The Most Important Thing to Know About Seeds in Midjourney V5
Vložit
- čas přidán 30. 06. 2024
- In this video, I'll show you the most misunderstood thing about seeds in Midjourney. I've seen a lot of bad advice out there on how to achieve consistency across your generations ... is it even possible? You tell me what I'm missing!
FREE PDFs
cameralens.futuretechpilot.com/
200styles.futuretechpilot.com/
cheatsheet.futuretechpilot.com/
Links:
Merch - FutureTechPilot.threadless.com
Discord - / discord
Creative Fabrica - www.creativefabrica.com/spark...
My MJ Gallery - www.midjourney.com/app/users/...
to have a better consistency...
1) generate the seed from your 4 pictures result (even if you like just one of them).. don't ask for the seed from your upscaled (preferred) version.
2) now you have the seed of your 4 pictures... upscale the one you like, and copy its image link
3) do a new "/imagine", paste your upscaled image link along with your original prompt (from poin 1) add some attribure you like to change and finish it with the seed you requested on the point 1).
Yeah there are a few tricks like that for sure! Thanks for taking the time to write it out
@@FutureTechPilot Also, I've seen the "Remix Mode" maybe is what we're looking for
@@EmanueleDiBacco naaaah. Remix mode sounds great in theory but it definitely doesn't work that well
it’s a big “thanks” from me🎉😊
@@FutureTechPilotif you know a few tricks like that….how about telling us?🙂
I've never thought seeds were for image consistency, but rather to reduce variables so that the difference in subsequent images will be entirely due to the changes you make in the prompt. This gives a better, more accurate understanding of what the prompt is doing. The subject of the photo may not be exactly the same, but it won't be nearly as different as if you had a random seed each time. At least that's how I understand it.
That's a really interesting way of explaining it! I guess I've seen a lot more advice suggesting that it was possible to be consistent with seeds and I made this video to try and clear that up. But I should make another video showing off what you just said. It's a good distinction and would help take some people to the next level if they practiced it. So thank you for the idea!
From what I’ve seen on Tokenized AI channel, if you upscale an image you like and then heart react to it, it kind of trains MJ on what you’re going for. Then you prompt with that same image you just upscaled, with the same description as before and the seed number. Repeat this process several times and it seems to help the characters be more and more consistent
It's not true though. It's a myth. I told him that once and he said he didn't care if it wasn't true lol
Midjourney has no memory. If you react to an image, it tells the system that it made a good image but it has no idea what you asked for and does not remember what you want in the future.
That MIGHT become a feature soon. But it is not how it works right now.
@@FutureTechPilot Interestingly. I kind of like the inconsistency and hope they don't change it. In fact, I'll ask them not to change it!
@@Vherie I know, but judging by the comments, it seems to be getting better (have been focused on other projects, so don't know). What's you assessment? Is it learning?
I really love the way you explain things in ur video. Easy to understand and very clear presentation with samples , etc. Keep up the good work . Learning a lot from you. Thank you..
Hey I'm glad my style works for you. Cheers buddy!
Thanks for the heads up. This is exactly the issue I've been having lately,, your timing couldn't be better.
hahah I'm so glad to hear that. Thanks for the comment!
wow! I've been using seed all wrong!. I thought it just didn't work because I was upscaling and using the seed number for the 1 upscaled image but when i used the seed from the grid i was able to successfully get variations of the model that I wanted, finally! :) Thanks!
haha that's great! I'm glad I could help. Thanks for the comment :)
Thanks for this, there's a lot of misinformation about seeds being a golden ticket that just isn't accurate or helpful, and people waste a lot of time rolling images with seeds that don't do what they think it does. Seeds basically just determine the starting "noise pattern" that Midjourney uses- when you make a grid of images, technically the seed is only for the top left image, and the rest are determined algorithmically based on the first seed. If I remember correctly, and I might be wrong, I think sometimes the seeds can switch placement with different prompts or something, but it's been a long time since I read an explanation on it by someone who knows, which is why sometimes in grids with the same seed and similar prompts, you get similar but relocated results. However, significant changes to the prompt are much more powerful than the seed is, so while it might hold some compositional coherency in similar prompts, it will be barely represented as prompts get more different.
In V4/V5, rerolling a job with a defined seed will actually give it a different seed since V4/V5 are deterministic so you would get literally the exact same image. A better way to get slight tweaks to an existing image is to add a very low weight junk multiprompt (i.e. ".::0.1") which will make things slightly different but not change the original image significantly for both V4 and V5, or to upscale again in V4, which will use a different seed for the upscaler, leading to potentially better or at least different results.
As for why the seed from the V4 upscale makes such different results than the upscaled image, using the envelope react to V4 upscales will actually tell you the seed for the upscale, not the image in the grid (which is again determined by the seed of the top left image in the grid). It's also why V5 upscales have no seeds, since they aren't actually upscales, just split off the the original grid.
haha there are so many layers to MJ, I feel like bad advice is going to always be around. It bums me out when I see it so I hope to pass along the truth in my videos. Thanks for the comment and your thoughts!
What's a noise pattern?
Thank you! Going to try the low-weight multiprompt. And off I go for a new adventure 🚢
@@LieOfOmission just the layout of the static noise each image starts off as
The painful truth, but I appreciate that honesty.
Cheers buddy
dude i just searched for this, thank you so much!!
haha funny how that works out. Cheers!
If you use images to create your character (I especially love to create characters based off mixed images of people), then you have to include the links to those images *along* with the seed number in order to have the consistent characters. The seed number does keep a consistent face, but only if you have the original links included. Without the links it'll create someone else entirely.
Good advice overall though. Very much enjoying putting my consistent character in different situations.
Do you have any advice on how to keep consistent drawing style of the characters clothes/hair?
@@FutureTechPilot So I can answer this in several ways but not straightforward (since I keep experimenting). The astonishing aspect about Midjourney, aside from its ability to create stunning work, is the combination of photos to create new people, new characters that don't exist.
So the ability to combine images to 'blend' a new human keeps the character intact from art style to art style (depending on image weights). But art styles always change based off prompts - because you can mix and match and blend art styles as well.
So to answer your question I still haven't honed it down to an exact science. But I use images to keep a character consistent between styles.
And the ability to use images to combine 1 or 2 or 3 people (or more) into another person with all their facial/hair features perfectly blended - as a photographer that absolutely blows my mind and I can't stop experimenting with it.
And again, consistency of a character between art styles is also kept (to a certain extent) if you use specifically the same images but alter art styles via prompts.
@@cagnazzo82 Hello Legendary, I'm also researching about it. Do you want to talk?
Thanks I found this very useful
Thanks for the comment!
I believe if you copy the link of the upscaled image, and then add it after imagine, with the same prompt following with minor tweaks, plus the seed of the grid, it creates pretty consistent looking characters. i use it for my channel
Thanks for sharing your process! I'll try that out and make a video about it soon
@@FutureTechPilot yea i have a bunch if you want to see proof. i mainly create consistent characters with midjourney. and no problem i like your videos
@@agelessone email me some pics if you'd like futuretechpilot@gmail.com
Have you try to use the variation feature? For Example ones you upscale an image you like, you can then make variations from that image, once you do that, a new grid will show up in this case (an instagram model) then you take the seed of that variation grid and from that make difference images of the same person (in different perspective, outfits or face expressions)
Hope that helps.
I just tried that and it didn't work for me! Kept giving me new people even if I used the seed from the variation grid
I too have seen this lack of consistency with seed. So what is the answer for creating consistent characters when making a comic book or graphic novel? Using an image prompt seems to help somewhat, but do you have any tips for greater consistency? Or am I asking MJ to do something it's not capable of right now? Also, any tips on different cartoon styles? Thanks for these helpful videos!
My personal opinion is that consistency is not possible right now. There may be some people who disagree with me, but I think the inconsistencies between image generations are obvious and easy to spot ... Midjourney might have a feature one day where you're able to dial in on a certain style, but I don't think there is any technique to tell a story made up of the same looking pictures.
As far as different cartoon styles go, I haven't experimented much! I stick to graphic novel aesthetic mostly but let me know if you come across any styles that really pop!
Do you know if you are able to upload your own photo to midjourney(like of your dog) and then make it into the style of an image you generated? Like cartoon style with paint splashes?
No unfortunately, not at the moment. But you can get kinda close - czcams.com/video/4FQgkyMGdVY/video.html - check out that video
The initial image styles you were able to recreate by the seeds in previous versions were more consistant than in v5? V5 has much more variables? It's just in previous versions, especially before v4, I'm finding with certain prompts I was able to get more specifc styles and have a good idea of what the outcome would be, but in v5 I'm finding that even slight changes to a prompt could dramatically change the initial images, making it harder to recreate the style I am looking for. Any advice?
Honestly, what you said is a really great summary of what V5 is like ... but V5 is going to get an 'easy' mode soon where it has more of a style in the same way V4 had a style .. once that happens, I think prompting will return to how it was
thank you so freakin much!
And thank you for the comment!
That is interesting. I didnt think about the implications of v5 rendering directly to high resolution before. That probably is also the reason that --video is gone.
Yah :( I thought --video was the coolest thing ever lol I really hope it comes back one day
Used to work in --v4, not in 5. You have to think as a prompt is "100% /4" you will understand in a week time.
That's a good way of explaining it!
Maybe you should try without Stylize at 1000. --s makes Midjourney prioritize styling over other things like consistency. I'm not sure if it will work, but I'm going to try it too. Good luck.
It's not a bad suggestion! Let me know what you find
Thanks for the video. My seed number doesn't show up when I'm using v5. Only if images have been created in v4. Any help is appreciated. Am I missing something? Thanks
You won't get the seed for a V5 upscale. Make sure you try it on the original grid!
Hi Nolane, Thank you for your content. I did everything but I can see the seed after that I put the enveloppe, I dont know why? than you.
haha they updated it after I made this video. Now you can see the seeds!
In a way this inconsistency is very human... Thanks for vid. I thought it was me, throwing in the wrong prompts.
haha Midjourney is very human in some ways for sure
Great teachers help you cut through bullshit. Great video. Give it a few months. Consistency is coming
you can see job id and a seed number in MJ5 in mobile version
Yeah they fixed it actually! Works on desktop now too
Thanks!!!!!!!
My pleasure :)
Thanks for claryfying, im not crazy
haha cheers man!
Model consistency probably won't be a true thing until V8 at the end of the year.
Patience is our virtue
im on MJ5 and do have my seed, but its not reading the command --seed 12345, I even tried --seed:12345, but all in all, its giving an error for the word "seed", thoguhts?
super strange ... --seed: definitely won't work so make sure you're writing (dash)(dash)(seed)(space)(####)
--seed #### should work fine.
Maybe try writing it at the end of your prompt too
interesting thank you
Happy to help!
This strange because I was trying to do an old witch and it kept giving me the same basic face. I had to work hard to get away from that face. Now I want the same Frankenstein face and have spent all day trying to do it.
consistency is the hardest thing to do!
Huh! Thanks for this. 👍
I read 'sameseed' in some places. So there's a difference between these or what?
And 'Job ID'?
This is all new. Thank you. Seeds always annoyed me because it just didn't help, perhaps once in a very blue moon. Also not a big fan of 'remix'. It has _always_ disfigured something, or lost quality or any number of things, even if I make a slight difference.
Agreed, the Job ID should do the trick...
'Sameseed' is amazing but hasn't worked since Version 3 lol
And Job ID is just a way to recall any picture as if it was a brand new generation. Like if you have the ID of an image from 5 months ago, you can use the ID and load the picture into your discord convo instead of needing to scroll back to find it
@@FutureTechPilot Thanks! Shows what I know! lol
@@AG_before 😂
Then how in other videos (by other youtubers) they use the seed phrase, and we actually *SEE* that the character stays consistent!?
There are two answers to this ...
1 is that they are using image prompts + (sometimes) multiprompts + simple character descriptions ..
and 2 - in my opinion, they are not achieving consistent characters. At least, I can spot the differences easily and I think it's bad advice to tell people they can get consistent characters when it's not really true. It can get close, but there are 4 billion seeds to it comes down to a lot of luck and persistence. So if it can't be done perfect, then I don't consider it a feature of Midjourney at the moment
Do you have a group on Facebook?
No I don't but that should probably be somewhere on my priority list lol
Putting less chaos, does that make it a consistent seed?
That's an interesting thought but the default chaos value is 0
@@FutureTechPilot thanks for the response .. okay so to get consistency of sorts one should fix a seed number and their 'variations' only would give us similar results?!? Cheers
@@ai.ai.captain My personal opinion is that consistency isn't realistically possible and it's not worth it to try. However, what you said kind of makes sense. You don't need to keep the same seed number, but yes when you hit variations, that will give you alternative versions of your image. However, there is no way to take a certain person and transport them into a new scene
@@FutureTechPilot very helpful my friend. Looking forward to your new videos ❤️
Seeds are consistent regardless of the value, but not in the way people think. Diffusion models work by training on images reduced to noise, and then using the observations learned by watching that process to reconstruct images out of noise. Seeds are the starting noise that the diffusion model uses when it starts to process the prompt, so seed 1 is always going to be seed 1 (unless something changes with how the random noise generator works, which might occasionally happen, might never happen, I don't really know that much).
However, the seed is relatively weak in the amount of work that the diffusion model makes, it's more like giving it a skeleton to build on, and so most of the time the seed is almost entirely unrecognizable by the time the prompt is done. There are many steps (I don't know the exact number) and each step the model reduces the noise away towards an image, and each step is a place for the model to make a different choice- and there are probably a hundred or so of those (maybe more, maybe less).
Similar prompts guide the noise in similar directions, which is why minor tweaks to a prompt end up with similar images, but significant tweaks end up very differently, because each decision along the way compounds. "cat --seed 1" is going to be very different from "dog --seed 1" even though the starting pattern is the same.
useful info, thanks! btw, ukiyo-e is better pronounced this way: oo-kee-YOH-ay. your japanese viewers will thank you
haha I appreciate the tip!!
it does not show a seed number
I thought the only way for consistency was - reference picture + prompt + seed number?
I think a reference picture helps with some details and similarities, but I don't think it's possible to get 100% consistent subjects in different pictures
But how to get SEED in MidJourney V5???
You can't for upscales. Only the original grid.
👋
🤝
No me des problemas, dame soluciones :p
😂 I'll try
Wrong.
Teach me