Playground AI Tried to Break the Internet. NEW Stable Diffusion Base Model!

a brand new image AI model that outperforms stable Fusion XL well that's what playground claims and it is free to use both on playground right now and locally to download so let's check it out and see if their claims are true oh and why was the baby jalapeno shivering it was a little chilly AI introducing playground version 2 a new commercially open model from our team that we trained from SC batch most notably our model was preferred 2 and 1/2 times more than the current leading open model sdxl let's quickly look at the comparison here so this is playground V2 and this is sdxl now okay I know what you're saying guys they did the comparison with the sdxl base model which isn't fantastic but I mean it is what it is we're going to have to um make sure that you guys compare it to the custom sdx and models and see if it can outperform them from the results that are uh showing here the playground V2 I mean it looks kind of cool and it continues on you says you can try them all on our website the playground well it doesn't say that here but you can actually download diffuser models on it uh it says read our full blog post on The Benchmark so we're going to do that so here's the the blog post so here's a couple of example images I'm going to zoom in a little bit for you guys and this is we're providing open weight for playground V2 an early preview of our efforts to make increasingly powerful graphic model and it says that this model is also available on hugging face we're going to get to that in a bit commercial use is permitted I'm going to show you how you can download it and uh we'll see if um you can get some something running from it here's a couple of cool images I mean looks kind of cool a little bit too smooth in my opinion I would love to see some photo realisms as well this is just those those kind of sdxl smoothed out images that um I think people like when they see a comparison it's like Oh I like that one but they get tired of it kind of quickly you know but who am I to judge I don't know and that's what they say here in early benchmarks have shown the playground is is preferred to half times more than stable Fusion XL when we talked about already again I would have loved to seen that with a better sdxl model than the base one there's some more data here if you want to go check that out this is basically just the the user data from the comparisons I'm going to link that in the description below here's again the same comparison we saw earlier and it says also the baseball weights are available in 256 and 512 pixel stages on hugging face however if you do some clever searching around you can actually find this one which is the 1024 aesthetic model so we're going to download this in a bit if you don't want to do that and just want to play with it now it is available on playground.com and when you log in you're going to have basically a user interface if you never used this before I mean the prompt is to the left here you have some settings to the right you have the size the guidance quality details seed whatever this's take some prompts here from the main page so let's take this one looks pretty realistic and this was made with uh table Fusion Xcel so just going to copy this we're going to take the negatives here oh that was the same ones okay cool uh did it have any special settings here they used ho at eight yeah we're just going to do Oiler a for now we're going to generate this we're doing a CG 3 which is uh low but that was the default value here so I'm just keeping that I lower the steps a little bit to the 30 instead of the 50 but I mean this is uh it's kind of coolish it's not as good as this one here this is very realistic this is more of a painterly mid Journey kind of vibe let's change this into raw photo of raw photo portrait of woman riding the bus cinematic film still color graded we're just going to leave the the negative ones and generate a new one here I mean it's a little better we're getting some good blur here on the background however I I think the fa is uh and the skin in general is a little too smooth a little too airbrushed eyes are a little weird but I mean it's just one image one seed can't judge it fully on that I'm expecting you guys to uh go test it out and let me know in the comments below what you think about it so far uh let's just take a prompt that's more stylist here let's take this blue no I would I liked a little longer prom but I mean sure let's try okay let's copy this we have blue eyes red black hair Rusty helmet crowds of medieval cat night we got some sort of a man here wearing a helmet with cat ears and a cat in in his lap as well was kind of kind of wild I I think we're going to head back into um the prompts here and find something where we have some more styling here so here we go you have some oil painting here this one's kind of cute just going to remove that so it doesn't mess with the prompt they oil painting vibrant colors very long hair exquisitly decorated medieval tapestry background y y y y y so this is more of a painterly style but but honestly I don't know what they were comparing to well obviously I know it was the base model but whatever I'm testing like all these images with custom XEL models when I take the prompts and put them in here they aren't fantastic now this do I think I think they had Oiler at 50 steps as a default fault so we're going to give them the benefit of Doubt a little bit so three was the CFG 50 was the steps and Oiler was the the sampler so this should be the default settings for playground it's not much better to be honest I have to tell you so let me see if I can get this running locally for you guys I'm not going to make any promises well I did get it to run locally just not probably not in the way that it's supposed to but I'm not a developer one of you guys can probably tell me how to make it better anyway uh let's head head into that now so we're tacking phase and files and versions here and if you go into the unit here you actually have some safe tensor so I lo downloaded this one which is the full 10 GB file you can download the fp16 it's not going to make too much of a difference and what you can do then is you can go into your uh comy folder models unit and then you're going to drop that in here I renamed it playground so I know which one it is and whenever you head back into um confy let me actually show you here let just load a default here now so if you get let me zoom in if you get unit loader you can actually load the playground model here and instead of uh loading the checkpoint here we're going to drag that into our K sampler now this isn't really how it is intended so bear with me I'm not a comfy expert by any means if you load a Exel model here just for the clip and the vae well you can load a vae manually if you want to vae loader load an sdxl VA drag that to there and if we run this now let's actually change the resolution here 1024 by 1024 it will actually load from the unit loader and use that model while creating our image now we will use the clip from our base sdxl or you know the other the older versions but honestly I have I don't know how to load the clip from the diffusers uh when it's like this um if you do help me out put it in the comments below this was uh how I managed to load it and you know it it kind of actually works and you can see let's set this to fix to change this to one here let's lower the C CFG to three which they had as a default we using Oiler again and if we run this just to get the Baseline here I'm going to show you guys here okay this is the image looks well actually pretty good and if we take the model then which is the Excel model my Sebastian's merch and this playground isn't being used and we're queuing this up you will see that we're actually generating a different image so that tells me we're on to something or well it actually semi works so let's see if we can um make some comparisons here so let's take this prompt here drag that into here going to take the negatives here put that into the negatives here now this is just the base text encoding or the prompt nodes you could load like like um I think there's the like sdxl base prompt encoders I test a little bit with that I'm not sure if that's better but what what my goal was with that was to try to replicate the results from playground into here now I didn't manage to do that even though you know I tried to get all the settings but we are getting you know similar looking results again is this the best way to do it probably not but but you know it kind of works works and it doesn't look terrible compared to the images in in playground and I mean if you look at this one here and that one there you know they're almost trying to do the same thing let's take this one here do that one and we had what 30 steps for that 30 I me this is kind of good we got a cat and the little helmet here and actually got the the ears out this time as well that was I was quite happy with this image now we have the playround mode loaded here let's take this one the ra photo of woman riding the bus let's load that up in there do we have a yes we had the negatives for that one as well and I mean we're getting similar rati results are they great no is it getting there maybe let me know what you think this was just released like half an hourish ago so I'm just just trying to find a way so you can um you know play with it again don't be afraid to show me a better way down in the comments I really want to learn as always have a good one see you

Channel: Sebastian Kamph

Views: 37,893

Keywords:

Id: nYR27Qhuk8E

Length: 10min 4sec (604 seconds)

Published: Tue Dec 05 2023