Midjourney Tutorials: Keeping Character Consistency
HTML-код
- Опубликовано: 6 окт 2024
- In this episode we'll cover the techniques of establishing a clear design language and staying true to it throughout the character creation process on Midjourney. We'll also delve into the methods of carrying that consistency into different art directions and scenarios.
#Midjourney #aiart #midjourneyart #midjourneyai #digitalart #art #design #tutorial
Follow Us
🌐 www.polyform.co/
👉 / polyform_studio
👉 / polyform_studio
It was not traning but an optimisation of seed. The name has only one influence. It is to prevent too much change. If too many textual things change, too much changes in the whole. It's like painting around a picture and the middle is your text. You can't train anything because nothing is saved on the server.
Thank you Adam. I truly appreciate this tutorial. 😊❤
Best tutorial on consistency.. Many thanks for sharing
Thank you for this tutorial. I am a newbie, and your explanation and examples of weighting were beneficial.
Very helpful video. Thanks
The thing I think everyone is looking for is being able to take a photograph or a drawing or a JPEG of a 3D model and image2image that into midjourney or stable diffusion, and, along with text prompts, having it spit out a new image in whatever style that’s prompted.
Now, you can KIND of get that with pictures using distinct silhouette. The problem is when you want dynamic posing, foreshortening along with interesting camera angles on the subject.
The Ai needs to be able to learn that all we want is for it to take the image we give it and alter the STYLE of that image, much in the way that a lot of these apps like PRISMA do, where they take your photo and turn it into line art or a painting. But we don’t want the IMAGE turned into that, we want the POSE and ANGLE turned into what the Ai has the capacity to do.
E.g. you should be able to put in a photograph of your punching at the camera and prompt the Ai via image2image with the text prompt “Batman punching in the style of Jim Lee, black and white” and have a render that essentially swaps out your image and pose with that of a Jim Lee Batman. Or a Pixar character. Or a Simpsons character, on and on.
It’s mind boggling that the Ai doesn’t have the capacity to do this yet, as it’s an incredibly simple ask considering the sort of apps that already exist.
You can totally do this by using your original image as a SEED image and then adding your stylization as a prompt, then use weighting to heavily weight your original image and lightly weight the stylization. In this way you could get the same image in the style of ANY artists.
@@adamkylewilson3080 I don’t think this is accurate (that I’ve seen).
Ai does a decent job of representing poses that are observable in silhouette, but once you start to try and get poses that require foreshortening and specific camera angles, things break down pretty quickly.
Again, at least as I’ve seen.
@@WhatDoesEvilMean check out this great tutorial on the subject and let me know what you think: linusekenstam.substack.com/p/tutorial-how-to-create-consistent
Great tips thanks I’ll def use!
That’s great tutorial thank you for sharing, just I wondered if you could apply that for a building ?
Hi Adam. You are awesome and I wanted to thank you for creating these Midjourney tutorials.
As I'm sure you know, Jhonny Jet's greatest superpower is his ability to change his tattoos in every image.
Do you know of any way to make that level of detail consistent?
Sounds like the basis for another tutorial! haha
There are ways for sure.
Tattoos are SO fine.. it would almost me more efficient to photoshop them on after the fact if we need them 100% accurate.
But the recommendation I would have would be to create the tattoo in Midjourney (or upload your own image) then use that image as a SEED for the character prompt. Might look something like:
/imagine Jhonny Jet is a cool guy with a chest tattoo, --SEED 00000, wearing a leather jacket...
One of the biggest issues I have is camera angles. Specifically getting the character to have his back to camera looking at a wall.
this video changed my life...
What about a combination of the image seed optimizations with image references? This is how i keep my faces looking the same. Thanks for the tip!
Great technique 👍
Kinda looks like Ronnie from the band Falling In Reverse
😂😂😂
Amazing vid and just a quick one: what if you upload the original image you created with prompt one and add something new?
We have a whole new video coming out with exactly that in mind!
Lots of good ways to do it but I would use that image as a seed and then heavily weight it against the new prompt you’re adding. It would be even more powerful if you took that 1st image into photoshop, cut the subject out, then reuploaded it as the seed image. Good luck!
Excellent review, Adam. I'm still concerned that there's significant variability even when it's trained so many times. Have you tried ControlNet with Stable Diffusion? From what I've seen, you can train that model for greater consistency, although with much more complexity of setup and running.
maybe you should mentioned where you got this idea with matrix and marvel comic from...this is the exact copy of the video from taokenized AI...
What? Oh damn lol I didn’t even realize I used the same example art styles. TBH I’ve never made a tutorial before. I use these methods and just watched like 10 tutorials in 30 mins leading up to freestyling this one.
I know the episode you mean. It’s made by Tokenize Ai (link below). Where he also uses Matrix & Marvel as art styles in his consistent character tutorial. I also love The Matrix and Marvel so just used em off the dome without realizing.
EVERYONE should watch and get inspired by Tokenize Ai. Some of the best MidJourney content on RUclips.
If you’re reading this and don’t follow him, head over to youtube.com/@TokenizedAI and give him a follow!
A.i is amazing i love this its like a cool game
I'm using MJ 5 I'm not seeing the seed after using the envelope.
Thanks for this informative tutorial, outstanding. I have a quick question if I may. So what if you have created a character design but did not give it a specific name other than the description of the image. How do you keep the image and still be able to upgrade it as you showed us? I am using the seed number but each new image looks nothing like the original? Thanks
I think giving the name is kind of the "secret sauce" to this method. Its basically saying "yes, this is what a Jhonny Jet looks like" again and again and again and again. Presumably this could work if you just ran the EXACT same prompt again and again and again. TBH I haven't experimented with that concept.
Interesting tutorial video. It does keep character consistency. I also just watched the storyboarding video you have. My complaint is with midjourney, not your video. You could paste your seed in with every prompt, but even then it doesn't seem adequate to actually storyboard a film. Or you would have to keep your prompts to something like "Pencil silhouette drawing of a medium shot." The realistic renders of people aren't desirable for a storyboard. So all up and down the software doesn't quite deliver.
Couple questions:
1 - Is Jhonny Jet now defined for other users, or just you?
2 - Does Midjourney now remember Jhonny Jet for you across sessions?
Hard to discover this without having a second Midjourney account... why don't you give it a try and let me know!
Neither, Midjourney don't learn with each interaction in real time. He is simply optimizing a seed that will generate the best Johnny Jet.
Midjourney only learns when they train a new version, which they are doing every few months. This works by improving the current model, v4 now for example, and retraining with the images with best rating by the users or with top visits in the gallery. In the case of this video nothing is saved on the server, so Midjourney don't have any idea of who is Johnny Jet, it's purely the seed that matters.
@@rafaelfigueroa2479 Noted -- thanks for the explanation 👍
Thank you.
Thanks for the info! Is this method superseded by --cref?
ermegerd this is gold!
Great video..any advice on combining three consistent characters in one picture?
i have a question, can you create multiple consistent characters? for use in different scenes? like multiple people in a single scene?
Hey there 👋 - yeah it's totally possible, though it's one of the most challenging topics with Midjourney right now. My favourite tutorial maker, Tokenize Ai, JUST released a great tutorial covering this topic I would recommend you watch here, hope it helps! - ruclips.net/video/znulbnRynOc/видео.html
At least give a credit to the youtuber who inspires you to create this video. I've seen that you modified some of it but the concept is still the same so at least give some credit. It does not decrease your credibility as a content creator because your quality of explanation is great. Thank you.
What? Oh damn lol I didn’t even realize I used the same example art styles. TBH I’ve never made a tutorial before. I use these methods and just watched like 10 tutorials in 30 mins leading up to freestyling this one.
I know the episode you mean. It’s made by Tokenize Ai (link below). Where he also uses Matrix & Marvel as art styles in his consistent character tutorial. I also love The Matrix and Marvel so just used em off the dome without realizing.
EVERYONE should watch and get inspired by Tokenize Ai. Some of the best MidJourney content on RUclips.
If you’re reading this and don’t follow him, head over to youtube.com/@TokenizedAI and give him a follow!
what if you want to change his clothes?
I'm using discord on a mac, I'm not getting the extra seed info when I hit the envelope is there some application level setting that I need to make this info show up?
Love this tutorial. I've easily done this 20+ times and midjourney is still giving me a lot of variability. Maybe my character is too complex? It's only a one sentence descriptor like yours. Do you have any thoughts?
Glad you enjoyed the tutorial!
Of course, results can vary depending on the prompt, and sometimes we either need to specify or simplify even more, depending on the results. Rerolling also another way to see what you may be missing.
@@polyformstudio Thanks for the reply! I think I'm going to try simplifying the descriptor. My characters also has a gas mask and that shoots out tons of variability.
Let us know how it goes and stay tuned! More videos coming soon 😊
@@veronikaabrams Maybe try defining a certain type of gas mask. Sometimes honing in on an EXACT version of a variable item will reduce variability.
I like that doll on the table behind the water class, what is its name please.
i did use the envelope on the picture but for some reason it doesn't work for me.
That's a bug. Report it. 🪲
using version 5.1, it keeps generating the same seed # every time with me, you kept getting a new seed number. How? I missed something, please help!
So now that I can't get a upscale seed in v5, am I forced to use v4?
Can you combine - - seed xxxx and - -ar 16:9?
cool bro but midjourney is not train at all by repeating like and prompts.
that envelope thing is not working for me is it maybe bcs im on my own server with bot as a friend or?
Make sure you're reacting to that specific post with the generic envelope, should still work while using the bot
@@polyformstudio yeah i did man and i dbl checked your video and did everything exactly the same, but when i click envelope it just puts envelope whitout giveing me seed,
Great tutorial, although having followed all your steps, Midjourney still couldn't produce a consistent character for me. It's so frustrating, paying a monthly subscription for Midjourney to waste my time. I cannot understand how it has the ability to produce some real outstanding images yet it can't do something simple like reproducing an identical image. I think I'll revert back to paper and pen
does naming the character really make any difference? I mean, you can't just type "Jimmy playing baseball" after training it, can you?
You cannot train the model like that, that's not possible, sorry to pop that bubble. Seeds are simply saying to the model to use a "fixed" set of point in the latent space, a way to initialize the model's internal state. If you are using same prompt + same seed number, your are just burning fast hours.
- "Same characters" is not possible, it will always be a "similar" one.
- Names matter, but not in a general sense. Try the names that generates similar faces. You can stick to adjectives as well.
- To get similar ones, prefer version 4, use same seed number, do not alter the prompt too fast.
- For version 5, try using one or two face portraits with seed number.
Looks pretty far from consistent. Lots of things change around the guy. But I guess it works for some cases.
Jhonny doesn't look as good as you do :)
It's quite impressive