SDXL Models Compared: See How 5 Top Versions Perform with One Prompt

hi everybody Stuart from Super machine here it's a very exciting time to be alive with all of the advances in artificial intelligence and especially image generation we're trying to keep on top of the latest developments and have rolled out sdxl across super machine we've not only released the base model we've had a go at training a few of our own models and we'll continue to do this in the future and we've also chosen a few of the best models which are open source available for you to use in super machine today but that leads a question and something that I've been debating inside about the stxl models specifically with the original stable diffusion models training a checkpoint on top was relatively simple and easy to convey the style that you wanted with sdxl because the base model is of such a high quality it is a lot harder to train a style or a specific way of doing things so in this video I just want to showcase a image prompt across all of the different sdxl models that we have in super machine and then I want to show you what I think is the future for creating images with sdxl and I believe this to be using Laura's on top of it and we've added a few of those so I will show you what they look like also so let's get into it logging into my super machine account to get to the sdxl models all I have to do is go from all models at the top to hitting sdxl you'll then see the five models that we currently have there is the base model here there is the two models that we have created ourselves super machine abstract XL and the Mystery model XL and then two of the favorites from the open source Community dreamshaper XL and mbb XL Alpha you'll see these are the settings that I've used for generating I used the 1024 by 1024 and I used this prompt here futuristic cyberpunk robot smiling mailbox design looking at the camera portrait I used this sampler the DPM plus plus 2m Karas and I used the guidance of seven and the steps of 12. what I really want to show you here is I generated four images from each of these models and just to show you that there's no real standout differences between them so dreamshaper XL provided these four outputs the mbb XL provided these four outputs I will say with the mbb XL you are getting more of a human touch where it's shown a few images of humans compared to just robots the mystery model XL is roboty like that and the the original sdxl looks like this and the super machine abstract looks like this I really like these first two they're kind of anime style and then we've got these other two which are more like photo realism style but I think what I'm trying to show here is unlike you had with the original stable diffusion fine tunes and the custom models there isn't one which you could say okay this is a bad model it's really the way that they've been trained produces good outputs I wouldn't say any of these images are specifically bad but it's a really how are they influencing the base model and are they making stuff which is better than the baseball so the base model is this one here and it's a case of are any of these a massive improvement over the base model or are they just influencing the style slightly and if I hit generate say 20 times would the base model give me some of the outputs that some of these fine tunes are giving me and the answer is possibly so there is a future for fine tuning these models but I think it's a lot harder and more difficult than it was before and I think there's still a lot of people trying to figure out the best ways to fine tune and it's not just a case of dump thousands of images in hit fine tune on a software or code come back in a couple of days and wham bam you have a new model I think there is a lot more thought that needs to go into it and this is definitely what we found when we did our fine tunes for the super machine abstract which is this one which is mainly trained on Ghibli style images so I'm super happy to see these two these two outputs here because it sort of shows that what we did kind of works there and then the mystery model XL was trained on uh Lexica aperture type images so it's interesting as well to see the output I can see how it has been influenced by that so it's yeah when we're talking about these models it's it's hard for me to talk about the models where I don't know the data sets that went into it but the ones that we train in-house I do know what went into it so I can see and really sort of um understand more what the image is out and what the images in sort of affect the outputs of the models so this is the first part of the video where I just wanted to do that basic overview and say you know each model is good they're all good now but they all offer something slightly different only a little bit different I think the the the benchmarks across different models is less now there's less differentiators and this is where I think Laura's are going to change the game with sdxl you're probably wondering what is a Laura and how does it affect the image generation well there are certain advantages to using Laura's when you are trying to generate images and also when you are playing around with different models one of the main advantages of a Laura is you can use it across multiple models so any of the sdxl models can use Laura's which have been trained on sdxl and the same goes for the base stable diffusion and Laura's trained on that the way that you train and Laura is to take a sample of images so say for example I wanted to train a Laura on a Japanese kimono so if I wanted to do that I would take multiple images say 30 images of Japanese kimono or women wearing this Japanese kimono and by doing this I would be able to then train a Laura which would know okay when I use the term Japanese kimono or kimono or whatever it will know okay this is what the user is after and this is what I should give the user so lures are very flexible in that regard to be used on top of Base models and you know if you were to use a Laura on top of an anime model compared to a photorealistic model you would have the same effect but it would look anime compared to photorealistic uh I think the future of sdxl is going to be using Laura's on top of the base model or any of these fine tunes which are pretty similar to the base model and then by doing this you'll be able to get different stars and different influences out so we've loaded two of these up onto super machine which you can play with today and we will be adding more in the future we have added a chibi style Laura and we have added a t-shirt design style Laura and I'm going to show you how they look and also how you can call them within your Generations on super machine the first model we've added is this chibi style XL and we are looking at Civic AI on how we use this so we have added this to Super machine and the trigger word is chibi here so how would I use this in super machine well I would come to my prompt at the top here I would put the word chibi at the front of it and I will put the Laura so we are putting in these uh brackets the Laura and we are putting chibi X sdxl and then colon one so this means that we want to give the Laura the full weight of one and this is our prompt so if I wanted to change this to something else I would change what I have highlighted here to whatever I want to change it to and by using this you can see how it affects the images so again I've done four generations for each of the models which we have for sdxl and you can see how the outputs look so this one is the dreamshaper XL we have those four outputs the mbb XL Alpha looks like this the mystery model XL really like some of these the original sdxl model and the chibi uh the super machine ass abstract XL model so you can see this uh I like this one he looks like a uh an officer it's pretty cool so these are the outputs that we've got and compared to these previous uh outputs without the Laura you can see that it's had a massive effect on what we are doing and what we are creating so that is the advantages of using a Laura because it completely changes what is possible with the pictures and just to go over again how to call it this one in particular use chibi at the front you could use this anywhere in the prompt really but I think at the front works well and then you need these uh brackets like this and you need to have the word Laura l-o-r-a and then you need colon and then you need chibi sdxl and then you need a colon and then you need one and then you need to close these weird brackets and if you have all of that then you will get images like this out so let's have a look at our second one now which is the T-shirt design so for the t-shirt designs I am using this Laura from Civic Ai and to call this I can either write t-shirt design or sticker and you'll see that it says use t-shirt design or sticker and I need to provide the details for the Laura so I will show you how you can do this in super machine so the way that I've done this here is I've put futuristic cyberpunk robot smiling t-shirt design Laura so I've put the trigger word here t-shirt design and then I've put the Laura code here so we have Laura colon t-shirt sdxl colon one all within these weird brackets so if you do all of that then you will be able to generate these images and what I just want to do as a test whilst I am making this video is just change the T-shirt design to be at the front and I'm going to generate four of these images using um the mystery model XL because I think that was the one which gave us the best outputs here and I just want to see how this looks so whilst this is generating and remember the sdxl models Generations can be a little bit slower than the original stable diffusion Generations it's something that we are working on um but you can queue them up and they will complete so just add them to the queue but here you can see the difference between the different models using this t-shirt Laura so this is dreamshaper XL you're getting these outputs I would say some of them would look good on the T-shirt some of them are more just uh generic sort of outputs here the mpb XL has taken it quite literally and it's provided a few people uh showcasing t-shirt designs so that is quite funny and then we have uh this one's quite nice that would be quite nice on the t-shirt and this one also the mystery model XL I think has done the best job on this I think this first one second one third one less so and the fourth one would make really excellent t-shirts and then we have the original sdxl where we have this one which I think would be a great t-shirt these two probably less so and then we have again the T-shirt design taken quite literally and finally we have the abstract XL where we have some interesting outputs I think this one would look quite nice on a T-shirt If you were to just extend this uh this background not have these buildings on the side um fairly interesting and then the one that I just queued up moving moving t-shirt design to the front has shown probably worse outputs so you know when you are prompting these things where you decide to put the words can have an effect on the outputs so by moving it so this is it we're at the front on the mystery model XL and then this is it at the end on the mystery model XR I think you'll see that these are a lot better output than these but hey this is why we run experiments so we can see these things and really understand what is working and what isn't so I hope that gives you some ideas of what's possible with stxl and also using Laura's for scxl within super machine I will put all the details of how to use these loras in super machine below this video so you can check that out for yourself if you haven't used sdxl yet please try it out we have these five models for you to experiment with within Ricky and you can also I try them with Dolores today if you are looking for specific models in the future or you would like to see specific lauras or you'd like us to create some then please let us know what you're after and we can take a look at that and see if we can make that possible I hope you've enjoyed this video feel free to leave a like or subscribe to keep up to date with what we're doing at Super machine and all of the latest developments we're excited to keep working on this technology and we're excited to see where it goes thank you

Channel: Supermachine

Views: 1,230

Keywords:

Id: rnuG9mDl2qA

Length: 16min 37sec (997 seconds)

Published: Wed Aug 16 2023