Playground AI Optimal Settings For Best Results!

well hello people of playground today I want to talk to you guys about your settings prompt guidance quality and details in Samplers I feel like some of you guys are kind of missing the point of those settings so I want to give you a few tips to get the most optimal results let's get started now first off we recently added some new Samplers for pro users so if you go to your sampler menu here you'll notice now that we have seven new Samplers there's one missing that will be active very soon but in the meantime these six will be available for pro users not to say the old Samplers are obsolete they're still very useful but there are some benefits to the newer Samplers for example I've got the older Samplers here listed at the top and the newer ones listed below for this demo I use the same seed with a prompt guidance of four and quality in details of 10. this is more of an exaggerated scenario but visually it was the best way to show you the differences you'll notice with ddim plms and Euler they're very soft and it looks almost incomplete plms looks sharper a bit more developed but again it doesn't look like the best image with these particular settings now if we look at Hyun dpm2 and LMS they're not as soft and although the image is fairly decent it still kind of looks incomplete right and then we have Euler ancestral and dpm2 ancestral again very soft dpm2 ancestral doesn't look good with the new Samplers DPM plus plus 2m to M Karis DPM plus plus sde sde Caris this is the one that's not active at the moment but it will be very soon if I zoom in a little closer here you see that the image isn't as soft as the previous ones we saw but kind of close to here in dpm2 and with the sde variants you see this one looks fairly sharp and this one looks a bit softer then we have the DPM plus plus 2m sde these are a mouthful and sde Caris same situation not as soft they seem a little bit more developed so generally speaking the new Samplers can process the image with lower quality and details in lower prompt guidance now it doesn't really mean that's the best way to approach this like typically you're going to use more quality and details at least 30 and up maybe 25 for stable diffusion 1.5 but I kind of wanted to show you this exaggerated setting so that you can see visually what's going on here I want to show you some quick examples with stable diffusion 1.5 here are the older Samplers at the top and the newer ones at the bottom with the prompt guidance of four and quality and details it 10 and the same Samplers with prompt guidance of 8 and quality in details of 20. you see the newer Samplers look a bit more developed now to kind of make sense of what's going on here I want to talk to you about something that's termed convergence convergence is basically when the image really doesn't change a whole lot regardless of more quality in details which is also known as steps in this example I have quality in details or steps between 3 to 30 with an increment of three and prompt guidance which is also known as CFG between 2 to 20. if we look at steps 3 to 15 prompt guidance two to six and by the way this is stable diffusion 1.5 no filters you'll start to see that there really isn't any great images here and the faces although it's under hood doesn't really show up even if we move up the chart part up to steps 15 to 30. we get some sort of an image but it just doesn't feel complete right and you can barely make out a face but the one thing that you'll notice is that they really don't change a whole lot from 15 to 30. now when we increase the prompt guidance 6 8 to 10 even with quality in details of 15 at the most we now start to see some faces show up but you see they happen at different points and then as we move further down the line here the face starts to become more prominent but again the images don't change a whole lot throughout the whole generation process that is known as convergence what is the point where the image no longer changes and I mean in a way where it's not a dramatic change smaller details may but the overall image doesn't change a whole lot that is known as convergence now if I I zoom out quite a bit you see that when we increase the prompt guidance even the image starts to change in color in style but it doesn't necessarily mean that if you add more quality in details that the image is going to get better now not all Samplers are made equal this is Hyun previously we looked at Euler even though it's the same image the same seed we can't even get the face to show up with a prompt guidance of six and adding more steps doesn't help but if we have a prompt guidance of eight we actually see the face start to show up with only quality in details of nine with the new sampler DPM plus plus sde which is probably one of my personal favorites albeit a bit slower than the other ones almost double the time even with prompt guidance of four to six and as low as six to nine quality in details we get a pretty decent image compared to what we previously saw now I could go a whole lot deeper into Samplers and that can be a 20 minute video let me know in the comments below if you want to geek out on that but now that you know the term of convergence and you see how prompt guidance and quality and details can affect the development of your image generation there's a couple takeaways that I want you to consider I don't know if you noticed as I was showing this if you'll notice plms here looks under processed or underdeveloped and of course with a prompt guidance of 4 and 10 we know that this particular sampler most likely needs more quality in details or an adjustment to The Prompt guidance so what I'm going to do is reuse the same seed and we're going to process this with different settings I'm going to select the same sampler and instead of four we're just going to bump this up to six maybe and increase the quality in D details to 25 I did use deliberate as the filter and we're going to generate here's the same image side by side and we see now it looks a whole lot better now if you ever come across where it looks like it's artifacting you just have to play around with either setting typically quality and details you don't have to go past 30 knowing what you know about convergence so the other Factor could be the prompt guidance it's either too low maybe even too high similar situation here with the newer sampler DPM plus plus 2 msde we actually had a prompt guidance of eight quality in details at 20. so let's do the same thing we're going to copy the seed let's put on the Rev animated filter it might have been a bit too much but we'll leave it for now and we're gonna bump this up to 30. let's make sure we select the right sampler and generate this image and here's the process image most of the filters that you see on playground the developers usually give a suggested sampler to use Euler ancestral is a pretty popular one but the majority of them these days recommend the newer Samplers but depending how it's trained those Samplers are best to use to get the best outcome there's a couple videos on filter breakdowns in both those videos I mentioned that there's a guide here that I posted for the best prompt guidance and quality in details now mind you these are suggested but there's also a section here for Samplers which Samplers work best we have our filters here on the side and some standard settings that you can start with and you'll notice a lot of them have the DPM plus plus variant let's take a look at a few practical applications just remember these are just suggestions you're free to experiment as you please but if you use a spreadsheet as a guide let's say we want to use rev animated for example let's collector have animated the suggested sampler is DPM plus plus 2m carius prompt guidance is a pretty wide range three to ten I like to start between five to seven and for quality and details you can go between 25 to 30 usually let's go to 25 and we'll generate four images as we look at the images here they turned out pretty well although I want something a bit more contrasty one of the other tips I can give you is that if you increase your prompt guidance you tend to get more contrast in your image deeper blacks that type of thing we can also increase the quality in details to get different results with that being said I'm going to change this to 8 quality in details to 30 and we'll go ahead and generate another four Images notice now that we have more prominent colors the Shadows the blacks they're a bit more intense yeah I would be totally happy with this set of images now let's try an s DXL model I'm going to change the height to 1024 and the width to 768. let's use mbbxl this is one of my favorites well I like a lot of the sdxl filters we'll leave the prompt guidance at 8 and typically for sdxl I tend to stay between 40 to 50. on occasion I may go to 60 to 75. it really depends on the type of image you want as always it will take a little bit of experimentation definitely a different result with mbb XL compared to rev animated different details in colors of course it's sdxl so we can do a higher resolution and personally I much prefer the mbb XL images in terms of the detail and style in the meantime if you happen to be new to playground AI make sure to check out any one of these videos on how to bring your own art to life until then my friends this is playground

Channel: Playground AI

Views: 9,429

Keywords: playground ai, playground ai tutorial, optimal settings, playground ai filters, stable diffusion, ai art, stable diffusion online, stable diffusion samplers, playground ai optimal settings

Id: wiPCSEu4YCQ

Length: 10min 59sec (659 seconds)

Published: Sat Sep 23 2023