FREE Stable Diffusion 2.1 Is The Biggest Disappointment Yet?

Key Takeaways

hello humans my name is Overlord and as you might know by now stable diffusion 2.1 was recently released and as you can hear by my voice I'm not exactly pleased with this newest release now I was really debating if I should make a video about it or not because well there isn't really a lot of things to say compared to the 2.0 that I haven't said before so in the end I decided to kind of make a in-between video kind of like a simpler video with less editing I suppose you can call this a ramp video in a way because I do have a lot of things to say about this newest release about the community about the future of stable diffusion now don't worry it will not be too negative it will just be my opinion after all but this will allow me to cover a variety of subjects that I want to talk about in one single video and also for the people who are here just to see how you can install the 2.1 on your own computer and what are the differences between the 2.1 and the 2.0 and the 1.5 don't worry I will also show you that and I will do it in the beginning of the video so you don't have to wait for the information all right this way I won't have in the comments some people telling me ah tlgw the actual content starts at 2 minutes and 45 seconds thank you for wasting my time okay Billy thank you so much for your service so yeah I suppose let's get right into the video so if you want to install the stable diffusion 2.1 on your own computer um there is basically two different models the 768 model and the 512 now personally I recommend that you download the 768 in my testing this is the one that is basically giving you the best results but still if you have some space on your disk you can basically download both of them and to do this you're gonna come here click on 1000 versions click on the non EMA prune.kpt file and I know it says here that that hooking face detected some pickle Imports and usually in that case when you see something like that if you remember my previous safety video you should not download any models that have some pickle Imports because they could contain viruses but this is a model released by stability AI I guess we can trust them I mean I'm really hoping that they're not putting viruses inside their own models because if they did we would be in deep trouble deep trouble yes so you're gonna click on the non-ema prune.kpt file and then click on download and you're gonna do the same for the 512 version so thousand versions non-ema pruned and then download then you're gonna click the two links in the description down below which will be a link to download the yaml files that you need to make this model work on your computer I've already renamed them for you you don't have to do anything so simply click here and click on download for the 512 and for the 768 so in the end you will have four files two ckbt files and two yaml files so you're gonna select these four files Ctrl X to cut them then go into your stable division web UI folder models stable diffusion and then you're gonna paste all the files right here then you're gonna go back in your stable diffusion web UI folder right click on web UI user.bat file click on modify or edit with notepad and here you need to add Dash Dash xformers and or dash dash no dash half because you need to run on this with food Precision now I really recommend trying out only with exformers and if it doesn't work for you you will then have to use the no half arguments otherwise you will just have black images and You're Gonna Save the file and launch stable diffusion now when it comes to the differences between the 2.1 and the 2.0 well as they said themselves uh the 2.1 is basically a very minor Improvement and I really want to emphasize on minor Improvement because there is really not that much difference between the 2.0 and the 2.1 it is a little bit better but really not by much so they basically said that they're not safe from work filters was way too aggressive so they basically decided to decrease it a little bit for this 2.1 version and this allowed to bring back some of the features that were present in the 1.5 such as art styles and celebrities that look a little bit more like themselves and also giving you the ability to create like this super for realistic architecture Landscapes images and one of the biggest things is the fact that now you can generate these super wide images with the 2.1 that you couldn't do before but unfortunately if you want to generate these kinds of images you really need a super super powerful GPU and I'm not even sure that you can generate this kind of super wide images with the automatic 1111 repository I mean technically you could already generate this kind of images using out painting before but apparently now you can generate these kinds of images in one go I mean that's cool you know but I don't really know a lot of people can generate these kinds of images with their local computer so now then is the 2.1 really better than the 2.0 um I guess a little bit not by much but my biggest problem with that as I said in my previous video is the fact that I personally still don't really find any good reason to use the 2.0 or the 2.1 compared to the 1.5 I still think that one 0.5 is way way better at creating images than the 2.0 versions and I know that a lot of people say that the 2.0 and the 2.1 are better at understanding prompts but from what I saw this really depends a lot of the kind of images that you want to generate so here's an example so here's a comparison between the 1.5 the 2.0 and the 2.1 where the person trying to generate a bunch of images of sculpture of marble sculpture wood sculpture metal sculpture with the 1.5 the 2.0 and the 2.1 you can see here for example the differences so for example here's the prompt highly detailed marble sculpture of a female Necromancer and you can see here this image in the 1.5 which is actually really really good and then the same prompt with the 2.0 and the 2.1 like I'm sorry yes the images look good personally I still find the 1.5 way closer to reality showing the face look actually really good here in the 2.0 and same with the 2.1 but when you actually look at the grain of the picture this picture is way more realistic than the 2.0 ones and the same thing here which is probably one of the biggest differences like a highly detailed wood sculpture of a female Necromancer this is the 1.5 where you can clearly see the grain of the wood with all the Ridges of the wood and this is the same prompt using the 2.0 and the 2.1 like like for me like this is not wood again you don't see any wood grain you don't see any details then this is a wood sculpture this is basically like a 3D rendering at this point it's almost like someone tried to 3D render this and forgot to apply the wood texture over it and you can basically see the exact same thing with metal sculpture which I'm still personally find that the 1.5 is doing way better than the 2.0 or the 2.1 now here's a comparison between the 2.0 and the 2 0.1 with different prompts cute brown puppy landscape painting etc etc and the differences between them is just like super minimal they're not better than just different pictures basically so it's really up to you to decide which one do you really prefer so for me it's not like oh yeah this 2.1 is so much better than the 2.0 because there's like so much more details no it's it's almost like this is a different seed made with the same model so for me the 2.1 is not really better it produces the same quality images as the 2.0 but they're just different basically but the best quality is still the same now one big question is does the 2.1 creates better hands and from what I saw the good news is that yes indeed apparently the 2.1 those produce better shaped hands compared to the previous versions but again the differences is very very minimal it really depends on the types of images that you're trying to generate and it really depends a lot on the seed again this is kind of like plain roulette this is a game of chance and if chance is always involved I really have trouble saying that yes the 2.1 is definitely way way way better than all the previous models the 2.1 is indeed better but not by much so as you can see here you have Emma Watson with her hands up in the air and if you look at her hands indeed you can see that most of them have only 5 Fingers which is pretty good but unfortunately again what I said in my previous video it still holds true when you look on the face of the celebrity this doesn't look like Emma Watson anymore and you can clearly see the big difference in quality when compared to the 1.5 I mean look at this this is Emma Watson and this is the same image this is done in 1.5 look she has two hands and five fingers but here she does look like Emma Watson and this is the same image with the same exact seed in the 2.1 which for me he looks way way worse than the 1.5 and you can even see like a beginning of some super long arm on the left here but this time the face just looks way way worse so again what I seen in my previous video still holds true today celebrities do not look like themselves anymore so also as I said previously since now we have a little bit more art styles back in in the 2.1 version now you can add more styles to your images but I mean seriously when you look at the final results here using a bunch of different artists I mean the the results are just not that great again I'm definitely sure that you can make way better images with better art styles and better details with the 1.5 like this is really nothing exceptional and to prove you that this is nothing exceptional here's two images made with the 2.1 they're using the exact same prompts this one is done with the 2.1 512 version and this one is done with 2.1 760n version and you'll be able to get this kind of results to be able to gain these kinds of art style using the 2.1 you need to use a bunch of negative prompts and I will leave a link here to the radio thread that show you the differences between art style using a combination of different negative prompts and using this prompt we have a combination of these specific negative problems you get these kinds of art images using the 2.1 and they look good they look okay but when you use this same exact prompt with same exact specific negative fonts with the 1.5 you get these kinds of images and I'm sorry but you know when you compare them side by side if I had to choose I would choose the 1.5 in a heartbeat I find them way way better way more stylized just way more aesthetically pleasing than anything that the 2.1 can give you so yes I'm still staying on my opinion than the 1.5 5 is definitely way way better at generating more beautiful images compared to the 2.1 and this is the proof of that now yes the 2.1 can produce some really cool images especially if you are into landscape or more like realistic Landscapes you can generate some really cool images with a lot of details that are really really super realistic and a lot of them really look like a real photo but personally I still think that you can produce similar images using the previous models so again I'm not super super impressed also if you don't want to install it on your own computer and you just want to try now the 2.1 for yourself super quickly you can actually use run pod and rent a GPU for literally a few cents if you come in here and select the template you have no issues between using the 1.5 and the 2.1 so you can simply click on 2.1 and then click continue to install stable diffusion 2.1 on your GPU and also here's another thing you can also use runboard to install invoke AI so I know that a lot of people have actually asked this question before in my previous invoke AI video if there was a way to install and use invoke AI on rompod and now you can and after trying out this is actually really really cool so you can simply select the invoke AI template on run pod and then use it directly on the website and all of that for a few cents which is really cool if your computer is not powerful enough to run it okay so now then what exactly do I have to say about this newest release now first of all I want to say something kind of like something about like the community in stable diffusion in general and that is the fact that we are acting a little too entitled you kind of need to understand that we get all of this for free okay guys like this is actually really really cool like this is something that did not exist a few months ago and now we get all of this new technology for absolutely free to try at home and pretty much anyone in the world with an internet connection even if he has a weak computer can actually use stable diffusion for free and generate some really really cool images so in a way I kind of feel bad for ranting or complaining about it when I know that this is an amazing technology that we received for absolutely free and this is just absolutely fantastic this is amazing and unfortunately I cannot help but feel a little disappointed anyway because now that we were kinda used to the fact that we get these amazing models already available like the 1.5 we are always expecting better and better models to come out in the future and for good reason I mean it makes sense that in the future we'll get better and better models than the previous ones which is the case with pretty much any other text to image AIS like mid journey and yet with stable diffusion this is not the case and yes as I explained in the previous video I know very well why they did that and I will leave a link in description down below for this video you if you ever watched it and I'm perfectly aware that in the near future we will get way way better results with the newest 2.0 models because if every few weeks we get a minor improvement over the previous one after a few months we might get something absolutely incredible that is way way better than what we have currently but I cannot help but feel that currently stable diffusion is way way behind next to image AIS like mid-journey and you see before the release of the 2.0 each time that you went on the stable diffusion radiant you will get some absolutely amazing images generated by the community in super amazing art styles that you've never seen before with lots of custom dreambooth models but then when the 2.0 released this was like a bomb that was dropped on the entire community and the community was split into two parts the people that loved the newest 2.0 release because the 2.0 was really good and generating very realistic images is images like this for example which look really really good but there were also another part of the community that felt like they were betrayed by stability AI a company that was always against censorship Pro Liberty Pro freedom of speech and in the end kept in to the demand to censor their future models and not only that but it also came at a cost because for a lot of people and me included they felt like the newest version was way way worse than the model that we got a few months ago which really did not make any sense and now the only images that you can generate with the newest 2.0 models are super realistic images like this or realistic landscape images but pretty much every image generated with the 2.0 or the 2.1 have exactly the same kind of feel a kind of realistic or ultra realistic images it is extremely difficult to create anything or in a completely different art style which was something that the 1.5 did very very well and you might think that this is okay and you can still generate pretty of cool images and it's true but all the text to image AIS can do the same thing and more now it really hurts me each time that I go on Mid journey and I see these kinds of images created by the community that are just way better than anything that you can generate with stable diffusion alone and I know that a lot of people are just super allergic to the fact that it is a paid subscription service and that they want everything to be free but at the end of the day poop is also free that doesn't mean that I want it on my plate and I would rather spend 10 or 30 dollars a month to be able to generate amazing images like this super quickly and super easily when stable diffusion is not able to even do like a quarter of that I also got a lot of pretty funny comments on my YouTube videos lately telling me that lately my thumbnails looked really really good and they were asking me how I did that in stable diffusion and the answer was very simple I did not use stable diffusion to make my thumbnails anymore so for those of you who want to know how I make my thumbnails or more like how I used to make my thumbnails is that before I would generate around 100 images in the stable diffusion and then choose two of them for the thumbnail and the reason I did that the reason why I generated so many images and I would spend so much time just generating a bunch of images is that around 70 to 80 percent of them would just be not good enough and yet now with mid-journey I can spend five minutes to generate amazing images that I can use right now directly for my thumbnail and they look way better than anything I could ever generate using stable diffusion and don't get me wrong I'm still a huge fan of stable diffusion I still think that stable diffusion is the most versatile text to image AI on the planet it is the only AI that you could train with your own images it is the only AI that you can use for text to image image to image in painting out painting create a new style train a new character Train A New Concept it is just absolutely fantastic when it works but when it comes to the ease of use and the quality of the generated images out of the box it is really far far behind what something like mid Journey can do and this is not even a discussion good luck trying to generate something like this with the newest 2.1 release but also for me one of the biggest problems with stable diffusion in stability AI is the fact that they're really bad at communication if you remember my previous video on the subject when I talked about the release of the 2.0 version and when I talked about the feature of stable diffusion and why they did that why they created this newest 2.0 version I had to piece together comments from emad the CEO of stability AI to try to understand why they did what they did so instead of having one single text explaining everything that they did and why they did that and all the problems they had to go through and the reasoning behind it they would rather write all those cryptic messages that we have to put back together to try to understand what they want to say I do not understand why they are so bad at communication it is just so easy it is so easy to explain why they did that and to appease the anger of the community so I'm just really really confused at what they're doing in the back end of the company because they are so secretive about it now don't get me wrong stable diffusion has a very very bright future and if every few weeks we get a minor increase a minor improvement over the quality of the previous one we could have easily something that could create images as good as me Journey currently especially once they release their dream Booth 2.0 that would allow the community to fine tune the 2.0 model themselves and create absolutely amazing models so I'm really waiting patiently for that but unfortunately as of right now stable diffusion is really really far behind thanks to image AI is like me journey and you also don't have to forget that meat Journey also is becoming better and better over time so maybe by the time that stable diffusion catches up to Mid Journey mid Journey will be even further ahead but this is not a problem for us consumers because in the end we'll still get something that is way better and better and better over time so I guess for now the only thing that we can do is just wait and see and again if you want to try now the 2.1 version yourself you can just download it on your own computer and try it out for free which is still super super cool and the real it folks now I know that this video was a little unusual so if you liked it or didn't like it tell me in the comments down below if you want to see more videos like this or not and I could maybe do something like this in the future and there you go thank you so much to my features and supporters for supporting my videos you guys are absolutely awesome congratulations also to this week's AR challenge winner yusu J for his awesome winter wallpaper submission which you saw behind me this entire video you saw jayview is also my loyal virtual supporters so big thank you to him and also congratulations and if you don't want to participate to an air all challenge what we do every week on Discord you can click the link in description down below to join my Discord server and maybe YouTube can be featured in the next video that being said thank you so much for watching don't forget to subscribe and smash the like button for the YouTube algorithm and I'll see you guys next time bye bye