Stable Diffusion Explained (BRAND NEW Art Generator)
Вставка
- Опубліковано 8 сер 2022
- In this video I give you a full tour for generating images in Stable Diffusion.
Check out my Website:
glibatree.com
Glibatree Discord (Glibacord) - Join the Community!
/ discord
Twitter:
/ glibatree
Instagram:
/ glibatree
TikTok:
/ glibatree
Stable Diffusion Website: stability.ai/beta-signup-form
Artist Resource: gainful-zone-eb1.notion.site/...
#StableDiffusion #Dalle2 - Ігри
Thanks for the detailed review and showing the different parameters. I am excited for this to be released, especially since you said it is easier than Midjourney.
Ive used Dall-e 2 for a little while now and i honestly like stable diffusion more because its so much more detailed and doesn't have nearly as many restrictions, and it's also free!
I agree with that a lot, though minus the Dalle experience 😂
Wait, why do I have a membership for it then? Lol am I on the wrong site
@@ATLAS-cy4xi its not free anymore
@@treudden ah
@@treudden there is a free version, you just get put into a queue. I generated a relatively simple image and it took around 4 minutes to create 4 pictures. The beta is paid with a membership
Very informative video, and very helpful, especially the part where you spoke about CGF_Scales. I didn't know very much about them initially, but after watching this I'm definitely going to be playing around with them more. I think my favourite part about Stability is it's open-source-ness and the lack of censorship. Another big thing is the cost, which is free and compared to DALL-E 2 it's a very strong competitor, especially for it being free, open-source software
I think the average person can’t draw as good as that, same as intermediate and amateur, but professionals could probably make something better , so the is like really good but in the middle of artistic potential . But it still is amazing lol
@Vortex Hash dang ai really gonna take over the world , thanks for the late night reminder that we are all fucked lol.
@Vortex Hash If you want to improve your AI images I would advice to learn basic art fundamentals (color theory, compoisition, lighting etc), these are timeless and having understanding of them will definitely show in your generated results
@@AtomicSlugg damn my comment started a whole argument , well if your not apart of the glibacord yet here ya go discord.gg/YVG3UNaQ
@@AtomicSlugg Most artists suck at professional level? How so?
It closes the gap? More or less. Seeing art made by hobbyists with it and seeing artists inputting their artwork into img2img - I'd prefer the output of the professional artist.
I cannot wait for openai to just call it quits and opensource their products too.
The dream!
So much for the name, it sucks that they initially were planned to be a non profit and decided to be a fake "totally not a for-profit" for-profit company. Real embarrassment of a leading company in AI in terms of their policies, I will say.
X years ago "hey get a job doing something artistic cause AI can't take those jobs"
Current year: "oh yea no they got us now"
Yeah, we really thought art was safe, turns out... No.
This suggests that creativity is very basis of intelligence and not the peak as used to perceive it.
eventually you still need a creative brain with some udnerstanding of art fundementals to create the best results with AI images, but yeah a lot of comercial illustration work will definitely be replaced since it doesn't really need any originallity to make it work.
Yeah it's interesting to realize that AI predictions were completely wrong with more creative fields having the most advances versus irl physical jobs. I still don't think it's anywhere close to job taking level though, it's too inconsistent and actually impossible to ask the ai to make slight changes but keep everything else intact. So it still lies as a neat inspiration tool and a way for people without the skill to draw what they want, still be able to get a beautiful image of the idea they had in their head
Always thought it was weird that most people never expected this. art fundamentals really just a simple set of "rules" so there's no reason an AI can't do it if it's fed enough information.
Thankyou for the thoughtful and helpful content! I am still on the waiting list for Stable Diffusion 2.0, but it is interesting to learn about the available parameters and see such compelling output. I was fortunate to recently gain access to DALL-E 2, and have had a great time making videos exploring what is possible with OpenAI's system, but I see different "personality" in each of the AI text-to-image synthesis systems I have encountered and each offer unique flavor to appreciate. I mean, there are some prompts today for which Craiyon (DALL-E mini) produces superior (if lower res) output than DALL-E 2 or MidJourney, from a subject matter perspective. Anyhoo, thanks again, excited to be exploring this crazy emerging creative and tech space that will surely shape all our lives in ways we can not yet even imagine!
CFG scale is the guidance of the model, basically, if the number is high, the ai will try to match your prompt more than if the number is low (it's what the devs explained during the launch presentation). Something interesting to note is that you can input negatives CFGs, for example -7 should output what the AI think is the opposite of your prompt. But as you saw in your experiment, the concept of matching more of less the prompt is pretty abstract, so the best thing to do is just to play around with the value and choose the one that gives the best results.
Thanks for that, it's useful to understand the intention behind the feature even if it's hard to understand what to do with it artistically.
cfg_scale is basically an amalgam of various parameters you may or may not be familiar with. In the most basic term it is steps and stylize, but that is if you are coming from midjourney level of knowledge.
From a diffusion model parameter setting it is closer to an amalgam of clip guidance, tv_scale, range_scale, sat_scale, eta, clamp, and probably cut_ic_power along with steps.
The coolest thing about all of these are the speed at which these perform. There is absolutely no reason not to use highest quality settings though for these with the default being lower to save costs.
Artbreeder has had something similar to this with all their influence sliders able to go below or above the standard ranges if you enter it in manually. Very cool feature to experiment with.
HEY !!!THK you . I was lost a bit that help me a lot :) am 2 weeks on and i love it
It’s interesting to see that the CGF scale seems to have a similar effect to Clip Guidance scale in Disco Diffusion. The stronger it is, the more the software “pushes” your image toward the goal. However, the effects of the high CGS are surprisingly different and seem to cause the duplication effect that a high ic_pow setting would have in Disco.
this this this.
@lastplusfirst good job on having a pretty good grasp of this. I actually just commented something similar above. :: kudos ::
I do art for a corporation and I can tell you my job is still pretty safe. Corporate style guides are way too ridged for AI to duplicate and what I do AI cannot do, yet. Even when AI can do what I do, I'll be the one telling it what to do and modifying it as necessary. What this AI stuff will do is raise the expectation for what "good" art is. If you can actually paint, you'll become more valuable as less talented people are able to use AI to compensate for lack of skill or vision.
As long as the world doesn't come to an end in the next 5 years or so, I'll be looking for actual artists who can physically paint, not digital, to do work for me. If you're good and talented, you'll still be able to find work. AI has just raised the bar.
It's just a tool in the end, even with AI generated art you will tell if somebody is creative, skillful and knows their art fundamentals.,
Digital art can be just as hard as physical art minus the annoyances of dried ink, no layers, barrier of entry, art supply costs, etc.
It just depends on the style. "Alegria" / corporate art may be a joke, but real art exists in digital form.
Why do you value physical painters - I’m curious ? I’m torn to go back and pick up my oils In this coming era
If you think your job is safe, think again. AI is only just beginning. You'll probably be redundant within five years...
@you Yes but this has only just got going. Wait to see what happens in even the next year let alone five....
Thanks for the video Gibatree, even if I don't get on board AI, looking at the output gives me plenty of inspiration for old fashioned manual generation using GIMP and Blender.
As a visual artist I think eventually these kind of engines are going to be a great resource for us to have reference images for concepts we are imagining. Poses, particular types of landscapes, etc etc
Agree, at least to make general tests, but even at its best you will still need someone to photoshop the end result to tweak it, and that is once it actually learns decent human proportions.
Yes, very much agreed. Like a search engine for inspiration but you're not stealing from anyone.
You talking to an AI bro, who probably does not have any kind of respect for art, and artists.
Thanks for the insights! Can't wait to be running this on my RTX 3090. The biggest feature I love about Dall-e 2 is inpainting, which lets me generate infinite clothing variations for a character(without changing the entire image, so you can take a photo, paint over someone's clothes and specify something new) for example or expand an existing scene. But stopped using it because $$$. So I've been playing with DD while I wait. Stable Diffusion looks like almost everything I've been hoping for. Though I imagine I'll be generating environments in Unreal Engine the same way in the future.
cfg scale controls how much of the existing image the model knows about as it draws in stages... especially for the smaller subjects, low CFG can mean that the model "forgets" what it has already drawn and causes repetition
That's really interesting! Thanks.
interesting way of explaining inner and outer cuts and ic_power. i like it!
damn, i thought it was another midjourney videos but this one is totally different
Great video thank you for all the effort you put into your videos >> very entertaining! 👌💯👀 I have been using Mid-Journey for about a week and Stable Diffusion for a couple of sessions >> got some very interesting results out of both and need to improve my prompting, so videos like this really help with some great pointers...
Where can I find the Artist tests you show here. Looks very useful.
Awesome overview, thanks a lot! Especially liked the experimental approach part.
The last screen seems to show a generation library site? Can you share the link to that?
Good catch, I didn't realize the link in the description was broken. Should be back there under artist reference.
All this image generation technology has got me so excited. I haven't been this excited since 2010 when I discovered minecraft! It will be another leap in human creativity!
hell yeah, being addicted to midjourney and DALL-E for the last month, it's so refreshing, fascinating and exciting. We're definitely entering an era of a new art period :)
Thanks for a great video! Tweeted it from the main acc :)
Thanks so much, I'm honored!
That art does look very good
Yeah it's getting really impressive
Personally out of the main three systems my preference is for Midjourney but I suspect everyone will have different requirements and likes
When do you think it’ll be released to the public? Also great vid I subbed :)
Welcome to the channel! I'm pretty sure very soon the model will be something you can download and run locally
I think c stands for how much the model should go in the direction that the prompt points it to. In other words it's guidance scale for classifier-free guided diffusion. I may be wrong though but your results kind of point for me towards the same direction. Because what's more dragon-y than a dragon? An abomination having key features of dragon multiple times (like 3 sets of wings).
If my hypothesis is correct then the more you increase the c, the less attention model puts into general image quality and the more it goes towards the direction pointed by the prompt.
So in conclusion, if you feel the model didn't follow your prompt exactly, try increasing the c parameter.
Stable diffusion will probably monetize with choose your own adventure style cartoons for kids. We are only a couple of years away from ai being able to generate whole saturday morning cartoons where kids can choose the story and it generates in realtime. Expect companies to micro transaction the crap out of it. I saw someone talking of examples of different options in Dora the Explorer where choosing different paths to take cost $.99 and then the ai generates a cooler story or additional content for choosing the diamond path instead of the dirt path for Dora to travel down. This is all very easy to do and each component is already available publicly. Just some company needs to come along a put it all together. Give it a year.
I mean AI Dungeon is incorporating Stable Diffusion for almost exactly this in the next few weeks (as I understand it)
@@glibatree interesting. definitely will have to check that out. I see you have some videos on it. I'll make sure to watch those.
I’m here , and btw Edward McGee you here?
As a 3d artist I would really like an option to map out the image coordinates in 3d space for export into Houdini etc
That would be amazing!
would also love seeing more ai tools come to 3d production
working on that, stay tuned ;)
Applied today. Wish me luck!
Good luck!
@@glibatree Thanks!
I'm guessing CFG is something about foreground. center fore ground. it emphasized the background more at low levels.
This would be great at generating stock images for still images for videos.
Very helpful,im actually in the server
Awesome! Glad you enjoyed it
Thanks
This ai seems to be way more versatile than midjourney can't wait to use it ! Is that a variation of disco diffusion or a complete new one ? And can we fine tune it as we like similarly as disco diffusion ?
As I understand it, it is by many of the same creators as Disco (and will likely replace it) but it is completely new training. It will absolutely be able to be fine-tuned, what you see here is the hot-off-the-press default model right after training. They've already released Source Code and I think the weights their using for the beta are going to be published soon (though if you're a researcher they will provide to you).
@@glibatree thks ! Currently from them all, disco diffusion is the mist advanced one in term of uniqueness. Just check the best artists on instragram it’s so much sharper and unique than midjourney for example. Because of all the parameters you can change. I ll test stable dif but is think they re gonna make it user friendly like midjourney so you won’t be able to create your own art with your own receip like in disco. Wait n see i guess
@@thibaudherbert3144 This is why I have always used Disco exclusively; I do not like just typing prompts in discord.
Are the controls easy to use, I was never good when it came to complicating controls like in Blender.
I have access to Dalle 2 and is pretty god, but i believe that will be more fun to play with stable diffusion..so i am still waiting.
Even just to be able to get 9 images per prompt again is amazing
Hard not to see the advent of this technology as heralding the end of human art.
I want you guys to step into the shoes of an artist listed as a prompt for this bot. Imagine a new show comes, and there's a hype train for one of the characters. Usually, that's a great opportunity for gaining followers and attention by making fan art. But, before you as the artist can sit down and put your pencil/stylus down, there's already hundreds of counterfeits of that character in the art style of your personal brand on the internet. Some of the fakes are already on merchandise labeled with your name. Some are shared on fake accounts that mix your real art with AI counterfeits. How would that make you feel? What would that do to your brand and your IP? To your prospects for staying an independent businessman?
The images are generated off of preexisting art, so its actually a derivative work already
Blown me away
Some recent changes: steps is limited from 150 to now 50, image count is limited from 9 to now 4, and there is a lot of blurring of false positive "sensitive" images. Very often innocuous.
Yeah, the problem with making videos about works in progress, very quickly becomes outdated.
I have Midjourney, Stable Diffusion, and Dall-e 2. Quite lucky I know. I mostly use them for artwork. I actually use and like Dalle2 the the least in terms of the quality and style it produces, but it does generate what you ask it to the cleanest. In terms of art quality though, compared to the other two, the art looks more like it was painted by an amateur - poor lines, style, strokes, fine details, etc. I'm also happy to pay for an AI, but Dall-e 2 is quite expensive. It takes some rapid iteration to come up with desired results on any of these and you can quickly burn through $15 on Dall-e 2. I do sometimes use Dalle-2 for edits, it's a great feature I hope we may get in the other systems soon.
That's really interesting, thanks for sharing! I'm super intrigued by in painting and would love to try it, hopefully they let me in soon haha
What is the website you were looking at towards the end?
It's an Artist study. So the names and pictures are how Stable Diffusion perceives the style of each person. Link in the description should work, but also it's broken once so lmk if I need to fix.
@@glibatree seems to be broken again
I got fortunate to be accepted into dalle 2. I am sick of the censorship and poor faces. This ai presented in this video intrigues me 10X more
I’m seeing awesome works from MidJourney so it seems they are pretty equal.
Midjourney for the win
What is the recommended hardware needed to run this ai ?
While it's on MidJourney, and when they move it online: anyone can run it on any hardware. To run it locally, it's a 4 gig file, that runs on 10 gigs of VRAM. They're also working on a release that runs on half of that: 2-2.5 gig of space, 5-6 gigs of VRAM. You also need Windows, since apple hasn't updated the pytorch.
@@glibatree Thanks for the explaination, when I have read on their website that they used 10GB VRAM while my gpu only have 8GB I was worried a bit ^^°
@@Mnemesia Yeah most machines are similar, and they want to release something most people can run so hold tight haha
I hate how some of my photos get blurred :(
Well know what I’m doing now
Exciting times with AI
You mentioned you liked the price, but looking in the transcript I could not find the price. (?)
At the time it was completely free, for everyone in the beta. Now it's open source, so free if you run it yourself. On Dream Studio, (on Stability's GPUs) it's 10€ for 1000 512x512 generations (50 steps)
I need to make an updated video now that it's really out, but v1.5 is releasing too and I want to cover as much of the news as possible
Still looks like something created in warhammer 40k warp.
Unable to join your Discord channel
Can it do inpainting?
Not yet 🤔 but I think it should eventually be able to
I never liked AI art since it always gave me an uncanny valley feeling, even Dall-e. however, I tried stable diffusion with Nightcafe and I genuinely like it. like it doesn't give me the uncanny valley feeling, the images look really good.
Also, if you use artist inputs who themselves have good faces, you'll tend to get much better faces out the other end... Likewise, if you specify things like Unreal Engine, CryEngine, Octane Render, Blender Render, etc. Some of those seem to have much better "face quality" when using those parameters. It seems to all be about the Artist Seeds, the prompts, and the modifiers & effects. Once you get the hang of some of the modifiers, it gets a lot better and easier.
AI generated art is shit. It literaly steels artists art, and there are cases where the artist watermark was even there.
How to Sign-up with this AI Tool, could anyone Help me Please
So, not "exactly how it generates art" but "exactly how you can generate art with it." That's two different things.
You not the first person to mention something along those line, maybe it's confusing enough I should change the title. The idea behind the title was "Exactly How to Use Stable Diffusion to Generate Art" but that felt wordy, and I wanted to clean it up. I'll try to think of something better that doesn't make some people think I meant "The Data Science Analysis of the Machine Learning Behind Stable Diffusion". lol
@@glibatree Cool. Unfortunately, as a computer scientist, that was kind of what I was looking for. I enjoyed the video anyway, though. 🙂
Royal Skies LLC (youtube channel) just announced he's going to do a whole series on how to get good prompts to make good art and stuff. He's a great educator who does two-minute clips on very specific functionalities in 3D art. If you or anyone else wanted to follow along.
Hmm, I was expecting to see "Exactly How Stable Diffusion Generates Art" not how parameter changes affect the output. I thought I would get an explanation of how the software learns and creates.
I can see how that might be confusing.. i really meant it as in how it actually generates it. So more like what you have to do to get art out of. I'm a fan- rather than a scientist haha.
Hey^^
Hey!
Midjourney is better for digital art but that’s not really my passion. DallE2 is doing more interesting work, like my oil work. I’ll have to try Stable Diffusion this sounds good.
Can you add the Dream bot to your own server?
Not unless you host it yourself, this video was before Stable Diffusion went open source
@@glibatree OK, ta. I've not found a video on how to host it yourself but I don't have a new PC anyway. :-)
I got excited for a min but not another discord ai.... ugh
rip image cache...
Lmao, fair. Webpage soon though 👀
-C5.0 Dragon :D
The images look so flat and awful, though. I can see how it has some extra coherence over Midjourney in some cases, but the compositions are boring, the colors uninspiring, values flat and not always cohesive. In that sense midjourney is leaps ahead.
perhaps better and more creative text prompting can resolve this
It will feature variations and inpainting though, which can help "fix" other AI's lack of coherence.
Absolutely, used in combination with other AI it's great
the examples he showed in the video werent good it can do much better
You say just put !dream bla bla bla but where?! Where do I put it?
Type it in the chat box
This makes me ask a question. Why make art?! If can just click generate??
This could potentialy lead to bad things...
Yes. Finally someone with rispect for art, and artists in the coment section.
Euler - "oiler" (German word)
To my eye, midjourney is leaps and bounds ahead of this, at least in terms of artistry. These images seem a little flat and cartoony. They don't elicit any emotion. But maybe it's just me and I'm completely full of shit.
Yeah, I think it's possible to get passed this with the right prompting. But the speed and fidelity alone is very impressive.
is that Felicia Day?
Yes, "Felicia Day as a Cyberpunk Assassin" haha
ai generated art is improving so fast!
thats why its scary sometimes
Absolutely agree!
now you can just install it on your PC no need for discord. its way better that way.
If you are ready to pay yourself (as i have exhausted my credits on dalle 2), you can take my ID, and change the password as you please.
What about copyright of all the works of the artists who made thefabulous illustration you use at reference, without even consider naming them.
They fockin closed access!
Yeah great if you like comics.
wow dragon body and a car that is all there , midjourney can't do that , try to do words , midjourney is about to get upgrades, you said this is easier than midjourney not sure what you meant you have to put in all info yourself ? ... but you have more freedom of the output that's great.
Oh I just meant that there are less parameters to learn. And actually now that there is a website, and it is open source it is actually much easier to start using.
still prefer midjouney, nothing produces works of at 99.99% of the time.
That's super fair! I think MidJourney has kinda it's own style and is much more artistic in general. I think they're being competition is great for AI art in general though
realized 10 minutes in that you don't have like 100k subscribers
Oh that's really nice of you to say. Almost to 1k though, I'm pumped that the channel is so doing well rn!
Dalle2 is objectively better. Has an actual platform, a bunch of tools, the output is more objective and it understands what you want way better. Stable Diffusion is just free, and free is not the same as better.
those aren't artowrks. they are just AI stitching images.
Have tried dozens of similar tools. The developers said that it's at the same level with Midjourney and Dall-E 2. Doesn't even come close.
I still think most if not all image generators are cheating (using existing images as is) more than they want people to know but stable diffusion is great. It has too many variables maybe even (different resolution creates a different image for instance) but this leaves lots of room for tinkering.
I know what you mean, it often feels like MidJourney is to painting what Stable Diffusion is to photo bashing.
Mid Journey is better
Ummm.... A Wyvern is a Wyvern not a Dragon, which is why your "wyvern dragon" picture is a Dragon not a Wyvern.
You may as well ask it to generate a 2 legged lizard with 4 legs.
Here is my problem with this YOUUUUUUU didn't make anything. this computer replaces lack of creativity or actual skill with machine learning. So the computer and AI is the artist the person punching the key words in is just that. Once this catches on it will kill art as a career. I mean I imagine opening a buisness making tshirts. why would I pay a artist when I can just punch some words in to a computer and get what I want for free. and if you think EA activision coca cola Warner are not chomping at the bit to replace thier whole art departments with a computer were an exec can just punch in" Hey AI make me an image that will make people want to give me millions of dollars" then you have been paying attention to virtually every other industry that is replacing its works with machines.
artists are fkd up
All the AI programs are bad.
AI art is the downfall of humanity.
Dalle-2 was not impressive at all. Totally overhyped....MJ is much better at this point.