The Best Refiner for SDXL - Stable Diffusion has NEVER been more DETAILED!

in this video we'll take a look at perturbed attention guidance this is a new method from our friends over at Korea University and Samsung Electronics now this method is already available as a Technique we can use inside of stable diffusion now the basic idea is that it Alters the way that stable diffusion looks at detail you can see some examples here of how it can work with control Nets image repair and if we go further down you can see some examples here so so they've got examples of conditional generation and this is the Baseline and if we just scroll through you can see the result with the new technique P that's before take a look at the dog right at the bottom left look at how it looks after p and look at how it looks before P there's a huge difference there in the results and we also get similar results or we get similar improvements inside of stable diffusion so this is conditional generation inside of stable diffusion and the four and with P it's a massive difference we can also see what happens when we apply CFG so this is with CFG and this is with CFG and P now the framework is one which is explained in this paper here goes into quite some detail and some of the examples they give are pretty pretty impressive I was particularly impressed by the results that they got with the control Nets because sometimes control Nets as you can see in these images here they can be a little bit they can be a little bit unpredictable but with the technique applied you can see that we get much more clearer results so it's all about detail there are options for stable diffusion web UI comfy UI it's all there let's take a look at this technique inside of comy UI now the technique comes in as to well basically couple of nodes we've got the P perturbed detention guidance node and this one may be inside of your comy UI if it's completely up to dat and uh if we go to manager we can also look for perturbed attention let's see let me just look at the installed ones there we have perturbed attention guidance for from this contributor here that is what you need to download if you want to have the the full shebang but you might find that the nodes are already inside of comfy UI if you have your comy UI up to date so we would search for perturbed and there is actually one which is slightly more complex and this one you probably won't need to use this one it has the same scale and it suggests three as the default and we also have adaptive scale unet block and unit block ID so you can go into the woods if you really want to with a lot of this detail I'm going to suggest that you probably don't need to actually use the advanced one you can probably just get by with the perturbed attention guidance simple node I've got two noes here and one is for uh the workflow that has got the refiner and we've got the refiner there sdxl and we've got the non-refined ones here now having run several tests I can tell you the the results are pretty impressive this is without this is without so we'll focus perhaps here on this image here you can see here we've got the guidance of one I figured out that one a scale of one doesn't produce much difference from the default and a scale of three produces pretty decent results this is without take a look at the detail down here at the front this is without and if we go to the one that has got the detail from the new method take a look at how there's more detail in this uh in this staircase here this is a beautiful this is a really beautiful prompt one I borrowed from one of our friends over at CIT AI oh by the way the prompt is from a user called Co on Civ Ai and a lot of what you get with stable diffusion depends on the prompt but you can use these different techniques like like the one that we're using here to to change the behavior look at this image here this is without P this is with ph there are subtle differences and we are really focused on the subtle differences rather than the well you could if you want to assess the overall look I would say the overall look of this one is pretty impressionistic it is really nice it looks very hurried as though someone just did it in a very quick quick way this one is more structured there's more detail we can actually see quite a lot of detail on the church before and after but you can see the detail kind of makes a little bit more sense it's more cohesive after and we can actually see detail here that wasn't there before in this particular section and the tree looks pretty impressive this is without the refiner this is the original image without the refiner so it almost is working like the refiner but not doing some of the weird stuff that the refiner does sometimes now just to give you a word of word of warning with this particular workflow it is fairly complicated it is used on the Mastery course over at udemy and it has got the model sampler tone map it's got a very complex node there which I've used to change the way the CFG scale functions you've got two things that are kind of affecting the the way the image is interpreted this sampler this particular node here and also the new node the pH G node and when you use it I would urge you to take a look at the number of steps and to play around with the number of steps to see what sort of impact the new node has when you change the number of steps I did actually do that and perhaps we can take a look at that a little bit later on this is a uh this is a wonderful image of a bird and here the refiner actually did a fantastic job in adding detail you can see the feathers there you can see this little patch of Light which adds this almost almost ethereal feel to to the image the feathers look fantastic the talons look a little bit messed up with the refiner but the detail at the back there which is prompted this is another prompt from Civ AI this is a beautiful image in both circumstances but I think the hair the feathers there look fantastic if we look at this without without the P let's see what it is one heal a difference I'll tell you that so that particular prompt comes from Princess art over at Civ Ai and it's a wonderful prompt no this is after applying pH this is afterwards and let's see what happens before look at that look at the lack of detail in the talons look at the feathers at the back there part of the prompt is just not actually shown in the image the way that the color just splashes everywhere and then when we move over to the new one just take a look at that the way the feathers come to life the color splash just emerges from the image we also have slightly better looking talents there although I'm sure if we were to put these in the refiner the refiner would probably mess them up again so I think this is a fantastic way of working particularly for people who use sdxl and maybe are a little bit wary of using the refiner cuz sometimes it does produce unwanted effect this is maybe a technique you should be able to get some some real Advantage [Music] from good I wish you were M wish you were [Music]

Channel: Pixovert

Views: 1,699

Keywords: Perturbed Attention Guidance, SDXL, Stable Diffusion, PAG, image generation, AI, machine learning, deep learning, Comfy UI, refiner node, image quality, detail

Id: 7NhKGdK8MTw

Length: 8min 28sec (508 seconds)

Published: Tue Apr 16 2024