledge on this subject. Either way in order to not actually sweat the small stuff, you shouldn't worry about dignifying yourself at all.
I think it would be even more amazing if we could communicate with the AI using layers. For instance, we could assign the first layer as the background, the second layer as the person, the third layer as the object they're holding, and the fourth layer as an additional object in motion. It would be incredible if we could then text our layered image ideas into a Photoshop project, manipulate the newly generated AI layers, and have full control over the final product. This could revolutionize the way artists create original art and provide a new level of creativity. I believe Adobe should seriously consider implementing this feature into Photoshop. Imagine creating a new layer, selecting the size, and having the AI ask us what we want to create on that layer. Then, we could manipulate each object in the layer, resizing and moving them as we please. This is a unique idea, and I can't wait to see it come to life!
One thing I've noticed about MJ is that if you keep using the same seed, the design is somewhat locked. So that it may ignore the prompt changes in favor of keeping the design mostly the same.
So in this case would you use --seed over --sameseed so that you get the variations? Or have you tried that yet? I feel like using just seed would get you what you want. Use the same number though. Hope this helps!
Nice catch! I wonder if a seed itself could influence variability. Statistically there would be noise generated with clusters of pixels at certain coordinates. At that point resolution also would start playing a factor
As a retired English teacher and editor, I'm having a blast experimenting with prompts in AI. I made a bumper sticker a long time ago: Words Matter (not a political statement but a professional statement). This is looking to be quite applicable to AI prompts. Like weighting, putting in 200 descriptive words has both diminishing returns and increased, unintended consequences. Being intentional about the words you include and looking at the relationship of words to each other (not order) within the prompt does seem to matter. As mentioned in the video, weighting words or phrases seems to take away (possibly unweight) other descriptors. The words Barbarian and Warrior, for example are likely to produce results where some images have armor, and others do not. Weighting (increase) Barbarian and Warrior may increase armor, but it still may not, and it may come at the cost of diminishing other traits (unweighting). I'm finding better results using words in such cases. Woman, warrior, leather armor with fine detail, etc., will increase armor in the results, add more detail to the armor, but won't water down the other descriptors in the prompt. Notice in my example that "leather armor with fine detail" is a phrase set off by comas. This seems to give the aI more to work with on the armor and keeps "fine detail" with the armor. Another example I discovered was on an image of a stone cottage in the woods. I had "wooden bench" in the original prompt, and most of the image results had a wooden bench up against the front of the house, usually next to the door. But when I edited the prompt and put the bench in a phrase, "wooden bench by a campfire", the bench always appeared next to the campfire and not against the cottage. This may seem obvious, but being intentional about word and phrase use in AI prompts helps you better predict the results. There is so much to learn about prompts from a language perspective, but so interesting to try and learn how the AI interprets what we write.
I totally agree. Being intentional about words is very underrated. Unfortunately, the fact that you can enter a bunch of random words and still produce a very appealing image is distracting from the fact that the art lies wtihin the ability to produce images that resemble exactly what you are aiming for. It's a bit like at the pool table. Sure, you can always get a lucky shot and it'll look totally cool. But the true masters are the one who call a shot and also make it!
In 5-10 years, schools will adopt all this as part of their curriculum. Having said that, I want to thank you for putting all this out here, 5-10 years early. I think I'll watch all your videos in just 1 long sitting. This one is already vid #7 for me.
I have watched all your videos and it's super fantastic! I've learned a lot of things and I would definitely recommend this to anyone who tried midjourney but does not know how to get around with it and how to maximize specific parameters. V4 prompt is nuanced so nobody can claim what is the right or wrong prompting. Please ignore the people who are spreading negative vibes, all of us are on the learning process. The only things that matter are you are consistent and continuously experimenting. I do love your approach so keep producing content like this as it is very helpful for sure! Feel free to experiment and prompt in your own way and I respect your structure and approach. This really proves that the knowledge shared really enhance my Midjourney prompting a lot! Thank you very much.
I can’t agree more!!! I learn the similar way. Try and learn. I definitely will try your prompt experiment and see how it turns out. We are learning through AI and AI learn from us too.
Your word order test didn't change the result because it was just a list of nouns and adjectives. If you had used a verb, the order would have mattered. For example, "cat eating a sandwich" and "sandwich eating a cat" contain exactly the same words in a different order, but give very different results!
Yep, I've discussed this with many others in the comments as well. This is a flaw, or at least a missing comparison, in the tests. Real phrases do get impacted. However, many are also underestimating how many people do not use phrases in Midjourney and mainly work with keywords. I'm not saying that's how they should be doing it, but it's what many are doing.
Also, the default weight is 1 so :: is 1. And --no is the same as ::-0.5. So "large office... worker suit --no glasses" is "large office... worker suit::1 glasses::-0.5". But check yourself that other punctuation does matter. Compare "large office... worker suit" vs. "large office worker... suit". English uses nouns as subjects or attributives to subjects, so "office worker" and "office, worker" are two different things in English, and you can hint that to MJ via normal punctuation, bit it doesn't matter which signs you use. Except :: which really has special meaning, and I propose to think of it as the equivalent of a paragraph marker in human text.
For testing grammar, you should make a sentence where the meaning is different depending on a comma. Use descriptions such that word order matters in English! Rather than just a bag of nouns and adjectives that you can figure out in any order, use actions with subjects and objects, or prepositional phrases that apply to specific other clauses.
Appreciate the comment, but this has been commented many times before. Yes, it does make a difference in that specific case. However, those cases aren't as common as you think. In fact, I'm seeing most people barely use any natural language in their prompts. This was the case when I recorded this (months ago) as well as today. You're not wrong in pointing this out. But punctuation still is largely inconsequential and MJ has even confirmed this.
ปีที่แล้ว +2
This is easily my favorite prompt-channel these days. Keep up the good work 👌🏻💪🏻
Lets not forget that the AI is constantly learning. I dont know about the I/F backbone of Midjourney in particular, but I guess using correct punctuation (while now be mostly irrelevant), might help in the future to receive more accurate/expected results. I mean, if you teach your children language and spelling they may very well understand your sentences without commas, but also may get problems with communcation in their future lifes. Hence, I personally believe it makes sense to globally agree on a certain syntax when communicating with AI.
Yes, proper syntax absolutely makes sense. The main reason why I emphasize that commas don't do much is because there are sooooooo many people out there who believe that commas actually break down your prompt into segments. But that's simply not true. The only way to split a prompt into multiple segments for a multiprompt, is by using colons (::) and potentially adding a weight to them, if necessary.
Stop spreading misinformation. "AI", of which these are not, are not "always learning". Ask David H himself, it's just a video game neural engine trained on billions of photos. It doesn't "learn" because it can't
It's true that adding "seed" would probably constrain the results somewhat but that doesn't negate the experiment. Still a good experiment; to be honest any experiment in this field and at this time is good. This thing is so new, and the variables so vast, that you could film a whole slew of videos on this and not get to the bottom of it.
Looks like they update something today, adding more --ar variations, like 16:9, but old seeds are broken - same string with same seed now renders absolutely different, and default style still --style 4b (i was thinking they have added --style 4c and now it's default, but no, you can't back to old algorithms).
@@TokenizedAI They noted, that only non-square images affected, but i don't like squares.. x) Also looks like "sunny" environments is more sunny (and contrast) now and smoke effects have more transparency and looks better.
I'm afraid I'm barely making progress 😔 Just so much to do between releasing new free videos and other stuff inbetween. I'm in the process of hiring a video editor, so I hope I'll make more progress soon as he takes some of the work off my plate.
My quick Midjourney v4 prompt guide: ➊ Use `::` to split the prompt into paragraphs. Their order doesn’t matter but their weights do: `office worker::` = `office worker::1`. Note: `--no glasses` = `glasses::-0.5`. ➋ Within a paragraph, simple punctuation splits text into phrases: `blue office worker` ≠ `blue office,worker` ≠ `blue,office worker`. Repeated or different punctuation like `blue,,,office worker` or `blue.office worker` with the same seed produce small variation (because they technically constitute a slightly different prompt). Repeated spaces are the same as a single space. ➌ Within a paragraph, word order influences grammar & leftmost words or phrases matter more: `worker,blue office` ≠ `blue office,worker`.
So you effectively disagree with almost everything in the video 😉 Well, except weights. I tested your examples, with and without a seed, and I don't see a difference that matches the significance you are attributimg to it. Again, I agree about weights, so I don't know why you are posting those here. You're repeating things that I've explained in the exact same way in multiple videos. It has also been confirmed by the official Guides (people in #prompt-chat) that MJ doesn't recognize commas and that they onla add noise. Feel free to ask them yourself. 🙂
@@TokenizedAI That noise has influence. They never said it doesn't. They call it "intriguing noise". Not as much influence as the paragraph splitting, but still. It definitely helps with the language processing. "blue office, worker" gives a prominent office because it's one of the subjects, and "blue office worker" gives a worker, because office is an attributive here (like an adjective = a worker who works at the office).
Ps. And I'm not here to criticize you. I appreciate your research, and your videos! I'm just trying to join the truth-finding, so that everyone is educated better. Your research & experiment, that's great. There is no shame in finding that some conclusions turn out to be inaccurate. Still, you're right to emphasize :: - this is not used by people much, and people often ascribe too much importance to order & punctuation. But I disagree that those factors "don't matter" :) Kerp up your great work!
@@AdamTwardoch Sorry, didn't want to come across as defensive. I was just genuinely confused because you listed things I've said myself. I agree that noise has influence. Of course it does. Every seed value is noise, so it's obvious that it influences things. The difference is that noise is just an uncontrolled input factor. Many people use commas not because it's just easier to read but because they believe that it separates their prompt into segments. This ascribes a level of "control" over your prompt and image which "noise" doesn't actually deliver. So yes, while commas and noise and noise can leads to dramatically different images, it doesn't necessarily mean that provide the level of control that people think it does. Especially since it's incredibly dependent on context. As you rightly say, multiprompts/weights are generally a better tool for control and fine-tuning.
I gotta say I'm interpreting the results differently. To me it looks like double quotes made a *huge* difference. Just because some of the women look the same, that's not the only thing that matters. In fact, it may be the least important thing-- as you said in the beginning-- depends what you are trying to achieve. From all the different punctuations, double quotes and no-spaces made the most changes and it would be amiss to pass them as "similar".
Yes, there's something different about the double quotes. It's definitely adding more noise into the equation. But here's my takeaway. If I can't tell what it exactly it does, then it not reliable and effectively useless to me. I don't want to be using conventions when I don't even know what exactly they do.
Also, about the quotations, for me it only seems to work differently when I put every single different idea in the prompt under quotations instead of just one part of the prompt, so that's kinda strange there's also the + sign people use in between the prompts but I think that also works just the same as anything else
The official position from MJ is: The AI can't read punctuation with the exception of :: and --. Any difference that anyone sees is just differences in noise.
This was an excellent prompt comparison video and addressed what I have been wondering about >> will watch your suggested vidoes >> thank you and Subscribed 👌💯👀🎥🖼😎🌟
I love how you go in deep on trying to understand how to better use the prompt to achieve better results but I don't think this video proves a lot because first you only show us portraits, i use midjourney to show designs or objects like gold frame, my point is maybe what you show to use don't apply to others prompts beside humans, Also I dint like how you compare images cant you put side by side ? let me thank you for trying to understand to learn how to use prompts like a pro noone else on youtube does that insane work like you and i love it.
Yes, that's a fair claim. But even if you take a scene or a landscape, the differences aren't huge. The only reason why you would perceive the small changes to be much bigger is because a scene can obviously contain more elements than just a portrait. But that's exactly why the portrait is a good example. The point of the video is NOT to say that there are no difference at all. Of course there are differences, because if you leave out commas, then your prompt is actually shorter than before. The point is to demonstrate that all of these claims about what certain types of punctuation do is mostly inconsequential. People worry about them way too much, when they actually do very little. And they definitely don't give you more control. That's just their imagination. Regarding the comparison of the images. You are right. I've gotten this feedback a few times now. I'll look for an efficient way to improve this. Thanks for the long comment by the way 🙂
The tip about the seed and remix function is super helpful. More in-depth on impact of syntax on prompts would be great as I was not entirely convinced of Christians interpretation and here's why... The AI has no need to prioritise one word over the other, with or without commas etc, due to the separate category each descriptor falls into. Parts that create the whole in a sense. For example, the words 'beautiful' and 'pink hair' are descriptors of appearance. The word 'woman' is a descriptor for the gender of character. The words 'barbarian' and 'warrior' denote the type of character. The word 'bavaria' tells AI the place and all that implies. So you can clearly see the AI has been given the following distinct ingredients for an image: 1. that there is a human character, 2. that the human is female, 3. that the female human is a barbarian warrior, 4. that the character is from bavaria. Further to this point, due to there being a much higher percentage of descriptors relating the female character it stands to reason why the character was not placed in a landscape image in bavaria but instead in a close-up portrait.
Solid analysis. In hindsight, I think I made the mistake of not putting enough thought into choosing a prompt that contained actual phrases. That would have prevented me from including the example where I switch around the positions of all the words because the order totally matters in such cases. With an assortment of keywords or concepts it usually doesn't. Regarding what you said about the "bavaria" part. When you craft the prompt with multiprompts and add sufficient weight to "bavaria culture", you actually do start to get Alpine scenery in the background and the woman looks a lot more like someone from the Alps. But in a regular single prompt it's simply not strong enough of a concept to actually show up in the image. The others are far more dominant.
@@kaizen_5091 i also notice that mj4 defaults to a 'iconic' image. Me as a concept artist and some1 whose looking to make mini cinematic movies, this is frustrating. Im bored now of seeing these images that have an emphasis on a main theme (human, landscapse, object). I hope mj5 addresses these and we can break out into a much complex canvas of mini events each with another lil story. Example : the famous indiana jones scene. Indy being chased over green fields by an amazon tribe , waiting is a pilot of a seaplane, he is fishing whist standing on the planes floats. I wait for that day with great anticipation :)
@@armondtanz I'm an Indiana Jones fan so I feels you on this one :) I'm excited to see what MJ5 brings also. I imagine Midjourneys deeper learning with continue to expand each and our ability to communicate with the AI, giving directives via prompts to achieve our desired effect will also become more realistic.
oh and about the same image over time for --seed is that this only is true, because david didn't update the system with improved ranking. Once he does that, the images will be a bit different. That happend more often in v3, but since v4 is as good, david didn't touch the running system as much. Remaking old images with seed is a great way to test if the system got an update ;)
@@TokenizedAI it always depends on the strenght of the update. But yes. After all an update is supposed to make anything better. That means changing stuff.
Thank you 🙂 Though it seems some people think I chose some bad examples. The key takeaway should be that punctuation is irrelevant, while grammar is only relevant if you're using full sentences or phrases in your prompts.
Your videos are fantastic. I would love to see one on combining images and prompts, including a deep enough, dive into weighting of images and prompts.
@@TokenizedAI I have! I like it. Still trying to figure out how to tweak weighting, I have been experimenting with combining many multiple images. Would love your take. I really appreciate your methodical approach to reverse engineering prompt magic.
Hello *Chris, would you allow me to make a response* video? FYI, *I will disagree from how you did your whole process and interpretations* , but please, don't take me the wrong way, we are all here learning and exploring together. :) Do I have your permission?
Sure, go ahead. I'll happily stand corrected. 🙂 But can you give me an example that I can actually test myself? Because based on what you're saying, I still don't see the problem. I'm focussing on what most people are actually doing, not what might the correct way of doing things from a "scientific" perspective.
I have a theory that once you have tried variations on a prompt MJ will settle on ideas that "it likes". And then, without using the seed command, when changing the the prompt significantly trying to nudge it in the direction I'm looking for... it will still give me similar results. It is like it remembers the recent attempts and has trouble letting go of the direction it headed in on the initial tries.
If I am correct, that might be skewing the results of your tests. I've tried using wildly different prompt words to get it to change significantly and it will still give similar results.
Last thought, I think the use of the same seed may be compounding the issue I'm talking about... because you're using the same seed, it may be remembering the idea it settled on and you would have to wildly change the prompt to get it to break from the initial idea.
I guess my counter-question to this would be: How else would you attempt to "measure" or "compare" the impact of a slight change in your prompt, if not by using a seed? As soon as you don't fix the seed, you will invariably get very different images than before, rending it quasi-impossible to reasonably ascertain whether you got different images due to the change in the prompt or due to a new seed.
@@TokenizedAI Don't get me wrong. I'm not downing on your methods. Last night I was working on a gorilla playing bass guitar. My initial prompts had the term cell shading in it. I was not liking the results. So, I got rid of that term. Mind you I was not using any seed command... yet it kept returning very similar results. I've noticed this behavior many times now. Sometimes I will be exploring looking for a particular look and once you've run the initial prompt several times with slight changes it's like it settles on a certain aspects that it is unwilling to let go of. In this case, it kept doing the cell shading idea even after I removed it from the prompt. That is one of my critiques of MJ. It is sometimes difficult to get it to spit out what you are looking for. While other times, the first or second prompt and it looks like a million bucks.
I didn't think you were downing my methods. I was genuinely curious if you had an idea how to better test this stuff 😁 Do you ever hit any of the feedback buttons after an upscale? I've heard that might flow into the model/algorithm as "signal" and may reinforce preferences. But I'm uncertain how impactful it really is.
You are incredible thank you so much for your help and in depth guide it helps me so much! On the 9 of February I am going to give a basic course on MidJourney where I live and your deep knowledge helps me as a beginning teacher in this beautiful world of Colab with AI to bring our imagination to life!
This guy makes pressing that Bell icon on YT sooooo damn easy! Im subbing and under 10k subs and taking a screenshot of this to make an NFT later for sale as this channel is going places! thank you. Great series on Midjourney man, beleive me, ive consumed a SHIT tonne on YT and your vids are gold.
You might be interested Christian in the series of vids I’ve just released on how to use A.I. generated images to improve your mental wellbeing? Not getting any views but still might be worthwhile to look at an episode? Just adds a new dimension to A.I. use?
I don't think it works how you describe it, and it wouldn't make sense that it does. What midjourney tries to do is extract populair meanings from the text, it uses bits of those sentences to find popular meanings and order does matter. For example, use: "praying, a mantis --seed 1" and "a praying mantis --seed 1". The first assignment will try to create a praying mantis in a praying posture, the second will just make a praying mantis. Midjourney doesn't understand a sentence in a classical sense, it extracts by association and tries to match that to what it has in its database. That is why an often used combination of words can dominate others and it is more creative with uncommon combinations.
I don't disagree. In fact, I fully agree. Midjourney doesn't understand grammar nor punctuation. So most of it is irrelevant. That wss the main point of the whole video. 🙂 May I ask what lead you to think otherwise? Genuinely curious 🤔
@@TokenizedAI What @merdanethubar-sarum9031 is saying is exactly the main problem of this video, and why I am asking to make a response. MJ is a Natural Language Processing (NLP), you created a sentence that makes absolutely no sense, with 5 contradictory Aesthetics references, a character reference, and a misplaced color. You can change this mess around and MJ will always give a similar result, and got even worst when you locked it all in a same seed... My friend, I love some of your videos, but on this one you are really missing the target and making people who still learning gets very confused... :(
@@AIArtistTV I think I understand the point you are trying to make. I chose these prompts to keep things simple. And tons of people write their prompts like that and they are happy doing it that way. In many cases, crafting full sentences doesn't make it easier for them, but actually the opposite. Are you saying that punctuation makes a difference, as long as you're prompting with full sentences? My own experience suggests the contrary
@@TokenizedAI by Christian Heidorn The main problem is that you are treating MJ as a random words processor, and MJ trying to find meaning. As the starter of this thread precisely refers to. What is the meaning of this promo you are using: "beautiful woman, pink Cleopatra hair, warrior princess, bavarian culture, barbarian warlord" This is not English, my friend. MJ is trying to guess what kind of woman you are trying to create. Got it? Yes, many are making this mistake of treating MJ as a random words processor, it gives results, but is very sad you make a video name as "TRUTH about Prompts in Midjourney" and perpetuated the most wrong way of doing it..:(((( Going further, thinking here, a prompt should be: /imagine prompt:character portrait, cleopatra as a bavarian princess with pink hair (THIS is an English prompt) The other one would be (because these are totally different aesthetics): /imagine prompt:character portrait, cleopatra as a warlord warrior with pink hair Get the best of both results and BLEND, maybe you have you had in mind. "Let's talk about the secret language of MidJourney...?" Is a prompt organized, clean and with proper english meaning on the subjet. :) I hope it helps. PS: And locking your whole process using the starting noise (seed), and expecting to see relevant changes, you locked and sabotage even more your whole experiment. :(
@@TokenizedAI Your video suggested that word order is also largely irrelvant. This isn't true because there is meaning in word order. Not in the sense of "the first is the most important" but in the sense of "Midjourney gives more weight to popular word combination when they are put right next to each other". I have given an example, just try it and see the difference.
So, some people think that I shouldn't be doing that. But I completely disagree. I'm looking at the prompts like a mathematical formula. Unless you fix the seed, it's going to change every time I enter a command. And we know that seeds are a major randomness-inducing variable in MJ (they add "noise"). So unless I use the same seed in every command (ceteris paribus), my experiment is absolutely meaningless. If I have a random seed, then my images are always going to look very different, no matter whether I change the prompt or not. Which defeats the purpose of trying to determine whether commas actually do anything or not.
Made a quick test during watching lioness --seed 666 lion ess --seed 666 l i o n e s s --seed 666 there spaces make a difference, however even in the last one a lion apeared (and it tried to write liones) I think future testing is needed. There are words that become new words with seperated spaces at the right place. They should give interesting results? Edit: made appletree --seed 666 apple tree --seed 666 they look nearly identical!!!!
Every tiny detail.....even a space, makes a difference. But only because it's interpreted as "interesting noise". However, that doesn't mean it actually gives you any extra control over anything. And that's what this is about.
Well thanks sir! Now we master each concept of prompt, seeds and text weights, I guess we eventually need to figure out an effective way to create multiple discord accounts: this 25 limit is annoying :p
Apparently you can win/earn free subscriptions if you're among the top raters for a specific day in the offiicial Discord. Though I'm not 100% sure about how that works.
I do feel that using :: even without a weight number to that part of the multi-prompt does make a difference in how much the AI focusses on it I get better results by multi-prompting the description of the subject, the background and then the style. This might also be fiction but as long as it keeps working, I don't really mind 😅
Surprised at how little difference word order made, given the importance of the attention layer in the underlying model. I'm presuming it is because word order tends to have very little importance in such keyword type lists, which in the initial training data would presumably be simply alphabetised. I presume if you are writing a verbose prompt it begins to matter more - eg. "Cat resting in a basket atop a tower at sunrise" versus "At sunrise, atop a tower, cat resting in basket"
Thank you so much for putting these videos out, I really appreciate your systematic approach that’s very much in tune with my own approach and for providing such useful prompting insights for the mj newbie that I am. 😂
Thanks for the kind feedback. The videos are a good reflection of how I experiment and learn via trial and error. It does require one to dig rather deep into the topic, which is somewhat rare in most TH-cam videos 😅
Wouldn't you get everytime the nearly same image by defining the first one be the seed image? So for my understanding you told MJ with your first Image what images it should generate no matter what you are typing, isn't?
That's not what a seed does though. Seeds are input variables that just define an initial "noise" input. "a blue house --seed 1" and "a green frog --seed 1" abviously do not render the same image, despite using the same seed. I'm not sure why, but a lot of people seem to misunderstand the purpose of a seed.
From what I noticed, mainly in image 4, when you had // you had some criss cross pattern in the back, with ( you had arches in the back and with " you had as you said more hair and pointy things. what do you think?
Yeah, but it's still meaningless. As per the MJ team, all it does is add "interesting noise". So you might get interesting results, but you'll never know why or what it means. And that ultimately makes all these symbols fairly useless. Unless you know exactly what something does, it's not really giving you more control over anything.
No. That's not what a seed does. A seed is just an input variable that is meant to inject random "noise" into the algorithm, so that you always get different results every time you enter the same command and prompt. For example, every time you hit re-roll, it changes the seed as well. Seeds do don't apply any "style" or anything like that. They are just one of many variables. The reason why you would fix a seed is because you want to "stabilize" the randomness of your outputs. In this case, I did it so that I can see whether the other changes actually change anything at all. In any normal situation you would do it because you liked a particular set of images for their composition and you want to fine-tune your prompt towards that same direction.
Hey nice vids, just 2 quick questions please. 1) This IA has some troubles with text right ? Do you have any tips for that ? Are do we have to use another software ? (If another one, do you have sugestion please ?) 2) Can we use :: and , at the same time ? To use weight on few words and not on others ? Thanks in advance !
1) Don't use it for text. Midjourney currently isn't capable of producing sensible text. Doesn't matter how hard you try. Anything that looks correct is pure coincidence. 2) Sure, since the comma doesn't actually do anything, you can keep it in there. I use commas all the time because it's easier to read.
There is another secret about text to image generators. The AI uses images from artists from all over the internet without their consent. Which makes it probably the biggest copyright infringement of the last two decades, or even the last century. All these generated images are made on the backs of artists, photographers and all other image content creators. I hope as many people as possible realize this before posting their generated images online, claiming "I made this".
I would the problem in your methodology is that it is you doing prompts over and over under your one login which the ai recognizes as all related. The test would be if your prompt put under your login and then also a non-related login give the same results, or near same.
NGL the more I dig into AI and programs/features available now.... I get kinda worried. As most of the population probably does. I feel so ... foreign for one. I just started basic coding, and thought it was crazy in itself... but now that I can type into CHATgpt all sorts of questions about coding, I am able to relate it to knowledge I already have. It's like having a person that has been coding 30+ years, right next to me, answering every question I have in full detail...that's just bonkers!! What do you think? Elon mentioned the risk of this already, but I didn't grasp what he was talking about until I discovered A.I. uses.
Careful with ChatGPT. Especially in areas where you don't necessarily have expertise. Very often it sounds super comvincing because it answers with such confidence. I've seen examples of it being horribly wrong but sounding very convincing. For example, it just made up Laravel functions on the fly that don't even exist. But because it used the typical Laravel conventions, I actually had to look it up first and confirm that it was bullshitting me. ChatGPT is currently more style than substance. But since most of the media types don't look beyond the surface, it's being hyped up way too much. Yes, it's extremely useful, but it's not an outright job killer in most cases. The clever people are just going to be more productive than they already are. I have friends who do a lot of writing in thier jobs. They seem worried, mainly because they don't understand the tech and I'd argue they aren't the most tech literate people anyway. If they bothered to actually care and get into the topic, they'd easily be amongt the first out of the gate. And yet, it grabs their attention for 2-3 weeks and then they return to their regular old ways without any consequence. Not gonna lie....they are the problem, not the AI.
🟠 Hi Christian this is a great video, thank you so much for going through this and clarifying. I have been wondering how to figure this all out, in which you have effectively clarified. One question still thou, i was not able to understand why princess::20 and culture::20 equal 93%? what does the numerical value of ::20 represent? It seems it is not percentage, otherwise then it would effect only 40% of the image instead of 93%? This was a bit confusing for me since im still learning // figuring things out. thanks! :)
If you have 5 segments in your prompt and three of them have a weight of 1 each and two have a weight of 20 each, then you assigned 43 weight points overall. So the two weights that have 20 each (40 total) represent 93% of the overall weighting (40/43=0.93) in the prompt. Each one has 46.5% importance. The remaining 7% are shared by the three other segments with a weight of 1 each. So in the greater scheme of things, they are almost irrelevant. Weights are relative, not absolute. So for example, these 2 prompts are identical im effect: red house::20 green garden::20 -and- red house::1 green garden::1
@@TokenizedAI Thank you for your great explanation. I have a feeling by your last name you may also be German like me, and is why i really enjoy your clear and well thought out style of explanation. Instantly subscribed after one video, fan for life! 😄😄 thanks again 😊
@@TokenizedAI I think this sentence "Weights are relative, not absolute" and the example is so important that it should be pinned at the top of your comments, just saying ;)
Yup, I'm German 😁 Thanks for the sub! I didn't expect the channel to grow this quickly. The first 5k were rather smooth sailing. But a fair bit of negativity is creeping in now. Need to learn to deal with it I guess.
I would, but I can only pin 1 comment and I need the current one to stay up. I've been getting a lot of flak for this video because some "uber experts" feel that I'm misrepresenting things. Unfortunately a lot of people are getting bogged down on details that are largely inconsequential. In all fairness, I may have tested with some unsuitable examples. But clearly, my main point is coming across enough: Commas don't matter.
Cool, thanks. The word order doesn't really seem to have any impact. I often use prompts where the topic is in the middle or at the end and it still works fine.
Word order DOES have an impact when you are constructing full sentences (that actually make sense). It DOES NOT matter much, if you're just using a string of random or loosely relatec keywords or partial phrases.
I was confused as well, but I guess their language model is able to identify stuff like this. Not sure how or why. I didn't go so far to actually mix up the letters though 🤣
Would be cool to compare results with the same prompt on different models on stable diffusion, e.g. the openjourney model claims to make you achieve similar results to MJ, thee is also a text inversion for stable diffusion (midjourney based). We all know MJ is better than stable diffusion. I discovered SD recently and playing with it a lot, and I'm quite satisfied with the results, it's like mining crypto, just I'm minng nice images on my GPU instead of paying MJ ;)
Midjourney v4 runs on top of Stable Diffusion. The main difference is in MJ's execellent use of embeddings as well as the natural language processing layer that it has placed on top of it. But at its core, any AI generator that's not built by OpenAI (e.g. DALL-E) is 99% likely to be using Stable Diffusion under the hood.
Hi, it's a nice tutorial. I get a lot of progress after watching your videos. I have a question: How to make the main object placed in a certain position on the canvas, for example at the bottom, or the top right? I'have tried several times but no luck, the object is always positioned at the center.
hi, I've seen the work of a user with id Julz#8275. I don't know how he is prompting to produce images with very high detail, and can place objects in certain positions.
@@TokenizedAII don't know how to get the prompt in MJ App. I've been looking for it but only the picture shows up, no prompt. I'm a newcomer to Discord and only been using MJ App for 3 days so I'm not very familiar with it. I am amazed by the work of other experienced users.
@@TokenizedAI yeah, on #newbies's channels there is always a prompt above every picture that is made. But the picture I mean is on channel #v5-showcase and #show-and-tell. There are no prompts at all, just pictures.
Perfect example, how you make a video with thesis in mind, and even if evidence is against you, you still arrive at the thesis. Even if img change, you disregard it.
I didn't say the images didn't change. I said the changes aren't as significant as some people claim. I appreciate the feedback though. Looking forward to a video with your take on the matter 🤗🤝
@@TokenizedAI I sounded rough, sorry for that. I really like all your other videos, good job on that! But this one... it just didn't resonate with me, as I disagree with the same result you showed. Don't let it discourage you :) Have a great weekend!
No worries. Appreciate what you're saying. Unfortunately, it seems like the main point that I wanted to make got lost since everyone seems to be focusing on 1-2 specific examples where (I admit, in hindsight), I may not have chosen the best examples. The point I was trying to get across is the fact that punctuation is largely inconsequential. So many people swear by using quotation marks, commas or even brackets, when they actually don't do anything (other than add noise). This has been confirmed by the clever people in #prompt-chat as well. I should have made clear that the order of the words obviously matters if you are using proper sentences in your prompt. I guess I just assumed no one in their right mind would put a real sentence through a blender 🤣
But, does each prompt have a memory from the preceding prompt? Because all of these tests are done in sequence immediately one after another does the AI remember the preceding prompt and, if so, wouldn't that effect the output?
i wonder if it is possible to build a dataset for AI. If any artist do not want their work to be used to train the machine, he can put their work into the dataset. Once Ai picks an image from internet, it checks if the image exist in the dataset. If an image is over 99% similar one of the work in the set, it will pick another one.
Problem is, I'm not sure that's possible. The AI model itself doesn't contain the actual image itself. It was merely trained with it. You would have to cross reference with the original dataset, rather than with the model itself.
I can never get buildings like skyscrapers, or anything with straight lines or get the perspective I want. Am I doing something wrong is this not possible yet ? lol All my buildings look like sketches from a drunk person. Are there also negative prompts in MJ? Edit: I played around with colons and its already helped with straight lines!
Hm... I wonder if in a 4-image prompt the relative position mean something, for example, your repeated women 1, 2, 3 and 4 always appeared in the same positions. In other words -- 1 would be tame, 2 epic, 3 complex, 4 went to hairdresser. Does character 4 always appear in position 4? Or is it just my imagination?
Does using the make variations button preserve the same seed or is that same as re-roll? Also, I think in one of your videos you used [brackets around the various prompt segments for image weights]::3 This was a great tip for helping me separate image segments in my brain.
Actually, I don't think I've ever used brackets. You must be confusing me with someone else. Variations will change the seed as well. However, it's important to remember that variations don't process prompts the same way as a regular /imagine command, since it sticks extremely close to the original.
@@spartiata71 How do you mean? Sorry for misunderstanding you. But I would be very curious what you mean by merging sentences together. If you could give an example with a prompt, I would appreciate it so much!
@@TokenizedAI So in order to preserve the same seed with variations you would have to turn remix mode and manually specify the seed, yes? Also, you are right. When I was talking about the brackets, I must have gotten my TH-camrs mixed up. I made it up to you by clicking that subscribe button! ;-)
I tried to get 2 characters like Batman and Joker or Superman. I always get same characters pics somehow changed like Joker but looks similar to Batman. Any suggestions how to improve?
@@TokenizedAI Yes true. Actually I entered - Batman vs Joker. Tried Batman::1 Joker::1 but the result was having weight of Batman or vice versa whichever is mentioned first
@@theboldbroadcaster Yeah, so the problem here is that when you use weights like that, it will always BLEND the 2 concepts together. So it's similar to what happens when you just add 2 image prompts. It always blends. It's possible to achieve what you want, but it's too complicated to really explain in detail here. One of my next videos is about this. For now, give this a try. It's not perfect, but a start. And yes, copy the entire prompt 😉 /imagine batman standing next to the joker in the streets of new york city, in the style of marvel comics:: the joker positioned to the right of batman on a new york city street, in the style of marvel comics:: new york street scene showing batman next to the joker, in the style of marvel comics::
Word order definitely matters when you have long strings of keywords in a very large prompt. It uses the first 5 or 6 keywords, and after that it randomly picks and chooses words to use.
It matters and it also doesn't matter. The lenght of the prompt isn't the driving factor. It matters when you use full sentences or phrases. It doesn't matter when you only use singular or very short keywords.
Wouldn’t they all look the same if you are referring to seed 1 each time-as it remembers that original seed? Or didn’t you remind it of the original seed between prompts?
If the only thing determining what an image looks was it's seed, then why on earth are we wrtiting prompts in the first place? Seeds are a critical influence, but it's not like they are the single most important thing.
@@TokenizedAI right, but if it took the seed as the base pic, would you even know if the commas, dashes, etc were making a difference? I guess I have always used a more narrative approach and v5 responds even better than 4 to my style. But very interesting…I have been curious about this for awhile. Thanks for taking the time to run the tests.
@@TokenizedAI thank you! And the repeating in prompts? Detailed hyper detailed extreme detail.. are those just rookie mistakes or do they have an effect?
"extremely detailed" might do something but largely depends on the rest of the prompt. Anyway you see complete gibberish prompts with lots of copied keywords, it's usually a newbie. Repeating prompts are a bit different though. Sometimes they're on purpose when someone's using multiprompts. Depends on what the user is trying to do.
I’m guessing you’re using V 4 here. My own experience with many thousands of V 3 images is that text order does matter, though I haven’t tested this with V 4. But I’m talking about style words, not nouns
Funny how everyone seems to be focusing on that one bit. I think I should clarify that text order absolutely does matter if you are using proper phrases and grammar. Which makes sense though. And I can't think of any reason why anyone would want to mess around with the order. So I don't really understand why some seem to be up in arms about the example I gave. Thing is, most people aren't reall using tons of phrases and if you only use basic keywords and concepts like I did in the video, then it doesn't really matter. Personally, my prompts are a lot more complex. But those aren't helpful for simplifying and explaining things. Especially not for Beginners. But it's funny how people seem to be missing the actual point of the video. My main message was supposed to be that commas and punctuation are completely irrelevant. 🤣
I think your confusing "+" with multiprompts. When you see multiple segments with a + between them on the Midjourney web app, then it means that the person actually used "::". That's how multiprompts are displayed there.
![F.HERO Ft. JSPKK x ลำไย ไหทองคำ x M-PEE - ไม่สนิทบิดหมด (Thai Riders Anthem) [Official MV]](http://i.ytimg.com/vi/kPJZ0UihBfk/mqdefault.jpg)
PUBLIC SERVICE ANNOUNCEMENT (just for context)
Some (very few) people seem to be upset by the way I've set up this experiment and they've identified errors in my interpretations. And that's fine! Please correct me whenever I am wrong. I'm just learning myself 🙂It doesn't matter how I set things up. There's always going to be an expert that will find something that I've done wrong. I'm sorry, but I just don't worry about the small stuff in life.
In many of these cases, they are honing in on very specific statements in the video as if they entire video was only about those specifics. And that's simply not the case. They are absolutely right in their assessment of those particular examples, however, they are choosing to ignore how the majority among you are actually writing their prompts. Yes, it might be mostly wrong. But so what! Yes, if you have advanced knowledge about those these AI models work, then you would probably not write your prompts with just a few non-sensical keywords. Or maybe you would. Who knows? It's all about trial and error, isn't it? I've gotten awesome images with just a few words. I don't see the problem in that. 🤷♂
The main message of this video is that the use of commas (!!) is not as relevant or powerful as many believe. This statement is absolutely true, no matter how you write your prompts. To claim that I am spreading fallacies isn't a fair representation of what this video does. Watch it and make up your own mind. Do your own tests. Verify or falsify it. Go crazy! 😂
My objective is to provide helpful insights to the 80-90% of you out there, who are currently struggling to create the images you want or need. It is NOT my objective to tell you how you MUST write your prompts or that you MUST absolutely conform to some sort of a convention that a small group of uber-experts believe to be the only true way. There's a reason that even at the very top echelons of science, people hold radically different views and routinely get upset about positions held by their peers. I don't really have the energy for that, to be honest. 😜
Let's not nitpick about inconsequential details. I'm a pragmatist. I'm a fan of the pareto principle. 20% of the effort is going to get you 80% of th way. The last 20% is going to cost you 80% of effort. I don't really see the point in that.
You are here to get results, not to debate the eternal granularity of a natural language processing AI. ✌🤗❤
I certainly don't want to go all poo-poo on your methodologies for sure, but I also think taking them as gospel is wrong as well. There is so much that goes on BEHIND the scene with MJ that we just don't know about. It isn't like Stable Diffusion where the model gets exactly what you typed, and then you get the exact output. With MJ what you enter is further processed by MJ before even being sent to the model, and then the output of the model is further processed before it is presented to you.
Furthermore we have all see how MJ tends to "lock in" features in subsequent generations as if there is bleed through from one to the other, regardless of the prompt.
Now all that said, I thought I would be able to prove how this methodology fell short by doing essentially the same tests in the OPPOSITE order, IE starting with a prompt with 0 spaces, just one huge word, and expanding out from there. But I'll admit when I am wrong and by golly I was wrong because right from the very first smashed together prompt it actually generated roughly the same thing. I don't have a clue HOW but it did.
That last bit made me chuckle 😂
Listen, I have no intentions of portraying anything as gospel. Quite the contrary. I'm showing in public what I've done and how I perceive the results.
Everyone should prompt whichever way they want and what gives them results. Let's s face it. No one knows what's really going on aynway. 🙂
Hi Christian, I found your content fascinating. Love that it opens up a focused discussion on the language learning curve of Midjourney AI or rather any creative AI. I commented separately on my take of the breakdown of syntax interaction in the video and would be interested in what you think. I think it would be great to see an even more in depth discovery sesh on this topic as I feel it is the key to unlocking potentially more control over the results produced.
@@kaizen_5091 Thanks for the feedback. Let me check out that other comment real quick.
Yeah the problem is more in how your title and thumbnail of your video conveyed a certain authority to the matter. It's nice that you tested some stuff out, but you still don't know what affects what and so there is no actual strength in this video.
There are a lot of experts that can teach you through some videos so maybe do a little more research, then come back and kill this argument; or be butt hurt that people pointed out your general lack of knowledge on this subject. Either way in order to not actually sweat the small stuff, you shouldn't worry about dignifying yourself at all.
I think it would be even more amazing if we could communicate with the AI using layers. For instance, we could assign the first layer as the background, the second layer as the person, the third layer as the object they're holding, and the fourth layer as an additional object in motion. It would be incredible if we could then text our layered image ideas into a Photoshop project, manipulate the newly generated AI layers, and have full control over the final product. This could revolutionize the way artists create original art and provide a new level of creativity. I believe Adobe should seriously consider implementing this feature into Photoshop. Imagine creating a new layer, selecting the size, and having the AI ask us what we want to create on that layer. Then, we could manipulate each object in the layer, resizing and moving them as we please. This is a unique idea, and I can't wait to see it come to life!
You sort of can. Check my community post on Layered Multiprompts.
One thing I've noticed about MJ is that if you keep using the same seed, the design is somewhat locked. So that it may ignore the prompt changes in favor of keeping the design mostly the same.
So in this case would you use --seed over --sameseed so that you get the variations? Or have you tried that yet? I feel like using just seed would get you what you want. Use the same number though. Hope this helps!
Problem is that --samseed doesn't work with v4.
This was one of his main mistakes on the whole process that he was trying to do.. :(
Was wondering the same, can be the topic for another video
Nice catch! I wonder if a seed itself could influence variability. Statistically there would be noise generated with clusters of pixels at certain coordinates. At that point resolution also would start playing a factor
As a retired English teacher and editor, I'm having a blast experimenting with prompts in AI. I made a bumper sticker a long time ago: Words Matter (not a political statement but a professional statement). This is looking to be quite applicable to AI prompts. Like weighting, putting in 200 descriptive words has both diminishing returns and increased, unintended consequences. Being intentional about the words you include and looking at the relationship of words to each other (not order) within the prompt does seem to matter. As mentioned in the video, weighting words or phrases seems to take away (possibly unweight) other descriptors. The words Barbarian and Warrior, for example are likely to produce results where some images have armor, and others do not. Weighting (increase) Barbarian and Warrior may increase armor, but it still may not, and it may come at the cost of diminishing other traits (unweighting). I'm finding better results using words in such cases. Woman, warrior, leather armor with fine detail, etc., will increase armor in the results, add more detail to the armor, but won't water down the other descriptors in the prompt. Notice in my example that "leather armor with fine detail" is a phrase set off by comas. This seems to give the aI more to work with on the armor and keeps "fine detail" with the armor. Another example I discovered was on an image of a stone cottage in the woods. I had "wooden bench" in the original prompt, and most of the image results had a wooden bench up against the front of the house, usually next to the door. But when I edited the prompt and put the bench in a phrase, "wooden bench by a campfire", the bench always appeared next to the campfire and not against the cottage. This may seem obvious, but being intentional about word and phrase use in AI prompts helps you better predict the results. There is so much to learn about prompts from a language perspective, but so interesting to try and learn how the AI interprets what we write.
I totally agree. Being intentional about words is very underrated. Unfortunately, the fact that you can enter a bunch of random words and still produce a very appealing image is distracting from the fact that the art lies wtihin the ability to produce images that resemble exactly what you are aiming for.
It's a bit like at the pool table. Sure, you can always get a lucky shot and it'll look totally cool. But the true masters are the one who call a shot and also make it!
In 5-10 years, schools will adopt all this as part of their curriculum.
Having said that, I want to thank you for putting all this out here, 5-10 years early. I think I'll watch all your videos in just 1 long sitting. This one is already vid #7 for me.
Your voice is alright btw
*Marks day in calendar in 2028 and 2032*
We'll see 😅 Thanks!
tons of tutorial but none is so to the point and full of substance as yours , amazing job , thanks a ton
I have watched all your videos and it's super fantastic! I've learned a lot of things and I would definitely recommend this to anyone who tried midjourney but does not know how to get around with it and how to maximize specific parameters. V4 prompt is nuanced so nobody can claim what is the right or wrong prompting. Please ignore the people who are spreading negative vibes, all of us are on the learning process. The only things that matter are you are consistent and continuously experimenting. I do love your approach so keep producing content like this as it is very helpful for sure! Feel free to experiment and prompt in your own way and I respect your structure and approach. This really proves that the knowledge shared really enhance my Midjourney prompting a lot! Thank you very much.
GM! Appreciate the super kind feedback, ser 🙏
I can’t agree more!!! I learn the similar way. Try and learn. I definitely will try your prompt experiment and see how it turns out. We are learning through AI and AI learn from us too.
OMG, 這是我這陣子以來看過最有用的影片，感謝你提供的寶貴經驗 🤩 我立刻把你的影片分享給同事們了
我之前操作時也對於到底要如何撰寫有效的Prompts感到非常疑惑。很多字詞的順序，“ ”和, 的使用等都不像許多人影片分享的那樣有意義，我做了一些嘗試卻歸納不出可能的pattern，因此很高興看到你也有這樣的想法。你的經驗與其他影片也 inspired 我有些新的方向，希望我們都能更加了解如何更有效的讓AI明白我們腦中構思的構圖配置，而不是常常像骰骰子碰運氣一般，再次感謝💗
♥️🙏🏻🤗
Your word order test didn't change the result because it was just a list of nouns and adjectives. If you had used a verb, the order would have mattered. For example, "cat eating a sandwich" and "sandwich eating a cat" contain exactly the same words in a different order, but give very different results!
Yep, I've discussed this with many others in the comments as well. This is a flaw, or at least a missing comparison, in the tests. Real phrases do get impacted. However, many are also underestimating how many people do not use phrases in Midjourney and mainly work with keywords. I'm not saying that's how they should be doing it, but it's what many are doing.
Also, the default weight is 1 so :: is 1. And --no is the same as ::-0.5. So "large office... worker suit --no glasses" is "large office... worker suit::1 glasses::-0.5". But check yourself that other punctuation does matter. Compare "large office... worker suit" vs. "large office worker... suit". English uses nouns as subjects or attributives to subjects, so "office worker" and "office, worker" are two different things in English, and you can hint that to MJ via normal punctuation, bit it doesn't matter which signs you use. Except :: which really has special meaning, and I propose to think of it as the equivalent of a paragraph marker in human text.
For testing grammar, you should make a sentence where the meaning is different depending on a comma.
Use descriptions such that word order matters in English! Rather than just a bag of nouns and adjectives that you can figure out in any order, use actions with subjects and objects, or prepositional phrases that apply to specific other clauses.
Appreciate the comment, but this has been commented many times before. Yes, it does make a difference in that specific case. However, those cases aren't as common as you think. In fact, I'm seeing most people barely use any natural language in their prompts. This was the case when I recorded this (months ago) as well as today.
You're not wrong in pointing this out. But punctuation still is largely inconsequential and MJ has even confirmed this.
This is easily my favorite prompt-channel these days. Keep up the good work 👌🏻💪🏻
Wow, thanks! 🤗
Lets not forget that the AI is constantly learning. I dont know about the I/F backbone of Midjourney in particular, but I guess using correct punctuation (while now be mostly irrelevant), might help in the future to receive more accurate/expected results.
I mean, if you teach your children language and spelling they may very well understand your sentences without commas, but also may get problems with communcation in their future lifes. Hence, I personally believe it makes sense to globally agree on a certain syntax when communicating with AI.
Yes, proper syntax absolutely makes sense. The main reason why I emphasize that commas don't do much is because there are sooooooo many people out there who believe that commas actually break down your prompt into segments. But that's simply not true. The only way to split a prompt into multiple segments for a multiprompt, is by using colons (::) and potentially adding a weight to them, if necessary.
Stop spreading misinformation. "AI", of which these are not, are not "always learning". Ask David H himself, it's just a video game neural engine trained on billions of photos. It doesn't "learn" because it can't
There a lots of MJ videos out there but yours are very informative, organized, and relevant. Thank you!
Wow, thank you! 🙂
It's true that adding "seed" would probably constrain the results somewhat but that doesn't negate the experiment. Still a good experiment; to be honest any experiment in this field and at this time is good. This thing is so new, and the variables so vast, that you could film a whole slew of videos on this and not get to the bottom of it.
Thank you! Finally someone who gets it 🤗
I'm very happy you mentioned the text weights in this video. They are EXTREMELY important!
Dude yes I been trying to make random one’s “Couch pumping gas at Gas station” keep getting Couch in in store lolzs this is very helpful.
Hahaha, I don't even want to try that in MJ. I'm sort of worried about the result 😂
Looks like they update something today, adding more --ar variations, like 16:9, but old seeds are broken - same string with same seed now renders absolutely different, and default style still --style 4b (i was thinking they have added --style 4c and now it's default, but no, you can't back to old algorithms).
Hmm, interesting. I'll have to check this out later today. Thanks for the heads up!
@@TokenizedAI They noted, that only non-square images affected, but i don't like squares.. x)
Also looks like "sunny" environments is more sunny (and contrast) now and smoke effects have more transparency and looks better.
I love your videos mate thank you when is your course coming out ???
I'm afraid I'm barely making progress 😔 Just so much to do between releasing new free videos and other stuff inbetween. I'm in the process of hiring a video editor, so I hope I'll make more progress soon as he takes some of the work off my plate.
Beautiful. Could you please make a lesson about basic things of effects like ultrarealistic/dramatic light/8k and so on? Would love to see that
Noted. "8k" doesn't really do anything by the way.
My quick Midjourney v4 prompt guide:
➊ Use `::` to split the prompt into paragraphs. Their order doesn’t matter but their weights do: `office worker::` = `office worker::1`. Note: `--no glasses` = `glasses::-0.5`.
➋ Within a paragraph, simple punctuation splits text into phrases: `blue office worker` ≠ `blue office,worker` ≠ `blue,office worker`. Repeated or different punctuation like `blue,,,office worker` or `blue.office worker` with the same seed produce small variation (because they technically constitute a slightly different prompt). Repeated spaces are the same as a single space.
➌ Within a paragraph, word order influences grammar & leftmost words or phrases matter more: `worker,blue office` ≠ `blue office,worker`.
So you effectively disagree with almost everything in the video 😉 Well, except weights.
I tested your examples, with and without a seed, and I don't see a difference that matches the significance you are attributimg to it.
Again, I agree about weights, so I don't know why you are posting those here. You're repeating things that I've explained in the exact same way in multiple videos.
It has also been confirmed by the official Guides (people in #prompt-chat) that MJ doesn't recognize commas and that they onla add noise. Feel free to ask them yourself. 🙂
@@TokenizedAI That noise has influence. They never said it doesn't. They call it "intriguing noise". Not as much influence as the paragraph splitting, but still. It definitely helps with the language processing. "blue office, worker" gives a prominent office because it's one of the subjects, and "blue office worker" gives a worker, because office is an attributive here (like an adjective = a worker who works at the office).
Ps. And I'm not here to criticize you. I appreciate your research, and your videos! I'm just trying to join the truth-finding, so that everyone is educated better. Your research & experiment, that's great. There is no shame in finding that some conclusions turn out to be inaccurate. Still, you're right to emphasize :: - this is not used by people much, and people often ascribe too much importance to order & punctuation. But I disagree that those factors "don't matter" :) Kerp up your great work!
@@AdamTwardoch Sorry, didn't want to come across as defensive. I was just genuinely confused because you listed things I've said myself.
I agree that noise has influence. Of course it does. Every seed value is noise, so it's obvious that it influences things. The difference is that noise is just an uncontrolled input factor.
Many people use commas not because it's just easier to read but because they believe that it separates their prompt into segments. This ascribes a level of "control" over your prompt and image which "noise" doesn't actually deliver. So yes, while commas and noise and noise can leads to dramatically different images, it doesn't necessarily mean that provide the level of control that people think it does. Especially since it's incredibly dependent on context. As you rightly say, multiprompts/weights are generally a better tool for control and fine-tuning.
I gotta say I'm interpreting the results differently. To me it looks like double quotes made a *huge* difference. Just because some of the women look the same, that's not the only thing that matters. In fact, it may be the least important thing-- as you said in the beginning-- depends what you are trying to achieve. From all the different punctuations, double quotes and no-spaces made the most changes and it would be amiss to pass them as "similar".
Yes, there's something different about the double quotes. It's definitely adding more noise into the equation.
But here's my takeaway. If I can't tell what it exactly it does, then it not reliable and effectively useless to me. I don't want to be using conventions when I don't even know what exactly they do.
Also, about the quotations, for me it only seems to work differently when I put every single different idea in the prompt under quotations instead of just one part of the prompt, so that's kinda strange
there's also the + sign people use in between the prompts but I think that also works just the same as anything else
The official position from MJ is: The AI can't read punctuation with the exception of :: and --. Any difference that anyone sees is just differences in noise.
@@TokenizedAI yep, the double quotations might be generating a lot of noise somehow, i will keep testing it out
This was an excellent prompt comparison video and addressed what I have been wondering about >> will watch your suggested vidoes >> thank you and Subscribed 👌💯👀🎥🖼😎🌟
Awesome, thank you!
I love how you go in deep on trying to understand how to better use the prompt to achieve better results but I don't think this video proves a lot because first you only show us portraits, i use midjourney to show designs or objects like gold frame, my point is maybe what you show to use don't apply to others prompts beside humans, Also I dint like how you compare images cant you put side by side ? let me thank you for trying to understand to learn how to use prompts like a pro noone else on youtube does that insane work like you and i love it.
Yes, that's a fair claim.
But even if you take a scene or a landscape, the differences aren't huge. The only reason why you would perceive the small changes to be much bigger is because a scene can obviously contain more elements than just a portrait. But that's exactly why the portrait is a good example.
The point of the video is NOT to say that there are no difference at all. Of course there are differences, because if you leave out commas, then your prompt is actually shorter than before.
The point is to demonstrate that all of these claims about what certain types of punctuation do is mostly inconsequential. People worry about them way too much, when they actually do very little. And they definitely don't give you more control. That's just their imagination.
Regarding the comparison of the images. You are right. I've gotten this feedback a few times now. I'll look for an efficient way to improve this.
Thanks for the long comment by the way 🙂
Ah man stop hating, why you always hating bruh.
David from MidJ said their language model is not great, so yeah commans don't matter. He saif they don't himself
The tip about the seed and remix function is super helpful. More in-depth on impact of syntax on prompts would be great as I was not entirely convinced of Christians interpretation and here's why... The AI has no need to prioritise one word over the other, with or without commas etc, due to the separate category each descriptor falls into. Parts that create the whole in a sense. For example, the words 'beautiful' and 'pink hair' are descriptors of appearance. The word 'woman' is a descriptor for the gender of character. The words 'barbarian' and 'warrior' denote the type of character. The word 'bavaria' tells AI the place and all that implies. So you can clearly see the AI has been given the following distinct ingredients for an image: 1. that there is a human character, 2. that the human is female, 3. that the female human is a barbarian warrior, 4. that the character is from bavaria.
Further to this point, due to there being a much higher percentage of descriptors relating the female character it stands to reason why the character was not placed in a landscape image in bavaria but instead in a close-up portrait.
Solid analysis.
In hindsight, I think I made the mistake of not putting enough thought into choosing a prompt that contained actual phrases. That would have prevented me from including the example where I switch around the positions of all the words because the order totally matters in such cases. With an assortment of keywords or concepts it usually doesn't.
Regarding what you said about the "bavaria" part. When you craft the prompt with multiprompts and add sufficient weight to "bavaria culture", you actually do start to get Alpine scenery in the background and the woman looks a lot more like someone from the Alps. But in a regular single prompt it's simply not strong enough of a concept to actually show up in the image. The others are far more dominant.
@@TokenizedAI Indeed. I look forward to more of your content. Thanks for reply :)
@@kaizen_5091 i also notice that mj4 defaults to a 'iconic' image. Me as a concept artist and some1 whose looking to make mini cinematic movies, this is frustrating.
Im bored now of seeing these images that have an emphasis on a main theme (human, landscapse, object).
I hope mj5 addresses these and we can break out into a much complex canvas of mini events each with another lil story.
Example : the famous indiana jones scene.
Indy being chased over green fields by an amazon tribe , waiting is a pilot of a seaplane, he is fishing whist standing on the planes floats.
I wait for that day with great anticipation :)
@@armondtanz I'm an Indiana Jones fan so I feels you on this one :)
I'm excited to see what MJ5 brings also. I imagine Midjourneys deeper learning with continue to expand each and our ability to communicate with the AI, giving directives via prompts to achieve our desired effect will also become more realistic.
@@kaizen_5091 i hope they dont just add another 5 billion scans. That would be terrible.
More complex prompting would be amazing.
Get well soon.❤❤
Thank you 🤗
oh and about the same image over time for --seed is that this only is true, because david didn't update the system with improved ranking. Once he does that, the images will be a bit different. That happend more often in v3, but since v4 is as good, david didn't touch the running system as much.
Remaking old images with seed is a great way to test if the system got an update ;)
So would that explain why we sometimes see really tiny changes to the image when I run the same prompt like a month later?
@@TokenizedAI it always depends on the strenght of the update. But yes. After all an update is supposed to make anything better. That means changing stuff.
Absolutely nailed it with the use of prompts, thanks for posting
Glad you enjoyed it! 😀
Love your methodical and logical approach to presenting the information. THanks so much
Thank you 🙂 Though it seems some people think I chose some bad examples. The key takeaway should be that punctuation is irrelevant, while grammar is only relevant if you're using full sentences or phrases in your prompts.
Your videos are fantastic. I would love to see one on combining images and prompts, including a deep enough, dive into weighting of images and prompts.
Thanks! Have you seen my video on combining images yet? Or were you looking for something more in depth?
@@TokenizedAI I have! I like it. Still trying to figure out how to tweak weighting, I have been experimenting with combining many multiple images. Would love your take. I really appreciate your methodical approach to reverse engineering prompt magic.
@@johnfromberkeley The more you add, the more it will try to blend them. You have to add a text prompt in order to exert more control over the image.
Hello *Chris, would you allow me to make a response* video?
FYI, *I will disagree from how you did your whole process and interpretations* , but please, don't take me the wrong way, we are all here learning and exploring together. :)
Do I have your permission?
Sure, go ahead. I'll happily stand corrected. 🙂
But can you give me an example that I can actually test myself? Because based on what you're saying, I still don't see the problem. I'm focussing on what most people are actually doing, not what might the correct way of doing things from a "scientific" perspective.
@@TokenizedAI On the @merdanethubar-sarum9031 reply, around 13 comments below I elaborated it better.
Dude you are killing me... I can't stop watching your videos. KEEP UP THE GREAT WORKS!
And I haven't even watched THIS one yet!
>play
Thanks 😁
@@TokenizedAI Thanks nothin,
dang you... I got things to do... like actually use some of this advice. :)
--seed 1 maybe a factor in keeping the images looking the same.
Could be. Question is....how else do you compare? Considering that changing the seed results in a different image anyway.
Compare "blue office worker --seed 1", "blue office, worker --seed 1", "blue office,,,,,.... worker --seed 1", "worker...... blue office --seed 1", and of course finally "blue office:: worker --seed 1" and its identical equivalent "blue office::1 worker::1 --seed 1"
I have a theory that once you have tried variations on a prompt MJ will settle on ideas that "it likes". And then, without using the seed command, when changing the the prompt significantly trying to nudge it in the direction I'm looking for... it will still give me similar results. It is like it remembers the recent attempts and has trouble letting go of the direction it headed in on the initial tries.
If I am correct, that might be skewing the results of your tests. I've tried using wildly different prompt words to get it to change significantly and it will still give similar results.
Last thought, I think the use of the same seed may be compounding the issue I'm talking about... because you're using the same seed, it may be remembering the idea it settled on and you would have to wildly change the prompt to get it to break from the initial idea.
I guess my counter-question to this would be: How else would you attempt to "measure" or "compare" the impact of a slight change in your prompt, if not by using a seed? As soon as you don't fix the seed, you will invariably get very different images than before, rending it quasi-impossible to reasonably ascertain whether you got different images due to the change in the prompt or due to a new seed.
@@TokenizedAI Don't get me wrong. I'm not downing on your methods. Last night I was working on a gorilla playing bass guitar. My initial prompts had the term cell shading in it. I was not liking the results. So, I got rid of that term. Mind you I was not using any seed command... yet it kept returning very similar results. I've noticed this behavior many times now. Sometimes I will be exploring looking for a particular look and once you've run the initial prompt several times with slight changes it's like it settles on a certain aspects that it is unwilling to let go of. In this case, it kept doing the cell shading idea even after I removed it from the prompt. That is one of my critiques of MJ. It is sometimes difficult to get it to spit out what you are looking for. While other times, the first or second prompt and it looks like a million bucks.
I didn't think you were downing my methods. I was genuinely curious if you had an idea how to better test this stuff 😁
Do you ever hit any of the feedback buttons after an upscale? I've heard that might flow into the model/algorithm as "signal" and may reinforce preferences. But I'm uncertain how impactful it really is.
You are incredible thank you so much for your help and in depth guide it helps me so much!
On the 9 of February I am going to give a basic course on MidJourney where I live and your deep knowledge helps me as a beginning teacher in this beautiful world of Colab with AI to bring our imagination to life!
So glad that my videos are helping people around the world 🤗
This guy makes pressing that Bell icon on YT sooooo damn easy! Im subbing and under 10k subs and taking a screenshot of this to make an NFT later for sale as this channel is going places!
thank you. Great series on Midjourney man, beleive me, ive consumed a SHIT tonne on YT and your vids are gold.
Humbled 🙈🤗💚
Convinced me to subscribe!
@@zerocarbdoc ❤️
@@zerocarbdoc Thank you so much! Will try my best to keep you subscribed too 😀
Try the Fibonacci prompt formula> beautiful woman::1 warrior princess::3 pink hair::9 barbarian world::99 wearing brassiere::999
Don't need to. I already know what'll happen ;)
oh you have this fancy lamp in the background !
You might be interested Christian in the series of vids I’ve just released on how to use A.I. generated images to improve your mental wellbeing? Not getting any views but still might be worthwhile to look at an episode? Just adds a new dimension to A.I. use?
I don't think it works how you describe it, and it wouldn't make sense that it does. What midjourney tries to do is extract populair meanings from the text, it uses bits of those sentences to find popular meanings and order does matter. For example, use: "praying, a mantis --seed 1" and "a praying mantis --seed 1". The first assignment will try to create a praying mantis in a praying posture, the second will just make a praying mantis.
Midjourney doesn't understand a sentence in a classical sense, it extracts by association and tries to match that to what it has in its database. That is why an often used combination of words can dominate others and it is more creative with uncommon combinations.
I don't disagree. In fact, I fully agree. Midjourney doesn't understand grammar nor punctuation. So most of it is irrelevant. That wss the main point of the whole video. 🙂
May I ask what lead you to think otherwise? Genuinely curious 🤔
@@TokenizedAI What @merdanethubar-sarum9031 is saying is exactly the main problem of this video, and why I am asking to make a response. MJ is a Natural Language Processing (NLP), you created a sentence that makes absolutely no sense, with 5 contradictory Aesthetics references, a character reference, and a misplaced color. You can change this mess around and MJ will always give a similar result, and got even worst when you locked it all in a same seed... My friend, I love some of your videos, but on this one you are really missing the target and making people who still learning gets very confused... :(
@@AIArtistTV I think I understand the point you are trying to make.
I chose these prompts to keep things simple. And tons of people write their prompts like that and they are happy doing it that way. In many cases, crafting full sentences doesn't make it easier for them, but actually the opposite.
Are you saying that punctuation makes a difference, as long as you're prompting with full sentences? My own experience suggests the contrary
@@TokenizedAI by Christian Heidorn The main problem is that you are treating MJ as a random words processor, and MJ trying to find meaning. As the starter of this thread precisely refers to. What is the meaning of this promo you are using:
"beautiful woman, pink Cleopatra hair, warrior princess, bavarian culture, barbarian warlord"
This is not English, my friend. MJ is trying to guess what kind of woman you are trying to create. Got it?
Yes, many are making this mistake of treating MJ as a random words processor, it gives results, but is very sad you make a video name as "TRUTH about Prompts in Midjourney" and perpetuated the most wrong way of doing it..:((((
Going further, thinking here, a prompt should be:
/imagine prompt:character portrait, cleopatra as a bavarian princess with pink hair
(THIS is an English prompt)
The other one would be (because these are totally different aesthetics):
/imagine prompt:character portrait, cleopatra as a warlord warrior with pink hair
Get the best of both results and BLEND, maybe you have you had in mind.
"Let's talk about the secret language of MidJourney...?"
Is a prompt organized, clean and with proper english meaning on the subjet. :)
I hope it helps.
PS: And locking your whole process using the starting noise (seed), and expecting to see relevant changes, you locked and sabotage even more your whole experiment. :(
@@TokenizedAI Your video suggested that word order is also largely irrelvant. This isn't true because there is meaning in word order. Not in the sense of "the first is the most important" but in the sense of "Midjourney gives more weight to popular word combination when they are put right next to each other". I have given an example, just try it and see the difference.
good video --was wondering is the double colon a midjourney thing or does it apply SD? and do the examples apply to all platforms?
I'm pretty sure that's a Midjourney thing. Especially since it's behavior has changed a bit between v3 and v4.
Found this video to be quite valuable. thanks!
Glad it was helpful!
Great video man!!! What is the purpose of adding a seed number upfront?
So, some people think that I shouldn't be doing that. But I completely disagree. I'm looking at the prompts like a mathematical formula. Unless you fix the seed, it's going to change every time I enter a command. And we know that seeds are a major randomness-inducing variable in MJ (they add "noise").
So unless I use the same seed in every command (ceteris paribus), my experiment is absolutely meaningless. If I have a random seed, then my images are always going to look very different, no matter whether I change the prompt or not. Which defeats the purpose of trying to determine whether commas actually do anything or not.
Made a quick test during watching
lioness --seed 666
lion ess --seed 666
l i o n e s s --seed 666
there spaces make a difference, however even in the last one a lion apeared (and it tried to write liones)
I think future testing is needed. There are words that become new words with seperated spaces at the right place. They should give interesting results?
Edit: made
appletree --seed 666
apple tree --seed 666
they look nearly identical!!!!
Every tiny detail.....even a space, makes a difference. But only because it's interpreted as "interesting noise".
However, that doesn't mean it actually gives you any extra control over anything. And that's what this is about.
@@TokenizedAI lioness vs lion ess changed the gender, there is influence. But it is way more limited than assumed.... see the appletree..........
Well thanks sir! Now we master each concept of prompt, seeds and text weights, I guess we eventually need to figure out an effective way to create multiple discord accounts: this 25 limit is annoying :p
Apparently you can win/earn free subscriptions if you're among the top raters for a specific day in the offiicial Discord. Though I'm not 100% sure about how that works.
I do feel that using :: even without a weight number to that part of the multi-prompt does make a difference in how much the AI focusses on it I get better results by multi-prompting the description of the subject, the background and then the style. This might also be fiction but as long as it keeps working, I don't really mind 😅
I'm a big fan of perceived progress/benefits 😂
Surprised at how little difference word order made, given the importance of the attention layer in the underlying model. I'm presuming it is because word order tends to have very little importance in such keyword type lists, which in the initial training data would presumably be simply alphabetised. I presume if you are writing a verbose prompt it begins to matter more - eg. "Cat resting in a basket atop a tower at sunrise" versus "At sunrise, atop a tower, cat resting in basket"
Yes, you are correct. Word order is much more important if you're using full sentences or phrases.
Thank you so much for putting these videos out, I really appreciate your systematic approach that’s very much in tune with my own approach and for providing such useful prompting insights for the mj newbie that I am. 😂
Thanks for the kind feedback. The videos are a good reflection of how I experiment and learn via trial and error. It does require one to dig rather deep into the topic, which is somewhat rare in most TH-cam videos 😅
the flow of this video is amazing,
Thanks :)
Wouldn't you get everytime the nearly same image by defining the first one be the seed image? So for my understanding you told MJ with your first Image what images it should generate no matter what you are typing, isn't?
That's not what a seed does though. Seeds are input variables that just define an initial "noise" input.
"a blue house --seed 1" and "a green frog --seed 1" abviously do not render the same image, despite using the same seed.
I'm not sure why, but a lot of people seem to misunderstand the purpose of a seed.
From what I noticed, mainly in image 4, when you had // you had some criss cross pattern in the back, with ( you had arches in the back and with " you had as you said more hair and pointy things. what do you think?
Yeah, but it's still meaningless. As per the MJ team, all it does is add "interesting noise". So you might get interesting results, but you'll never know why or what it means. And that ultimately makes all these symbols fairly useless. Unless you know exactly what something does, it's not really giving you more control over anything.
But doesn't the '--seed 1' parameter influence the outcome of the painting in terms of style?? Doesn't it practically apply its own style?
No. That's not what a seed does. A seed is just an input variable that is meant to inject random "noise" into the algorithm, so that you always get different results every time you enter the same command and prompt. For example, every time you hit re-roll, it changes the seed as well. Seeds do don't apply any "style" or anything like that. They are just one of many variables.
The reason why you would fix a seed is because you want to "stabilize" the randomness of your outputs. In this case, I did it so that I can see whether the other changes actually change anything at all. In any normal situation you would do it because you liked a particular set of images for their composition and you want to fine-tune your prompt towards that same direction.
Hey nice vids, just 2 quick questions please.
1) This IA has some troubles with text right ? Do you have any tips for that ? Are do we have to use another software ? (If another one, do you have sugestion please ?)
2) Can we use :: and , at the same time ? To use weight on few words and not on others ?
Thanks in advance !
1) Don't use it for text. Midjourney currently isn't capable of producing sensible text. Doesn't matter how hard you try. Anything that looks correct is pure coincidence.
2) Sure, since the comma doesn't actually do anything, you can keep it in there. I use commas all the time because it's easier to read.
I like your approach, you make it easy to understand.
Thanks 🙏🏻
💯 Another great video. I've been struggling with syntax lately, this is just what was needed. TY again!
Practice makes perfect!
Hope you get well soon :)
Thanks. I hope so too. This one is actually quite stubborn. But I think I'll be fine in 2-3 days.
That was very comprehensive and useful. Thanks!
You're welcome!
Can you make a tutorial to learn how to create your own 2d assets for platform games like Rayman Legends or Donkey Kong Country?
If you'd see my list of video topics I want and need to cover, you'd be terrified 🤣
It looks to me from your video, that quotes make a focus on detail on what is in quotes..
Maybe, but which detail exactly? I can't really say that it does this in a reliable way. So it's not really useful.
Amazing channel, amazing video. Thank you for sharing all this. It is reeeeeeeeeeeally helpful.
My pleasure!
There is another secret about text to image generators. The AI uses images from artists from all over the internet without their consent. Which makes it probably the biggest copyright infringement of the last two decades, or even the last century.
All these generated images are made on the backs of artists, photographers and all other image content creators.
I hope as many people as possible realize this before posting their generated images online, claiming "I made this".
Thanks for feeding the algorithm 🤗
@@TokenizedAI It is more important to spread this mesage and discourage the use of AI in order to produce a final artwork product.
Also thank you for putting on the heart to push up my comment
Feel free to comment as many times you like. I like and heart literally everything 🙂
I would the problem in your methodology is that it is you doing prompts over and over under your one login which the ai recognizes as all related. The test would be if your prompt put under your login and then also a non-related login give the same results, or near same.
This assumption (that your results are somehow specific to your account) have been refuted and debunked on multiple occasions by the MJ team.
The double colon didn't even give you what you wanted
You asked for pink hair and it gave orange
If you see orange, then you may want to do a checkup with your eye doctor for Tritanomaly. No joke.
@@TokenizedAI Sure, Jan
@@shockadelic How friendly
@@TokenizedAI 🎵I'm not here to make friends🎵
@@shockadelic Evidently
NGL the more I dig into AI and programs/features available now.... I get kinda worried. As most of the population probably does.
I feel so ... foreign for one. I just started basic coding, and thought it was crazy in itself... but now that I can type into CHATgpt all sorts of questions about coding, I am able to relate it to knowledge I already have. It's like having a person that has been coding 30+ years, right next to me, answering every question I have in full detail...that's just bonkers!!
What do you think? Elon mentioned the risk of this already, but I didn't grasp what he was talking about until I discovered A.I. uses.
Careful with ChatGPT. Especially in areas where you don't necessarily have expertise. Very often it sounds super comvincing because it answers with such confidence. I've seen examples of it being horribly wrong but sounding very convincing. For example, it just made up Laravel functions on the fly that don't even exist. But because it used the typical Laravel conventions, I actually had to look it up first and confirm that it was bullshitting me.
ChatGPT is currently more style than substance. But since most of the media types don't look beyond the surface, it's being hyped up way too much. Yes, it's extremely useful, but it's not an outright job killer in most cases. The clever people are just going to be more productive than they already are.
I have friends who do a lot of writing in thier jobs. They seem worried, mainly because they don't understand the tech and I'd argue they aren't the most tech literate people anyway. If they bothered to actually care and get into the topic, they'd easily be amongt the first out of the gate. And yet, it grabs their attention for 2-3 weeks and then they return to their regular old ways without any consequence.
Not gonna lie....they are the problem, not the AI.
@@TokenizedAI Thank you for the response! My mind is literally melting haha. What a time to be a live.
🟠 Hi Christian this is a great video, thank you so much for going through this and clarifying. I have been wondering how to figure this all out, in which you have effectively clarified. One question still thou, i was not able to understand why princess::20 and culture::20 equal 93%? what does the numerical value of ::20 represent? It seems it is not percentage, otherwise then it would effect only 40% of the image instead of 93%? This was a bit confusing for me since im still learning // figuring things out. thanks! :)
If you have 5 segments in your prompt and three of them have a weight of 1 each and two have a weight of 20 each, then you assigned 43 weight points overall.
So the two weights that have 20 each (40 total) represent 93% of the overall weighting (40/43=0.93) in the prompt. Each one has 46.5% importance. The remaining 7% are shared by the three other segments with a weight of 1 each. So in the greater scheme of things, they are almost irrelevant.
Weights are relative, not absolute. So for example, these 2 prompts are identical im effect:
red house::20 green garden::20
-and-
red house::1 green garden::1
@@TokenizedAI Thank you for your great explanation. I have a feeling by your last name you may also be German like me, and is why i really enjoy your clear and well thought out style of explanation. Instantly subscribed after one video, fan for life! 😄😄 thanks again 😊
@@TokenizedAI I think this sentence "Weights are relative, not absolute" and the example is so important that it should be pinned at the top of your comments, just saying ;)
Yup, I'm German 😁 Thanks for the sub!
I didn't expect the channel to grow this quickly. The first 5k were rather smooth sailing. But a fair bit of negativity is creeping in now. Need to learn to deal with it I guess.
I would, but I can only pin 1 comment and I need the current one to stay up. I've been getting a lot of flak for this video because some "uber experts" feel that I'm misrepresenting things. Unfortunately a lot of people are getting bogged down on details that are largely inconsequential.
In all fairness, I may have tested with some unsuitable examples. But clearly, my main point is coming across enough: Commas don't matter.
Super useful! Thank you for making this video.
You're very welcome!
Cool, thanks. The word order doesn't really seem to have any impact. I often use prompts where the topic is in the middle or at the end and it still works fine.
Word order DOES have an impact when you are constructing full sentences (that actually make sense).
It DOES NOT matter much, if you're just using a string of random or loosely relatec keywords or partial phrases.
Best video I’ve seen on Midjourney and I’ve watched a bunch. Thank you. 😎✌️
Thanks!
Thank you for your research and explanation, really appreciate it! Safes me a lot of time 🙂
My pleasure!
deine Tutorials sparen mir viele fast hours, danke!
Bitteschön 🙂
Example where you removed spaces makes ne very confused.
What if you would mix letter order?
I mean sequence and separators mean nothing...
I was confused as well, but I guess their language model is able to identify stuff like this. Not sure how or why. I didn't go so far to actually mix up the letters though 🤣
Would be cool to compare results with the same prompt on different models on stable diffusion, e.g. the openjourney model claims to make you achieve similar results to MJ, thee is also a text inversion for stable diffusion (midjourney based). We all know MJ is better than stable diffusion. I discovered SD recently and playing with it a lot, and I'm quite satisfied with the results, it's like mining crypto, just I'm minng nice images on my GPU instead of paying MJ ;)
Midjourney v4 runs on top of Stable Diffusion. The main difference is in MJ's execellent use of embeddings as well as the natural language processing layer that it has placed on top of it. But at its core, any AI generator that's not built by OpenAI (e.g. DALL-E) is 99% likely to be using Stable Diffusion under the hood.
Hi, it's a nice tutorial. I get a lot of progress after watching your videos. I have a question: How to make the main object placed in a certain position on the canvas, for example at the bottom, or the top right? I'have tried several times but no luck, the object is always positioned at the center.
hi, I've seen the work of a user with id Julz#8275. I don't know how he is prompting to produce images with very high detail, and can place objects in certain positions.
If you know his username, why don't you just look at his prompts on the MJ web app?
@@TokenizedAII don't know how to get the prompt in MJ App. I've been looking for it but only the picture shows up, no prompt. I'm a newcomer to Discord and only been using MJ App for 3 days so I'm not very familiar with it. I am amazed by the work of other experienced users.
The prompt should show up right below the image. Kinda difficult to explain here in a comment.
@@TokenizedAI yeah, on #newbies's channels there is always a prompt above every picture that is made. But the picture I mean is on channel #v5-showcase and #show-and-tell. There are no prompts at all, just pictures.
Perfect example, how you make a video with thesis in mind, and even if evidence is against you, you still arrive at the thesis. Even if img change, you disregard it.
I didn't say the images didn't change. I said the changes aren't as significant as some people claim.
I appreciate the feedback though. Looking forward to a video with your take on the matter 🤗🤝
@@TokenizedAI I sounded rough, sorry for that. I really like all your other videos, good job on that! But this one... it just didn't resonate with me, as I disagree with the same result you showed.
Don't let it discourage you :) Have a great weekend!
No worries. Appreciate what you're saying.
Unfortunately, it seems like the main point that I wanted to make got lost since everyone seems to be focusing on 1-2 specific examples where (I admit, in hindsight), I may not have chosen the best examples.
The point I was trying to get across is the fact that punctuation is largely inconsequential. So many people swear by using quotation marks, commas or even brackets, when they actually don't do anything (other than add noise). This has been confirmed by the clever people in #prompt-chat as well.
I should have made clear that the order of the words obviously matters if you are using proper sentences in your prompt. I guess I just assumed no one in their right mind would put a real sentence through a blender 🤣
@@TokenizedAI I agree, as Natural Language Processor, v4 should treat punctuation as a 'interesting noise'. Not much more.
Any chance you can make a similar tutorial for Blue Willow?
I might...one day. But my focus right now is on MJ.
The best explanation for MJ only at your channel ❤❤❤
Wow, thank you!
great video. thanks for doing all these tests!
Glad you like them!
But, does each prompt have a memory from the preceding prompt? Because all of these tests are done in sequence immediately one after another does the AI remember the preceding prompt and, if so, wouldn't that effect the output?
According to what I've been hearing from people close to the MJ team, they do not.
So, in short brackets, commas, word order, are useless, but quadpoint + numeric weights and seeds are the real thing.
Commas are not necessarily useless if your prompts use full sentences. If they don't, then yes, they are mostly useless.
i wonder if it is possible to build a dataset for AI. If any artist do not want their work to be used to train the machine, he can put their work into the dataset. Once Ai picks an image from internet, it checks if the image exist in the dataset. If an image is over 99% similar one of the work in the set, it will pick another one.
Problem is, I'm not sure that's possible. The AI model itself doesn't contain the actual image itself. It was merely trained with it. You would have to cross reference with the original dataset, rather than with the model itself.
7:07 top left is perfectly balanced 🙂
I can never get buildings like skyscrapers, or anything with straight lines or get the perspective I want. Am I doing something wrong is this not possible yet ? lol All my buildings look like sketches from a drunk person. Are there also negative prompts in MJ? Edit: I played around with colons and its already helped with straight lines!
Hard to say without seeing your exact prompt. Yes, negative prompts exist.
i really like your videos, keep them up mate :))
Thanks a ton! 😀
Hm... I wonder if in a 4-image prompt the relative position mean something, for example, your repeated women 1, 2, 3 and 4 always appeared in the same positions. In other words -- 1 would be tame, 2 epic, 3 complex, 4 went to hairdresser. Does character 4 always appear in position 4? Or is it just my imagination?
There is some consistency and some randomness. Overall, I have no idea what the logic is. It seems that some seem to rotate, while others don't.
Does using the make variations button preserve the same seed or is that same as re-roll? Also, I think in one of your videos you used [brackets around the various prompt segments for image weights]::3 This was a great tip for helping me separate image segments in my brain.
Actually, I don't think I've ever used brackets. You must be confusing me with someone else.
Variations will change the seed as well. However, it's important to remember that variations don't process prompts the same way as a regular /imagine command, since it sticks extremely close to the original.
Unfortunately the brackets don't work also when it comes to merging sentences together...
Why would you want to merge them together in the first place? Like, what are you trying to achieve that's not possible with multiprompts/text weights?
@@spartiata71 How do you mean? Sorry for misunderstanding you. But I would be very curious what you mean by merging sentences together. If you could give an example with a prompt, I would appreciate it so much!
@@TokenizedAI So in order to preserve the same seed with variations you would have to turn remix mode and manually specify the seed, yes? Also, you are right. When I was talking about the brackets, I must have gotten my TH-camrs mixed up. I made it up to you by clicking that subscribe button! ;-)
Thank you very much for this very interesting and informative tutorial.
Glad it was helpful!
I tried to get 2 characters like Batman and Joker or Superman. I always get same characters pics somehow changed like Joker but looks similar to Batman. Any suggestions how to improve?
Does your prompt look like this (very simplified)?: /imagine batman::1 joker::1
@@TokenizedAI Yes true. Actually I entered - Batman vs Joker. Tried Batman::1 Joker::1 but the result was having weight of Batman or vice versa whichever is mentioned first
@@theboldbroadcaster Yeah, so the problem here is that when you use weights like that, it will always BLEND the 2 concepts together. So it's similar to what happens when you just add 2 image prompts. It always blends.
It's possible to achieve what you want, but it's too complicated to really explain in detail here. One of my next videos is about this.
For now, give this a try. It's not perfect, but a start.
And yes, copy the entire prompt 😉
/imagine batman standing next to the joker in the streets of new york city, in the style of marvel comics:: the joker positioned to the right of batman on a new york city street, in the style of marvel comics:: new york street scene showing batman next to the joker, in the style of marvel comics::
Thank you
THank you for this!!!
No problem 😊
Word order definitely matters when you have long strings of keywords in a very large prompt. It uses the first 5 or 6 keywords, and after that it randomly picks and chooses words to use.
It matters and it also doesn't matter. The lenght of the prompt isn't the driving factor. It matters when you use full sentences or phrases. It doesn't matter when you only use singular or very short keywords.
Don't apologize...just don't get sick
Thanks 😅
Wouldn’t they all look the same if you are referring to seed 1 each time-as it remembers that original seed? Or didn’t you remind it of the original seed between prompts?
If the only thing determining what an image looks was it's seed, then why on earth are we wrtiting prompts in the first place?
Seeds are a critical influence, but it's not like they are the single most important thing.
@@TokenizedAI right, but if it took the seed as the base pic, would you even know if the commas, dashes, etc were making a difference? I guess I have always used a more narrative approach and v5 responds even better than 4 to my style. But very interesting…I have been curious about this for awhile. Thanks for taking the time to run the tests.
This video is gold and very well done. You earned a subscription from me. :)
Awesome, thank you!
ok followed for once a tutorial of real top notch quality, good work !
I appreciate it 😊
@@TokenizedAI this is really high quality content, very clear, concise and very well explained, and I truly learned new things !!
thank you again :)
Thank you!
🤗
I used to write whole novel of adverbs and didn't get what I want 😂
🤣
Is there a reason I see prompts with, for example: 8k three times in the prompt ? Why not just use 8k once?
"8k" and anything like doesn't do diddly squat 😉
@@TokenizedAI thank you! And the repeating in prompts? Detailed hyper detailed extreme detail.. are those just rookie mistakes or do they have an effect?
"extremely detailed" might do something but largely depends on the rest of the prompt.
Anyway you see complete gibberish prompts with lots of copied keywords, it's usually a newbie.
Repeating prompts are a bit different though. Sometimes they're on purpose when someone's using multiprompts. Depends on what the user is trying to do.
I’m guessing you’re using V 4 here. My own experience with many thousands of V 3 images is that text order does matter, though I haven’t tested this with V 4. But I’m talking about style words, not nouns
Funny how everyone seems to be focusing on that one bit. I think I should clarify that text order absolutely does matter if you are using proper phrases and grammar. Which makes sense though. And I can't think of any reason why anyone would want to mess around with the order. So I don't really understand why some seem to be up in arms about the example I gave.
Thing is, most people aren't reall using tons of phrases and if you only use basic keywords and concepts like I did in the video, then it doesn't really matter. Personally, my prompts are a lot more complex. But those aren't helpful for simplifying and explaining things. Especially not for Beginners.
But it's funny how people seem to be missing the actual point of the video. My main message was supposed to be that commas and punctuation are completely irrelevant. 🤣
Why didnt you test + becaurse when you use + it showing veird on the web site MJ. It decides sentences t different part of the prompt. Try it.
I think your confusing "+" with multiprompts. When you see multiple segments with a + between them on the Midjourney web app, then it means that the person actually used "::". That's how multiprompts are displayed there.
@@TokenizedAI ok thanks.