5 EASY WAYS TO MAKE BETTER AI ART | Stable Diffusion 2023

hello everybody my name is Binks and today we're going to talk about five different tips for stable diffusion to maximize quality productivity and efficiency in your AI workflows the world of Open Source is extremely useful for people like us who like to Tinker on our own computers and especially for free sometimes it can require a little bit more work to get the images and the the results that we're going for but I'm just going to give you a couple different tips that can work together and try to make your process a little bit easier and just before we get started here I'm using protogen version 3.4 it's a photorealism model I have a prompt here you can pause and type this in if you want I also have the negative problem from my negative prompt videos just link and down in the description and up below if you're interested very short and very useful I'm going to use the DPM plus plus SD Cara sampling sampler with 30 steps I'm going to leave the width and the height at both 512 and I'm going to turn the config scale down a little bit to something like a six all right and with the settings that we have in already we're able to generate an image but it's not exactly desirable so that leads us to our first tip and if you follow nothing else from this video just simply do this check the restore faces box I'm going to go ahead and regenerate an image now that I have the restore faces box checked and I'll show you just how big of a difference it is and already just by checking that box we have a much better result now there are some things that are changing in between all of these images which lead us to our second tip which is increase the batch count to iterate faster it's no secret that with stable diffusion we are essentially spending a load of time just typing in prompts and moving sliders and just changing tiny settings here and there to see if we get what we want but you might be able to generate your target image with the settings that you already have by simply turning this batch count number up to something of at least two and clicking on generate again we can generate two different variations of the same type of image and which is closer to the direction that we want now as a quick note this obviously will take double the time because it's generating two images but that extra 15 to 30 seconds especially when you're generating low resolution images is totally worth it to save you some time in the long run so now with that option turned up we've generated two images and we can see right away if we had generated this one on the right as our first attempt we might have started changing settings and changing a prompt when in reality we didn't need to because this image on the left is in my opinion much better than this one on the right and we could work off of that to continue getting what we want all right great now we're going to go ahead and move on to tip number three which is when you're generating your starting points you want to try to use the XY plot features to test things more quickly to access this you come down to the script section of stable diffusion choose XY plot and in here we can start to look at some of these settings there's an X-Type and a Y type these are represented of each individual axis of The graft and how exactly one of my to find these values it's important to note that you do not actually have to use both the X and the Y and I'm going to demonstrate that here by changing our X type over to steps and setting the value between something like 30 and 50. now right off the bats table diffusion doesn't understand how it's meant to increment up so what I'm going to do is put a 10 here in square brackets and just tell it that I want to generate 10 images between 30 and 50 steps I'm going to make sure that draw Legend is checked and go ahead and click the generate button all right and as you can see as this is finished generating we can see what exactly the step increase does we put that number 10 in Brackets which means we have 10 images here that were generated and they all range between 30 and 50 steps on this particular sampler some of the illustrations are actually kind of getting better towards the end as far as the facial structure goes but most of these work well and in fact something like 30 or 32 works just fine for my liking the other advantage to this is of course there is still a y-axis so what I'm going to go ahead and do I'm just going to bring this down to a number like three I want to generate three images between um let's say 20 and 30 and then on my Y type I want to actually vary the config scale and what we'll do here is we'll set this quite low to something like 4 and let's make a cap out at an eight instead of square brackets let's open up some parentheses here and let's do plus um 0.5 and so what this will do is this will go between four and eight and it'll scale It Up by 0.5 each time we want to again make sure draw Legend is checked and we'll go ahead and hit generate one more time a few moments later and actually like earlier it's important to note that when you go to generate this many images it is going to take a significant amount of time depending on the hardware that you have if you don't have really maxed out Hardware it's going to take longer than it would if you had a 40 90 and a top of the line processor what I'm actually going to do is change this between six and eight and I'm only going to have it increase by one every time just to kind of speed this process up a little all right and here we can see just how powerful the XY plot is we have on the x-axis we have steps 20 25 and 30 going up just like we told it to and then on the y-axis we have the config scale going from six to seven and we have our images laid out in a nice grid so we can kind of see how the differences change one of the important parts about AI that I see a lot of people don't quite understand is that it's more powerful when you keep tip number four in mind which is that it's an iterative design process what I mean by iterative design process is just you're supposed to start with nothing get something that's closer to what you want take that image and begin modifying that image to get closer and closer and closer in an iterative way towards your goal so let me give you a quick demonstration of that we're just going to grab one of these images from our generation here and we're going to head over to image to image and we're just going to drag our favorite of those into here and now that we have this image in here you can see that we have a prompt and a negative prompt space and more or less the same settings as before except there's this new one called denoising strength um what we're going to do is actually just copy and paste everything from the previous screen on here I'm going to set it to the same sampler and this one was from the 30 columns so we're going to keep it set on 30. width and height is fine config scale it was a 6.5 and this denoising strength is a factor that we haven't seen there's a nice little tool tip here determines how little respect the algorithm should have for the image contents at zero nothing will change in at one you'll get an unrelated image so this is what I mean when we're going from one image to a new one in an iterative process so what we can do is let's say we want it to be about half new and half old and we can change this to be a silver crown let's do rubies as the stones let's get rid of medieval architecture and we'll say um Victorian architecture and we'll get rid of all of this stylized text here as well and again you don't have to do exactly what I'm doing this is just a demonstration for how to use this stuff on your own and like I said before I'm going to turn out my batch count to two so we can get a good idea of kind of what is going to be generated and go ahead and click generate so as you can see in a kind of before and after here um the noising strength value there isn't doing well in our config scale also isn't doing well so when you move from text to image to image to image a lot of these settings aren't going to be quite right for your application I can tell from experience that I need to turn my config scale up a little bit and I actually want to turn my denoting strength up just actually down just a little bit to something like 35 and I want to make sure that we still have restore phases checked and we'll let that run again okay and as you can see we have a much better result so we're getting kind of a shift in the crown it's not massive the faces are a little bit different in between them and as you can see the background is changing we want to give it a little bit more denoising strength and a little bit more freedom and I'll demonstrate what that would look like and as you can see with the much higher denoising strength much more freedom essentially we're giving to the AI that the image is much different the faces even between these two images are different and most certainly different between the original the background imagery is different on both the clothing is different the necklace is different the hair is different and the crown is different but it's still kind of in the same vein of what we were headed for so let's say you were headed for more of a result closer to this and then you would click on this image and you'd say send to image to image you really like this a lot but you wanted to kind of change it to be something a little bit more maybe you wanted to go back to the diamonds and we wanted to add in um that she has uh long eyelashes let's put that in and then we'll give it the same amount of freedom and let it rip and great now we have two more images to kind of compare and see what exactly it is that we want um I'm gonna go ahead and choose this one on the left here and we're going to send this to extras where we will talk about our final tip tip number five which is just simplify your upscaling my tip and this you know mileage may vary on this don't take it for what it for everything that it's worth is I use swin IR four times and I resize by four times by eight times whatever I want just one upscaler that's all you need set it to four click generate Let It Go save it download it upload it do whatever you want and just leave it alone as you can see we have a great upscale result there this image isn't exactly perfect I wouldn't release this anywhere but just for example's sake as you can see the upscaling it's much higher resolution than it was before and then if you want you can click Send to extras again and click generate and it'll upscale it another four times if that's what you want all right I think that wraps it up for this video I just wanted to say thank you for the support on the last two videos especially on the negative prompts video it's been mind-blowing Thanks for the new subs and thanks for the views I appreciate it a lot if you enjoyed this and found it useful go ahead and check those other videos out like subscribe I read all the comments so leave a comment if you have any questions and I will see you guys in the next one thanks so much for watching

Channel: Binks

Views: 25,836

Keywords: ai, stable diffusion, midjourney, how to use ai, ai tutorial, ai art, ai art tutorial, protogen, photorealism ai, free ai art, chatgpt, machine learning art, how to make nsfw art, ai nsfw, twitter art, civitai, lexica art, stable diffusion better models, better ai, improve ai art

Id: E96EtKitug8

Length: 11min 20sec (680 seconds)

Published: Sat Jan 21 2023