Stable Cascade vs Stable Diffusion XL

hey guys this is Kevin for pixa.com and in this video of stable Cascade we're going to start off inside of stable defusion and I want to explain something about how I went wrong completely this is stable diffusion sdxl one of the early workflows that I developed in comfy UI and with this one we have a lion it's in a nice kind of R rural setting and this is the version with the refiner model and uh to my mind the one with the refiner model looks a lot better and that's one of the reasons I still use the refiner model even if a lot of people have stopped using it now this kind of complex workflow just suits comy UI perfectly and what I wanted to do was to take some of these images which I developed very early on in sdxl and to test them inside of uh the new stable Cascade and it turned out to be a disaster and I learned something along the way so I'm going to tell you what we learned and we'll take a look at what stable Cascade actually is and why it's so different to stable diffusion now this is halfway to our destination this is the state stability AI page for stable Cascade only came out yesterday and if you saw my previous video on the news about this you'll have all the information about how it actually works and what the comparisons are and all that good stuff 20 GB is what they're recommending for this one it is designed for very high quality and no the 20 GB is not the storage space it's the memory that you need for your vram for your n video card and if you look at the files you can see why we've got a massive number of different files that are available some of them very large and these will produce the best results so this is probably one of those guys that will probably be used differently to stable diffusion because not everyone has an RTX 4080 or 4090 and that's the kind of device that you're going to need to be able to really get the best performance out of this for a lot of us probably sdxl will continue to be the best option let's take a look at what you can actually do now because it is uh the hardware requirements are pretty challenging I think most of us are probably going to be playing around with the spaces on hugging on hugging face hugging face spaces so these are some here and you can choose one or another one and uh I've had different levels of success with these options and here we have our destination or at least part of our destination this is one of the spaces I tried out which really produce nice results here we have the results and you can see perfect text this is the kind of thing I just wouldn't try inside of stable diffusion cuz it just doesn't do this type of thing I would have to go to do 3 or something here however the prompt is very simple and what I wanted to do was to create something that looked like this so I wanted 3D Stone text stable and uh that worked perfectly you can see it's spelled it perfectly and it looks like Stone beautiful flowery overgrown sculpted bold brush Strokes didn't get the brush Strokes blue texture background and you can see this beautiful full design here which has the correct spelling of the words but it's also got the words in this sort of stone uh it's like it's sculpted another one and another one and another one these were all from the same four sets and you can see in the background we've got that sort of weird Watermark thing happening but it actually makes it look nice so 768 x 1024 and what I did was set the guidance scale to 15 the prior inference step to 50 and the decoder inference step to 50 that seemed to work well for text and another example here we have some more text here we've got text stable made from Marble Shadow and depth minimalist texture and it's made the stone into text reading stable here the stable is cut into the marble and uh maybe not quite as satisfactory here but with a really nice and I think perfectly accurate reflection you can see that there that is really nice and then we have this one here again we've got a little bit of reflection going on there and then not so successful and maybe just one more example once again we've got the the the text stable and here we The Prompt is just text stable beautiful flowery overgrown impressionist style bold brush drugs blue background uh blue texture background we've got the blue texture we've got the text and it looks okay looks perfect here I mean the way it chooses the font it's pretty amazing and this one just looks perfect with the with the text and it the way it looks almost handwritten that's amazing that that's really really impressive and this is the kind of thing I just wouldn't try to do inside of sdxl cuz it just doesn't render text that well but with the right settings I found this one actually worked really well now let's take a look at some examples that maybe didn't go so well or went even better than I expected this is the prompt that you saw earlier on it looks not quite as nice as the one inside of sdxl but the prompt is correctly rendered this is a sphere inside a Swiss town on a cobble street it's rendered better than with sdxl but I think the sdxl one looks kind of nicer even if it's not just as accurate this is a a really challenging one where we have um stable Cascade trying to draw a girl who's looking into a beautiful universe that is through a portal this is something that sdxl did really well really really well and I actually tested it with Bing and also with doly 3 doly 3 managed to get it just about uh sdxl was perfect this guy here struggles quite a lot so you can see we've got Devastation happening there the girl is supposed to be in a devastated area she is here and she's supposed to be 13 does look a little bit younger than 13 and then she's looking into an area which is also devastated so it has difficulty understanding context so here we have a devastated area here we have a beautiful landscape and it doesn't quite get that distinction um the way that the reflection work is by and large awesome and I do like the aesthetic if the aesthetic was what I wanted it would be perfect another one at no point did didn't actually get the the meaning corrected it did not understand this is a portal looking into a different Universe sdxl produced some amazing images uh I might link to a video where we discussed SD xl's approach but I gave it lots and lots of attempts and every single time it did not quite produce the result now we'll take a look at some more results and as you can see here we've got more of this go and I tried to give it a good go and at no point it didn't actually give me the result that I wanted here we have the spaceship or Airship steampunk Airship and what I realized was that trying to use the same prompts that I used inside of stable diffusion just didn't work uh inside of stable diffusion you got to use different prompts uh here here we have a lighthouse this is a very simple prompt just asking for a lighthouse looking asking for a beach it looks so much better than with stable diffusion you can see this little detail here uh where it's kind of weathered and then the text again I had to do quite a number of renders to get the text right if it's not expecting a particular word it does struggle with that word and putting it into a context but I wanted a signpost with picks of bir on it and uh it got it it got it about 50% of the time uh here it gave the word as a sort of um I don't know as a footnote but we have the beautiful uh white and red Lighthouse it understood the instruction to produce a white and red Lighthouse the stripes there and here we fell into some confusion this is a poster showing a Roman Center Senator on the beach at Sunrise and it's put the Roman senate but not the senator and it did that twice we've got the nice Title Here I asked it to produce a movie poster and it's got the little credits down here and uh the the title up there as well the Roman Senator we've got someone who could be a Roman s there but it didn't give me the character I was looking for and I did find keeping the prompts nice and simple was really useful to get this guy to work properly future here pix oft this is supposed to be a steampunk Airship it's not an Airship it's some kind of signpost it took two ideas a signpost and an Airship and you combined them into one and again this isn't the idea is to keep the the prompts simple the simpler the prompts the way the more it actually understands I'm just going to spit speed through and we'll talk about the these six uh images here I asked it to draw a woman in impressionist style and it did that and uh again and then I asked it to give the woman a red suede jacket or make it a girl and it did it and then I thought okay the red jacket is perfect each time what about we give it a background color make it blue and it got the background color perfect and this is again something that you really struggle to do with sdxl so I found with this one that the less you treated it like sdxl the more it kind of produce the results that you want you treat it as something completely new and you get different results and it has strengths and weaknesses but the strengths and weaknesses complement uh those of sdxl

Channel: Pixovert

Views: 10,411

Keywords: Stable Cascade, Stable Diffusion XL, AI Image Generation, Machine Learning, Comfy UI, AI Art, AI Technology, Generative AI, AI Comparison, AI Rendering, AI Model Analysis, AI Creativity.

Id: -ICRqD3pYXw

Length: 10min 46sec (646 seconds)

Published: Wed Feb 14 2024