Emulating Artist Styles: The Ultimate Guide
HTML-код
- Опубликовано: 31 май 2024
- 📙 Midjourney COURSE mastersofmidjourney.com Beginners to Advanced
🚀 FREE Midjourney Cheat Sheet tokenizedhq.com/freebies/mj-c...
🔗 FREE Promptalot Extension promptalot.com/extension
🔗 FREE Supporting Material tokenizedhq.com/freebies/vide...
Folder: 2023-04-22 - Emulating Artist Styles
🌐 Check out the full blog post:
tokenizedhq.com/midjourney-ar...
-
📰 AI Newsletter 👉 Your Inbox tokenizedhq.com/newsletters/ai
🐤 Follow me on Twitter / chrisheidorn
💼 Follow me on LinkedIn / christianheidorn
📸 Follow me on Instagram / christianheidorn
💬 Join the Tokenized AI Discord tokenizedhq.com/invite/discord/
🎬 Prompting Methods & Experiments Series 🎬
Playlist: • Prompting Methods & Ex...
Part 1: Blend Images with Reverse Engineering • Blend Images with Reve...
Part 2: Emulate Artist Styles • Emulating Artist Style...
📺 Recommended Related Videos & Playlists 📺
Commands, Syntax & Parameters: • Commands, Syntax & Par...
Prompting Methods & Experiments: • Prompting Methods & Ex...
Character Design Series: • Character Design with ...
Storytelling: • Storytelling with Midj...
Watch Me Prompt: • Watch Me Prompt in Mid...
This Brand Does Not Exist: • This Brand Does Not Exist
Monetizing AI Series: • Make Money with Midjou...
-
I'm really curious to see how people will respond to this video.
I personally see 3 potential reactions:
1. Some will be intrigued by the underlying questions
2. Some will think it's pointless (e.g. waste of time)
3. Some will be snarky, pointing out how these models were originally trained
If you fall into the first category, then I think you'll find some of my results surprising and maybe even fascinating.
I originally asked myself whether it was possible to emulate or replicate an artist's style in Midjourney. Obviously, we all know that we can simply reference the artist's name in the prompt. But what if you wanted to avoid doing that? I can think of many reasons why some may not want to. You might not care, but hear me out.
So tried to figure out a more or less formalized process for crafting a custom style prompt that would achieve a similar, albeit not perfect, replica of any given artist's style. And while the process did not work equally well in all examples, some of the results were surprising, to say the least.
Things got very intriguing when I started to compare prompts with the same seed. And I already know what some of you may be thinking: "Well, you used the same seed. Duh!". That's not how seeds work. If the only thing controlling an image was its seed, then why on earth are we adding stuff to our prompts!? Well, because it's not all just in the seed!
So trust me, some of the examples I'll be showing you today are downright shocking. In my opinion, it raises some very interesting questions that deserve some of our attention.
When we reference an artist's style in our prompts, is it really using that artist's style? Or is it just an amalgamation of the style of the particular art movement that the artist belongs to?
If this piqued your interest, then make sure you tune in!
-
⏰ Timestamps ⏰
00:00 Emulate Artist Styles
00:30 The Reference
02:24 Extracting a Style's Essence
07:19 Apply the Style
09:32 The Showdown
-
#midjourney #midjourneyai #midjourneyart #aiart 
Наука
This man is a GENIUS! In a few minutes he (basically) recreated a famous artist's style without referencing his name in the prompt. The images are strikingly similar, and for some of the images I even thought that the created style was better than the artist's style - just my personal preference. Thanks for this awesome video!
🙏🏻
Thanks so much for this video, Christian. It's especially helpful to see your thought process and how you set up the reverse engineering experiments. Lots of lightbulb moments!
Thank you Kimberly 🙏🏻
GPT4 is so polite. It warms my heart.
Chris: Thank you sooo much. You have no idea how enormously useful your video is. ✨✨✨✨
Heh, I think he used GPT 3.5
About time Christian, I've been waiting for your next video!
2:51 'Don't lie. ' 🤣
Great video! Fantastic way of referencing a style. Take it a step further and mix artists' styles and we're onto something even more. The inspiration is endless. Thanks for this. 💯
That detail is far more important than you think 🤣
Thanks a lot fornyour incredibly informative videos Christian. I really love the experimental approach that you convey in most of them and how concise they are.
Experiments is what this is all about
Thanks for your videos and tutorials… appreciate your works as always, keep it up Bro!
AHHH this person is up set with you tube!!! I never got notified about this video!!! Thank you Mr. Heidon 😊
You're welcome
Christian, another excellent video. I'm loath to using an artists name because it's their art after all. I do like to follow their "school" of art though. I've been doing a kludgy job of researching the artist style to come up with how they were trained and using the characteristics of the "style" they were taught to create similar looking art but your method is genius and so much faster. Thank you for sharing.
Agree on all points
Thank you! It was very informative.
Impressive! Those examples look almost identical!
This is a great video - once again! 💯 thank you for this cool tip!
Pleasure
Great work, man. Thank you for this.
My pleasure
Thanks for this! I noted that artists style applied using their name work well although most of them add elements that are not always appropriate for what we want to achieve: some will depict a certain period, or favor one ethnic group or even add people to a scene. You are providing us with a way that has a great potential to get around this! Thanks again!
Yeah, that's I insisted on no references to characters or superheroes. Otherwise it would constantly add them.
Just discovered this vid. I'm not finished watching it, but I found it interesting that you have an ethical hangup over prompting using an artist's name, yet you are perfectly willing to prompt midjourney into giving you output that uses his style. You must know full well that midjourney will still be accessing work from his portfolio in the dataset , so in the end, seems like a lot of effort for very little gain. I do respect the process however, and discovery is key, so if you stumble into a method of gaining consistency across subjects as a result, that I can get behind.
It's not an ethical hangup. I have no problems using artist styles. I just don't see any value in doing so because I find it far more interesting to figure out how to achieve an effect with words.
If I reference an artist, then I have no idea how to reproduce parts of their style. I'm not interested in pretending to be a particular artist. I'm trying to create new things from different bits and pieces.
Fantastic - so interesting - and useful! 😊🙏
In all last 4 pictures the subject from the images created using directly the artist's name is bigger, it fills the 2/3 of the picture. To make the two sets of pictures even more similar, you can add something like "close-up"
Christian, excellent video as usual, plus I think you've hit on a central idea for systems like MJ
You could identify the concepts of anyone's art by using /describe on a large selection of their works, then identifying the N (10, 20, ... your choice) most common phrases out of all the phrases used for each work by all evaluators.
And I think you've hinted at embedding -- since the training images were probably described by human evaluators, the "in the style of" set of artist style descriptions could be essentially embedded vectors of smaller concepts.
Yeah, the last point is exactly what I was thinking.
Nice choice for Art Adams, he's long been a favorite artist of mine. I strangely once found myself in a conversation with Art and Mike Mignola where they were talking about how great Bill Sinkevich is. Art mentioned he once told Bill about a New Mutants cover he always loved, and Bill offered to give him the original. Art accepted, but only if Bill would choose an original piece from Art that he could return in favor. Art received the New Mutants cover, but never received a trade request from Sinkevich.
Art jokingly concluded that it's because Bill doesn't like any of his artwork, and to this day, that is the only thing that drives him: To make a piece of art that will impress Bill Sinkevich!
Ha! Beautiful anecdote 😁
This method makes me feel better about using certain styles 👍 thanks.
Awesome 👍
Many thanks! I learned a great deal including (1), how to use CGPT to optimize prompts for other apps / use cases, and (2) the value of a precise style guide to help improve my prompting.
Great to hear that 🙂
Excellent!
Hear that? A horde of Artist REEE into the void now that we can deconstruct their art styles.
😅
People don't like being robbed?
That's new to me.
Brilliant and next level informative. 🤔
great vid! tnx!
I totally love your videos! What platform/ software do you use for video editing?
I use the Affinity suite
Thanks Chris, for this video, It is so informative, but I'm lazy to ask all of these questions 😅 to ChatGPT so I took your style prompt and asked chatgpt to generate a similar but for Van Gogh, and It works very well 😍
I think Van Gogh is probably a bit too easy though 😅
can you get the art style details from an image with an unknown/unfamiliar artist? I want to use midjourney to create a collecrion of images based on my personal drawing style but i dont known how my style be described.
Hi @TokenizedAI, I saw this first in Twitter, what can I say? Thank you very much for share your work with us, doesn't matter if there will be controversial, you always give us the ideas and how we can replicate this is more than anyone here can ask. So thank you very much.
My pleasure
This is the first time I've ever gotten a good result from AI generated art. Your tutorials are brilliant, while still being realistic and hype free. Thanks!
Thank you so much 🤗
Hi thanks for your great content, I have see you with Philip Ander, Now I am subscribed, I have question about text for Print on demand, Philip say Dall E 3 is better than Midjourney text to image what you think about it?
Have a good Day
I tried to get an illustration in the style of Quentin Blake and got hilarious results. Still very limited for certain styles.
Nothing's perfect
I have an observation that might help you moving forward with your understanding of MidJou
If you pay attention to MidJou within Discord during the creation process , you get a handle on the evolution of the art.
Typically, a base coat of color.
Then a proportioned shape as per our promts.
A general outline, with simplistic details
Then a more polished phase to the art . . .
This is one method artists are taught to the creation process.
That said, the colorization of Mr. Adams and the AI work popped out at me right away.
Strip away the pencil marks and the images have an under tone of color that is similar.
P.s. I was a pencil and Ink artist in my younger days.
All Black and white, but using the lead to shade and shadow.
Agreed. I have often seen this process in action when watching a YT video of someone does a speed drawing.
You mean the diffusion process?
@@TokenizedAI no, I mean YT videos where an artist is painting, for example street artists painting a portrait of a person
@@MrJoe1199 Yeah, I got what you meant. My comment was in response to the OP. 🙂
In the beginning I thought you were going to use MidJourney's "Describe" command. Wouldn't that get similar results much faster?
Not necessarily because /describe returns a bunch of nonsense words too, which don't do anything. MJ has stated themselves in their FAQs that it is not a source for "well-formed" prompts.
Thanks.
To give you even more context, I have worked on a Childrens Book for 3 months, it was my very first book ever created really put a lot of work into it. For my second book I was about to make a coloring book, however, after 3 days of trying out Midjourney's different Coloring Page promts I can say that they are not clean, fully understandable from customers perspective those Amazon's feedbacks. In other Words Midjourney's Coloring pages still need a lot of Cleaning work, you have to vectorize and clean the unnecesary shades..etc you end up working on a low content book to much.
The problem isn't MJ though. The problem are lazy KDP sellers churning out huge amount of crap. 🫤
Midjourney works with what it is trained on.
I like very much your videos, perhaps I have a topic for your next video. There are so many videos about Midjourney Coloring Pages how easy they are to create...etc, however I do know for a fact that Amazon already has a lot of bad feedback from customers buying AI made Coloring books, which are very bad quality wised. I checked personally hundreds of these feedbacks to identify the issue with those Books. Based on customer feedbacks they are no clean, mainly because of the shades or they are too detailed. I hope you will find a workaround by advanced promting to create Really - Clean Quality Vector kind Coloring book pages. I'm 100% convinced despite Amazon aproves nearly any kind of coloring books on their website the majority from the Ai space is not yet 100% Book quality and they never rank.
I'm interested by your online course, but I can't fond the duration or the number of videos for every chapters. I wanted to have an idea about the time needed to follow it.
Hi, the course is currently just over 5 hours of content (new stuff is also added over time). Each lesson is usually between 5-15 minutes, so the time commitment is low for each learning session.
The 2 biggest modules right now are 2 and 3 on the Foundations of Prompting and Advanced Promoting Methods.
I hope that helps? 🙂
You should always use the artist's name if you're going to change *nothing* about the art style. Learn to be efficient. Now, if you're going to change element(s) from the art style, this could be useful.
That being said, this was useful for me to learn about how an art style comes to be. I didn't think MJ could replicate the art style using "key words", which is probably gonna make me experiment to find unique art styles.
It was an experiment, so efficiency is completely irrelevant in this context and also wasn't the objective.
And if someone simply doesn't want to use the artist's name, that's their prerogative. No need to impose something that's entirely up to them, whether efficient or not.
@@TokenizedAI I'm not trying to impose anything, but it probably came off that way because of the way I type. It's just a suggestion, you're free to follow it or do whatever suits you.
On a side note, I just started using MJ a few days ago, and it seems to be struggling to create *somewhat specific* scenarios. Is this common?
@@TokenizedAI One more thing. I've been watching a bunch of MJ videos, and some of them had a "max upscale" option, but I can't seem to find it as it seems to be removed from the list of features. Is there a better way to upscale in HD now while retaining the picture's quality or is that just gone?
Since you've only been using Midjourney recently, you'll quickly notice that things change quite a bit with every version. Technically, MJ is still a beta product.
I recommend reading through the announcements in their Discord to get a better idea of how things have evolved. Or just look at the official documentation to see what is available in each version and what isn't.
Oh, the controversy! While watching, I considered whether laws might protect those who avoid using artists' names and whether doing so would make a difference. This is especially relevant if an AI detector can measure the percentage of a style used, even though such technology is not yet available. I imagine something will be created in the future along these lines. I also wonder if using an artist's name could ever be considered copyright infringement (there are good arguments both for and against this). So, it's unclear where the law will stand - it's possible that using an artist's name could end up being completely legal.
I think the most compelling argument that AI-generated images are NOT infringing on any copyright is how AI diffusion works. Given that diffusion models create unique images from random noise grains and apply your prompt to those grains, it seems less likely that infringement is occurring. From my experience, reverse image searches via Bing, Google, Yandex, or Tineye, etc. always show that the images the AI creates are unique and not copies. On the other hand, the biggest argument that they are infringing relates to taking away an artist's market share while operating in their same genre. However, the overall debate seems to favor how AI works, making it more likely that AI will prevail legally.
When GPT-4 computer vision arrives, it might be the ideal way to describe an artist's style without mentioning their name. However, describing the style might be too similar to using the name anyway, because they may develop style percentage detection in the future, which may not offer much legal protection. Additionally, the MJ describe feature emulates styles effectively. While avoiding using an artist's name may appear more politically correct, the output differences may not be significant in terms of how copyright law ultimately evolves, especially when style percentage detection becomes a reality.
Are artist who are complaining about AI being hypocritical? Rutkowski clearly takes influence from Frazetta? One thing to consider is that artists who claim AI is infringing on their art may themselves be “infringing” on certain aspects of artists who came before them if they were to use their same standard of infringement they are claiming the AI image generators are committing. So, at what point is supposed copyright infringement actually natural inspiration and influence that artists have always had? AI is just so proficient at inspiration and influence that it looks bad politically, but in reality, AI image diffusion models work in much the same way human artists do when they're inspired. Do they not?
You’re right this is dripping in controversy, so it will be fascinating to see where the law lands on this issue. It could go either way, but I believe AI technology has a stronger argument for not being considered infringement considering how it works, scientifically. Another approach companies could adopt is creating completely royalty-free datasets of imagery, hiring artists to create images for the dataset, and granting the subscribers 100% legal rights to create the images using AI. This seems to be the approach Adobe is taking. However, a criticism of this approach Adobe is taking is that it assumes creations from platforms like Stable Diffusion, Dalle, or Midjourney are copyright infringement, which is not yet clear AT ALL. Personally, I don't think it's infringement; it's just advanced technology that does what we've always done - using inspiration and influence to create new art. It just does so much better due to machine learning, neural networks, and AI - and so right now in history that skill the AI has is so high level it looks bad politically, even though in reality its no different than what artists have always done.
Well, according to the hordes of AI haters on Twitter, I've effectively taught people how to copy the art of other artists. Which is kind of weird because that's what they said about these diffusion models already. So which is it? Now they're bothered because I explicitly avoid mentioning an artist. But for some reason they can't tell me who owns the term "intricate lineart".
It's complicated 🤣
The problem is the Ignorance of the Lawmakers and the fact the Artist pushing these lawsuits are either misinformed on how Diffusion works making reference to a non existent Dataset (the Models are trained on Datasets, not built into them) or purposely lying to the judges so that they can get this technology banned.
@@TokenizedAI Keep supporting Open Source AI generators. Even if Stability AI is sued to Bankruptcy. Their work is on the net ready to be used thanks to Open Source.
Honestly. This method is actually future proof because The Latest Base Stable diffusion model WILL NOT ALLOW YOU TO NAME THE ARTIST when using their art style.
This Method allows us to fight Censorship the Artist will harrass Stability AI into Implementing.
So this method Bypasses any "safeguards"
Hey Christian, any idea how to do this with an artist that is really obscure? Someone not recognized by ChatGPT?
I don't think it works it the artist isn't known to ChatGPT
Interesting video, what would be really nice is to use a certain style to apply to images, and translate them into that style, while this works, the outcome is not consistent and it also varies depending on what image you want to apply it
Nothing is ever going to be consistent and that good. How boring would it be if everything was literally done with the push of a button?
@@TokenizedAI in my eyes, it wouldn't be boring at all, imagine having the style of Batman the Animated series, and use it to generate new adventures for the characters, even for pure fantasy and it would be a lot of fun, or being able to create self portraits using the Simpsons artistic blueprint...it would be amazing, and highly probable to happen in future iterations..
@@alterverse_ai It would be boring because anyone could do it. Value lies within the things that are difficult, not the stuff that anyone can do.
@@TokenizedAI that is the beauty of AI being universal, anyone can use it, I might not know to draw cubism, or suprarealism, but with my creativity and AI I can do it :). Where the real originality comes into play is when you can generate a style of yourself that only you know how to achieve it, until then, you are merely imitating others :).
Could you show in the video how to improve midjourney pictures quality to get them ready for printing?
To do that you must improve DPI of photo at least on 300, you can do that online, also in prompt add ''print'' , ''for print'', ''sticker''
I can, Photoshop.
MJ can only go so high. At that point you need to use an AI upscale tool. There are so many. A free one I've used is bigjpg.
Christian explains a little in this video. He uses Upscaledotmedia but Jonathan Roy's bigjpg also works. ruclips.net/video/KV8OFn59Xco/видео.html
Wow
thanks a lot, can you make a video of how to make an ai aet based on a random image we find, how ro extract info from it and how to replicate it, most of the time the images i find have no artist name which is why I am asking if it is possible
Just use /describe?
@@TokenizedAI ok i forgot about that😂
Isn't it the case that v5 is better with actual descriptions instead of just using keywords? That's at least the official word.
Define "better"? It just works better with more detail. Doesn't mean that is must be natural language, as you can clearly see in the video.
Here’s the problem. MidJourney limits the number of tokens it will process. An artist’s name is a single token. Your prompt contains 15-20. That SEVERELY limits the number of tokens you can use to describe the scene. It’s ok if you just was an interior of a spaceship. But if you want to describe the subject in any detail, you would be better to just use the artist’s name.
Sure. But obviously that wasn't the point of the exercise nor is it to copy an artist's style verbatim. The objective is to extract elements of the style.
And MJ's token limit is roughly 77. Yes, the first tokens are most important but the rest is still being taken into consideration.
👋
Artists always borrow other artists' styles as long as they make their own work original.
Agree
I get where you're going with this, but referencing Adams in ChatGPT is still referencing the artist, even though it wasn't in MJ. I think the best way and more "pure" way is to be generic when asking about comics in ChatGPT. Just generalize the question based on the style of comic you want to create, not the artist. Then break it down when you feel some keywords that you feel are the best fit. And honestly, I'm not here to criticize or knock the method - I enjoy your content. Just commenting on the goal of this video. But good job on all the content.
I disagree. If I were to ask an expert on about the best way to describe someone's art and then used those words, would you call that referencing an artist? I wouldn't. I could have gotten the same information my simply reading and researching public information. In which case I'm not referencing anyone at all.
@@TokenizedAI Well, I was going off the assignment of not referencing the artist at all but still being able to capture the look and style. You can spend time using key descriptors to get a look similar to any artist without using the name at all. Would be like showing a piece of work to someone that is unfamiliar with the artist and ask for them to describe the art in detail. Again, not saying it's wrong or anything about this is wrong - I quite like what you did. But just simply getting artwork to look similar without using the name at all would be a fun task and definitely more challenging and rewarding :)
Oh yeah, that would definitely be considerably more challenging. In this case, I was more focused on outcomes (within boundary conditions) rather than the process.
I also think that despite the fact that you and I would probably find it very interesting, I don't think it would find mass appeal. At least that's what the stats tell me 😅
@@TokenizedAI I totally get it and I find myself going down a rabbit hole trying anything and everything - and you def got a good thing going and know your demographic - that's why you're good at what you do here and why I can't do what you do here. haha. Great content all around.
reverse engineering. 👌👏
This is a hell of a lot of effort to then refer to yourself repeatedly as lazy when the simplest method is to refer to the artist. Neither someone's name nor their art style can be copyrighted so anyone full of consternation about this can be safely ignored and ridiculed. Just use the artist's name and move on.
True. This method gives the added benefit of having someone's style, and then changing 20%, 30%, 50% or so on. Choosing the exact element that you want to tweak and how.
You can combine two or three artists styles in a premeditated and orchestrated way, rather than getting back a randomized mix up of the styles.
It also improves your ability to craft original prompts, as your mind is being exposed to the specific elements that contribute to the style.
Doing this overtime with different styles would make one a very skilled prompt Engineer capable of creating unique, consistent styles that other people won't be able to replicate
1. I consider this "research"
2. It allows you to craft derivatives of these styles.
Don't compare the effort. Looks at the options it gives you.
this is completely useless in stable diffusion since the model or the lora mostly dictate the style :(
The video is about Midjourney. What on earth made you think that this could be done in Stable Diffusion?
@@TokenizedAI 🤷♂️
I love the line "Don't lie" used with Chat GPT :))))