10 Stable Diffusion Models Compared!

all right we've got a fun one today we're going to test out 10 different generative AI art models most of us are familiar with all the models from stability AI like stable diffusion XL but there's a whole universe of models that have been fine-tuned and trained beyond that for either different aesthetic values or even textual embeddings that help improve the way that it follows prompts I thought it would be fun to take the exact identical prompt and run it through 10 different models and we can see how each one of them looks at the end I'm going to throw these on one of my websites and you can go through and actually vote on which ones you think are the best here's a list of models that we're going to test today it's Proteus V2 SSD 1B playground V2 stability AI sdxl this is kind of the Baseline Juggernaut XL fine-tuned model on top of sdxl Juggernaut XL v9 anime XL which is fine-tuned specifically for anime and cartoons Kandinsky 2.2 real viz XL version 2 and dream shaper X XL turbo links to all of these models are down in the description so you can go and download those and try them out on your own computer if you have the ability to do so if you don't have the ability to do so head over to pixel Dojo AI all of these are loaded up and ready to go what better way to test these and just go through some of these examples so the first one we're going to do is this photo of a red-haired girl this one is beautiful lady freckles big smile Ruby eyes short hair dark makeup hyper detailed photography soft light Head and Shoulder portrait cover we're looking for two different things here one the prompted hearing so how well it actually follows the detailed instructions I provided in the prompt does it actually make the eyes Ruby colored for example I haven't had too much luck doing that in the past the other piece of this is the aesthetic how does the image actually look at the end is it visually pleasing those are the things we're going to look for we're going to start with Proteus V2 and click generate it's going to generate two images for each model we're we're going to keep the same sampler scheduler everything else identical between Generations the only thing that's going to change is the model that's being used and here are our first two images you'll notice that both of these are freckled red-haired girls very short hair shoulder up Follow The Prompt very well you'll notice that both of them have Ruby colored eyes this is something that in my previous testing I had a lot of difficulty with sdxl getting that exact precise prompt following and I've been really impressed overall with this particular model not only does it come back with really good quality results it's also very fast next up is SSD 1B this is actually a fine-tune stable diffusion XL model that has 50% fewer parameters and is 60% faster at generating an image you'll notice right away this isn't as high a quality as Proteus you can see that the face lost a lot of detail it doesn't look as realistic it doesn't have Ruby eyes so it didn't catch that part some of the teeth over here on the left hand side image kind of missing just not that same quality you get out of some of these other models the third model we're going to test is playground V2 what's interesting about this is it was actually trained and fine-tuned with 30,000 images from mid Journey so it's been purported to have much higher aesthetic quality score than stable diffusion XL but we'll be the judge of that and here's a result it only comes back with one image instead of two I think maybe if we Tred to different scheduler we might get a higher quality result but it's just there's a lot of artifacting it's kind of out of focus really oversaturated not one of my favorite images next up on the list is stability AI stable diffusion XL this is kind of the OG the base model that a lot of these others were trained upon and here are the sdxl images you'll notice they're not quite as saturated as some of the others they aren as sharp as some of the others either I think by default it's sort of softer images that end up coming back from sdxl the cool thing is sometimes I'll take these because I I like the aesthetic of them and I'll run them through my image upscaler and it sharpens them up and it adds a little bit more detail and you end up getting a really nice picture now looking at these though it did follow most of the prompt instructions you've got a freckled woman she has short hair you'll notice that neither of these images have Ruby colored eyes though do Juggernaut XL version 8 Juggernaut XL was trained and fine-tuned upon stable diffusion XL just like a lot of these models were and this is the eighth iteration of that there's also the ninth which we'll test out after this and really what they did is they tried to fine-tune it to get a higher aesthetic score something that's a little bit more visually pleasing than the images that come out of sdxl you'll notice right off the bat that it's a lot sharper a little bit more refined so you'll see this one adheres to The Prompt fairly well it looks like you've almost got Amber or reddish colored eyes so it's a little bit better than sdxl in that regard and otherwise just nice high quality photos here if you zoom in you do have a little bit of weirdness it looks like there's maybe something going on with the teeth here you could probably do an image to image and kind of refine this a little bit but there is that at least let's see how Juggernaut XL Version 9 compares to version 8 oh that's interesting it doesn't like this prompt something's going on here all right so you've got the Ruby eyes uh this is these are kind of terrifying looking honestly so you know you don't see the abnormalities in the teeth like you did in version 8 although there's this weird like masking around the mouth and the eyes and then the skin just looks like too wet and glossy other bits of this are nice like the hair looks ultra realistic I'll give it that and then the eyes are very detailed and they're definitely Ruby so this followed that prompt adherence really well it's just I don't know I I don't like the overall aesthetic of this compared to version 8 kind of creepy and when you zoom in you can see that even more looks like there's this masking around the mouth like it must do some sort of fine tuning there specifically I don't know odd one this next one's called animag XL you can see it was trained on a whole bunch of anime images and it should come back with some really nice quality results now you can't compare this obviously directly to some of these others because they're trained for more photo realism than this one but depending on the type of project you're doing this could be a good alternative and here are the two result from animag you can see both of them have those beautiful Ruby eyes really high quality results freckles everything else same shoulder up view that we're supposed to get from this image so again you can't really compare this directly to the photo realism of some of the others but if you're going for that anime look really should check this model out next up is Kandinsky 2.2 Kandinsky is interesting because the images that come back have this like darker I don't know it's like a surreal ISM to them they really nice teeth uh it looks nice overall it's it's just a very particular aesthetic and if you're going for this look it's great it's just slightly different than a lot of the stuff you're going to get from the other models now you'll notice it didn't quite adhere to The Prompt fully neither of these models have Ruby colored eyes they both have red hair but overall it's it's a really unique look and you can especially see that when you zoom fully into these images now one of the other things I will point out it looks like it gets a little bit too precise in its patterns you can see the freckles have almost too much of a a normalized pattern to them they're not quite realistic in that sense like you'd never really see somebody that looks quite like this it's almost like too symmetrical next up is real viz XL version 2 here are the two results from real viz XL pretty high quality the eyes are a little bit odd to me they're not quite as realistic as I've seen from some of the other models that we've used it also didn't adhere to to The Prompt as well she has dark brown hair which is short which is good so it followed that piece still no to the Ruby eyes I don't know just an interesting overall look but not quite as realistic as some of the others you almost get some of that more symmetric pattern in the freckles as well like we saw in kinsky it leaves us with dream shaper XL turbo and being a turbo model you can typically get away with running fewer inference steps in order to get the same highquality image out of it but like I said we're not going to touch that we're going to leave everything as is all right just like a couple of those other models that we've seen before it's sort of overly saturated overly stylized I don't really like this one either you can see there's way too many freckles in some of the places it's not as symmetric as some of the others so that looks better but it's still kind of bizarre the way the skin looks it just doesn't look realistic to me the hair the lighting all these other details are pretty good it looks like it actually captured Ruby eyes in the face as well but I don't know there's just something about these that does not look right and look if you're doing something other than photos of a person you might get better results from some of these other models like here let's test dream shaper with a dragon instead of a girl and see if you get back a better result and here you go depending on what you're looking for this might be a better result than you get back from some of the other models I think that's sort of the bottom line here a lot of these models are trained on very specific types of images specific data sets and so they're going to do a better job at producing some images over others so it really comes down to your prompt in the art style you're trying to create I think for me personally Proteus V2 sort of stood out as one of the leaders in this and that's something I didn't really expect going into this because it's not a model I've heard a lot about outside of my testing the last couple days Juggernaut XL is something I think most people have seen it's one of the default models for Focus for example and like I said I'm going to drop all of these images on a page so by the time this video goes up you'll be able to head over there check them out for yourself and go ahead and vote in a poll tell me which one you think is the best leave me a comment and let me know which one you liked out of these don't forget to grab and download your favorite model or head over to pixel dojo and you can run all these no problem as always thank you all so much I'm Brian love it and remember all your Tech are belong to us I'm the virtual profit in the tech town breaking down AI wearing the crown from BAS to complex never let you down all your tax ain't I earning the renown

Channel: All Your Tech AI

Views: 8,162

Keywords: stable diffusion, stable diffusion tutorial, stable diffusion xl, ai art, how to use stable diffusion, stable diffusion ai, ai image generation, stable diffusion art, stable diffusion prompts

Id: XsG_-UlS_NI

Length: 10min 35sec (635 seconds)

Published: Fri Mar 01 2024