Revealing my Workflow to Perfect AI Images.

Video Statistics and Information

Video
Captions Word Cloud
Reddit Comments
Captions
hey hey I'm going to show you my Advanced workflow in stable diffusion something that I've learned with a lot of experimentation and a Little Help from My Friends healing Maui and everyone else in my Discord I'm going to teach you how you can create amazing images in stable Fusion as well with more control and better quality done well almost anyone else in any application so let's dive straight in and I'll show you as well now I had a hard time getting the kids to sleep before recording this and I don't know what to do are they supposed to go to jail for assisting arrest alright so I'm going to be showing you my workflow with control net and doing multiple people and just doing Advanced images and getting great results in general so I'm prepared to control net input which is an image from the series The Last of Us so I'm not going to spoil anything don't worry I'm going to take this I'm going to input into control net down here now if you don't know how to install control net check my previous video where I go through it in detail I also put a link in the description below for that so I drag that in here and I made sure that this image is 512 by 768 so I'm going to do the same here I'm going to same here or actually this needs to be the other way so we have with 768 and enable and I want to use the Canon actually what I do is I mix between the Canon depth and I try what's best for the particular image that I have and it's always a little different I usually start at weight one and might go lower or higher depending on the results and I start hashing out lots of images with Euler a at about 25 and then I move on to SD Karis but let's start here and then I'm currently using the protogen model and I'm gonna load a preset that I have this is a friend of mine named Healy who has written this presets so I'm just gonna add in here we want I think I'm gonna remake this into a sort of a Blade Runner futuristic sci-fi thing so I written here man woman and futuristic Blade Runner sci-fi neon City I'm gonna up the weight and do this by holding Ctrl and arrow up then I'm gonna just try one image at first foreign and I'm fairly happy with the settings don't mind the faces for now we're gonna fix those later so just uh we just want like a general composition so now that I know that the settings are all right I'm gonna put it to like here like nine images and then I'm going to generate I'm going to speed this part up and now we have a lot of images to choose from the people are generally the same what I'm looking for is the composition in the background and if we have an interesting image because I want to use the colors to keep working with this in image to image so I'm thinking either the top or the first one here is pretty good it's a lot of colors up there now This Little Light here is messing up the backpack but it's not a big concern I also like what's going on down here because of the big backlight going on there and also this with the purple is very nice but I'm gonna try this one down here so take this and then you're going to send that to image to image that will save all your settings and now we're going to change to SD cards we're going to up this to about 40 steps and we're going to increase this now what I prefer to do is use an aspect ratio calculator because if you just you need to have the same aspect ratio when you well we don't need to but the image will be a little better if you do now I have 512 by 768 which is an aspect well the other way around actually 768 by 512 so we have an aspect ratio of three to two we want to double this so this would be 1024 and then we get the other number which is 1536 so we're gonna up this to 1024 and this two 15 36. we're gonna leave restore faces styling everything I'm going to lower the denoising strength to about 0.6 and I want to run this a couple of batches well first let's run one and see if we need to use control let now this turned out pretty well even the faces are all right actually and what I usually do is I do this with a large number of batches maybe three four at the time and then keep going until I find something I want so I'm gonna do that actually just for the sake of this video even though I'm fairly happy with the results here so we're going to change this to four and I'm actually going to lower a little bit because I had a very long time rendering this and as I'm recording this at the same time my computer didn't really like that it's gonna go to 832 and I'm going to check back here 832 that would mean 12 48 for the other one now we don't have exactly that so I'm just gonna stick to 1216. and let's run this again and the progress rendering is much better now even though I am recording so now I can actually keep the recording going as it's rendering now with every new image that you get here from the four that we wanted you'll get some variations and that's what we're looking for we want the variations just to find an image that we like so now we're finished and we have four images to choose from as you can see there are some variations in each image the backpack here changes a lot this doesn't seem to be a backpack at all and this uh well I'm not sure we have some tech on the arm here and here which I think is is very cool and I personally like what's going on with this image so I'm gonna stick with it even though the backpack is a little wonky here it's open so I mean we could keep going do more batches but I'm fairly happy with this now the reason that we're doing a larger resolution here in the image to image step and not in the beginning is that it takes such a long time for the first step that we did previously which was the text to image here was a much smaller resolution in Euler a and that way you can hash out a lot of images quickly and find the composition then when you've found the composition you move on and do what we did here and improve on the m image now we're going to improve even further and in this step we're going to in paint now the image that we have here is fairly good already but you can see here in the eyes that is not perfect and the same here so what we do first is we're going to send this to extras and if you want to you can upscale four times but I'm gonna stick the two for now I'm not gonna bother with any of the upscalers but um particular one that you prefer I know some people with that use this style a lot use the Lang sauce I've used ESR Channel a lot previously for paintings swing IR for example is very good with intricate details but can be very intense you can then get two too sharp you know that let's upscale this then we're gonna send it to in paint and now we're going to improve on the parts that we need to fix which would be well actually eyes here we're gonna impaint the whole face just to get a higher resolution on it take this part we're just going to say man face and of course you can make a much more detailed prompt if you're looking for something specific in the face and now we're in painting here it's important for this step if you're doing what I'm doing you want you know this is a face on top of a face I just want it better then I'm using original then it takes the value that's below if you take Latin noise for example that's good if you want to replace something because then you get new noise that stable Fusion can work with but for now we're going to use original and we're going to in paint at Full Resolution and we're going to keep the same 40 steps with Keras and we're just going to leave everything here except the batch account I'm going to lower that to two and the denoising strength it depends how much you want the image changed so if you got an image that you're fairly happy with already you could keep this lower I want the face changed quite a bit so I'm going to keep it to 0.6 but I don't want a completely new new thing there so let's generate this you can see you get a close up here and that's because we checked in the Box here in painted full resolution and instantly before it's even finished you can see that eyes are much better and the face is generally more detailed than what we had in the beginning and you can actually see it compared with the ear here and the hair which is still low rest in comparison now you could repaint the whole head that way you just paint over the hair here and there as well let's check and see what we got so this is the first one and this is the second one and I'm gonna go here with the first one so I'm just gonna take this I'm gonna drag it to the left I'm gonna press the little thing here to remove our mask and then we're going to continue with the woman here or which is actually the child in the original it doesn't look like a child anymore it's just a short short person so I'm thinking we could try to make this into a robot I'm going to change the prompt to woman android robot face and I'm gonna up the strength a little bit because we want some more changes now and we can see that we have a robot face emerging here and the detail and resolution is just stunning just imagine that this is a small part of this huge image now let's see the results the first one here and the second one the second one is more resembling to the original human has some hair but still some mechanical details in the back and this is a full-on robot we're gonna take the robot here and next up I'm thinking we're going to change the little displays or controls here on his arm and again we're going to drop this here press this to remove the mask we're gonna draw here let's try something like display control unit now I don't know what that is but uh we'll find out and we don't want a lot of denoising here we still want to retain it's going on we basically just want to improve on what's here most of the stuff is here so it's just gonna basically upscale make it better I think this one has some great potential I like that the letters and letters here the n and y now they are almost perfect now they don't really mean anything but for the sake of the image maybe it's the name of the control unit or the name of the company it doesn't matter so let's take this um see yeah I think this is super cool and here you can see our final image now what we're seeing here is a large image and this is just at 59 of its original size now the actual size of this is about 2 400 pixels by 1600 pixels so we can see if we zoom in here to 100 percent it's a huge image and you can see where we didn't paint it the detail here is amazing as well as here I'm the robot android face here now you can keep adding and in painting as much as you want I think this is a great way for you to get great quality and large images a lot of people just try to use a large resolution from the start and that's not a very effective workflow I think doing this together with control nuts will yield you the absolute best pictures available as of right now now this might change tomorrow you know how it is with AI art and the progress I hope you learned something and you can use this in your workflow and thanks to Healy and Maui for helping out with the channel and images in general if you enjoy this content like And subscribe if you want to I have lots of videos um tutorials on AI art as always have a good one see ya
Info
Channel: Sebastian Kamph
Views: 312,310
Rating: undefined out of 5
Keywords:
Id: 4u-Ytioi3DM
Channel Id: undefined
Length: 13min 31sec (811 seconds)
Published: Sat Feb 18 2023
Related Videos
Note
Please note that this website is currently a work in progress! Lots of interesting data and statistics to come.