00:00Hello and welcome back to another Reality Check VR video

00:01That's right, this is the ultimate ComfyUI guide at the moment, with image-to-image refinement, upscaling, sharpening, and even Laura's built-in tools as well

00:12So, we've got all these little additions, and hopefully you're going to enjoy it

00:15I spent the entire night getting it working the way I would want it to work for my own workflow, so hopefully it's something you can use advantageously for yourself

00:23And if not, you can even adapt, modify, or change it if you need to in any way

00:29So, I'm going to go ahead and jump right in, and we're going to see exactly what I'm talking about right now because it's a beauty, it's a masterpiece

00:36That's right, no, it's… Kind of messy, and you can certainly move these things around as you desire, but I put it in a way that I felt was the easiest for me at the moment, for my workflow

00:48And I don't know about you guys, but I'm finding ComfyUI not only to be fast and highly effective, but also fun

00:57The node-based style is really something that I'm digging

01:00Even though I find Automatic1111 to be more useful for certain things, like ControlNet and other tools such as DreamBooth, I like ComfyUI so far with what it has

01:09I want to learn how to do more with it, so as I continue to learn, I'm going to make more and more tutorials, and hopefully you guys can learn some more with me

01:16 But all right, so right here in front of us, as you can see, we have down here on the..

01:19Below, we have the image-to-image section

01:21This is where you can drag and drop any of your images, but we also have a text-to-image section over here, so if you want to do text-to-image, that will work as well

01:29Let me go ahead and move my camera over there

01:31What we do is, right now, it's initially set up for image-to-image, but I'll show you how to switch it to text-to-image

01:39To get everything working, we have a few simple configurations, and I'll show you exactly how to do that right now

01:45On this page, you put the image in the image-to-image section, and it will generate your base image here

01:50Above it, it will then produce a refined image, and on the right, it will give you a sharpened and..

01:55Upscaled image, and that's it

01:59So, uh, I'm just amazed with how good this looks

02:03I mean, come on, are you kidding me? This right here from SDXL 1.0, it's beautiful

02:10I mean, you can even zoom right in and see some of those textures, which obviously look better than others, but that's phenomenal right there

02:18The fact that we can get this level of clarity from a very quick image is impressive

02:22It took my computer, using an old graphics card, about a minute and a half, or maybe a minute, to render this

02:29And after the initial render, it just gets faster and faster

02:33One thing I really like about ComfyUI is the fact that if you are doing something like keeping..

02:39The same seed, and you're just changing something, like, for example, if I'm just changing the upscale, and I keep everything else the same, every time I hit that Q button, it doesn't have to do everything else over again

02:48Sometimes it will only take four or five seconds to redo those last changes, and then the upscale image will sometimes only take a few seconds, literally one, two, three seconds, as opposed to a minute or two

02:59So I find this program to be very, very fast, but as I said, there are a few configurations to get going

03:06So let's go ahead and get those, and just to show you right here, this is where I started

03:11This is the first image I created, and I actually created this in Stable Diffusion yesterday, and then I..

03:17I brought it over into SDXL, and I was able to get images like this

03:21You know, from this image to image here in ComfyUI to get images like this, which are just astounding

03:28So I mean, the level of detail right there that you can see in this character, the kind of stylization that we've got going on here, obviously it's very artistic

03:38It doesn't look real, but it's just so cool what we're able to achieve very quickly

03:44And I believe this one was kind of a denoise of 0.5, you know, to get something like this, which was just—like I said—just amazing

03:50So, all right, let's go ahead and get our configurations going

03:55The one thing that you're going to want to go ahead and get for your ComfyUI is, uh, this thing called Duffy, and that is the Comfy

04:05UI for few math and Mata nodes, and the reason is because my configuration

04:08One thing that we do is, when we're using image to image, we don't want to have to do any math, and the one thing about Durfu is it does it for you

04:18So all you have to do is just go right here and click this download button

04:21Once it downloads, it will go into your Downloads folder

04:24Let me go ahead and take a look at that right here

04:28So here we have it, and all you need to do is take this and add it into your ComfyUI folder, into the 'custom nodes' folder

04:45Right here, you can see I've got it in the ComfyUI custom nodes folder, called 'ComfyUI Durfu' or the 'comfy UI modded nodes.' So once you just drag and drop it right in there, you might need to restart

04:54Restart the program, and after you do that, it will be able to work

04:58So, make sure you get this file right here, and that will be the one thing you need to ensure this configuration is going to work on your computer, okay? And while you're in the ComfyUI folder, there's one other thing we can do, and that is get all of our models to sync up

05:11If you already have them downloaded—for example, for Automatic1111 or any other program that uses a Stable Diffusion-type checkpoint model—you can actually use them without having to duplicate them in different folders

05:26All you have to do is go into your ComfyUI folder; you should already have a file that says extra-model-paths.yaml.example

05:33If you do not have this file right here, don't… Worry, it should have downloaded, but if it didn't, if you go right over here to GitHub and scroll down, you'll see ComfyUI

05:44Not only is it right here and you can download it directly, but if you go down to the config file section and click on that, it will give you the ability to just go over here

05:55Right here, you'll click the download button, and then you'll have the file downloaded

06:00For those of you who want to change your models path, you need to right-click this file here and select "Open with," then choose Notepad and click OK

06:11Once it opens, you'll see at the top it says "Your base path." All you have to do is change the base path to where it is installed on your system

06:20It should have one more, one in there, base path, and then right here

06:22So, that spot is where we're going to have to replace the information

06:27And so, if I go into my A1 Master down here, into the Stable Diffusion folder, I can either go up here into the title bar and click Ctrl+C to get it that way, or I can go back, shift right-click, and hit "Copy as path." Either way, I go over here, and I'm just going to paste it

06:46I'm going to get rid of the parentheses, and I'm going to go ahead and hit "Save all." All right, and now that's it

06:54Because I’ve got that working in there, it's going to automatically pull my models and things from that other folder and get them exactly where I need them to be

06:58So, perfect

07:03Now that we have the Defurfu installed, these modded nodes..

07:07Once installed, we can go right over to our area

07:10Of course, all the pictures that I'm going to be showing you guys are going to be in the description, and all you have to do is actually drag and drop these pictures into this area

07:19So, let's go ahead and clear everything up real quick

07:22I'm going to go ahead and hit "Clear." Yeah, clear it all up

07:25 If we go into the ComfyUI GitHub, we can see they have a spot here that says "ComfyUI examples," and if I click on that, they have all kinds of crazy things

07:36And when I say crazy things, I mean crazy things

07:39The ComfyUI basic tutorial, and I thought, hey, I should get the basic tutorial, right? And I didn't know what I was getting myself into, or a kid in class, and your TTSO, you go through all the Stable Diffusion

07:50They show you a bunch of tutorials, of course, my camera's in the way, and they display screens like this, which is great because they give you examples of how to basically build the components themselves

08:01But you don't need to look right in there

08:02For example, I wanted to do image-to-image, so I clicked here on image-to-image, and very simply, it shows you how to do it

08:10If you just zoom in on that, you can get a really close look, and you can start to see the kinds of things you'll need

08:16You need to load a checkpoint, you need to have a CLIP model, you need to have a K_sampler, a decode step, and a spot where the image outputs

08:23You also need to be able to load the input image

08:26So, all right, we've got the… Building blocks right there, not bad, not bad

08:28All right, if it looks pretty good

08:30Okay, I totally forgot that I made these ones, but yes, I was testing, and as you can see, it goes from certain styles to getting more and more similar to completely different ones when you change the denoise to almost the same, or when you put the denoise back to a closer strength

08:46Oh yeah, some people mentioned this in chat last time as well, but the SDXL spot right here has two images, and if you actually just copy these images or save these images to your hard drive, you can just drag and drop them into here

09:01For example, here's the one bottle; if I drag and drop it, we instantly have a refiner model with the ability to refine the prompt

09:08Separately from the base prompt, nice, right? And I actually use this for inspiration on how to do my step control, so it was very helpful for people to see this one

09:23If I drag the other one in here, you can see it's just a little bit different from the other bottle

09:26So, anyway, I'm going to go ahead and drag my first, uh, my first creation of the day from scratch, and that was this one right here

09:34What this is is just a basic spot where we can take a prompt, a beautiful portrait of a shining, golden AI robot

09:39And with this one, it does not have a refiner; it just has a spot where you can click denoise and then queue it

09:52So, let's go ahead and give this one a try

09:54That is, uh, that is one good-looking..

09:55Result refined Text: Even without the refined result, this image looks stunning

10:02I mean, it looks really, really good

10:04So, this workflow, just by itself, for speed and efficiency, and just for what it is, I really like it

10:09I'm going to go ahead and put this one in the description as one of the main workflow results

10:14As I mentioned, we're going to look at a few of these other ones as well

10:20Oh, you can see I was testing some stuff; I started dragging these other pictures, and you can see how my workflow changed a lot over time

10:27Yeah, I started changing it quite a bit

10:33All right, I can tell I'm getting a little too tired, so I need to hurry up and finish this video, and later I can make a more in-depth explanation

10:40For anybody who wants, and I'll use all the comments in this video to help me know exactly what I should include in that more in-depth video, and I'll put that out right away as soon as possible

10:49Also, be sure to check out—I've got a Patreon link in the bottom

10:53I'm just now starting my channel back up, and I want to try to get things going as quickly as possible

10:57The more support I get, the more time I can dedicate to these videos

10:58So, I really appreciate anybody who wants to help support the channel

11:04Anyway, enough spiel—let's go ahead and show you guys exactly what we've got going right in here so that you can understand it as best as possible

11:12Over here, in these purple little boxes on the left, we have, obviously, the best sizes as is

11:16In that other node, I just copy and pasted this over, and then down here, I put: if you want to change the style, you can do that right here

11:25To have image-to-image, make sure the latent from the VA encode is set to 'latent image.' VA encode latent is set to 'latent image.' See, that's what it's talking about—these two things right there

11:36And then, if you want to do text-to-image, make sure that latent from text-to-image is set to 'latent image.' So, okay, for text-to-image, if I take this latent and I put it on there—well, now it's a text-to-image

11:41Simple as that

11:45And I can create the height and width right there

11:50Then, down here, it says for image-to-image, start steps will determine the denoise strength, and this is kind of the important thing right here

11:58What we have is over here, the starting steps

12:00When you are doing an image-to-image in ComfyUI, you want to make sure you tell it how close to the original image you want it to be

12:09Do you want to keep 20% of the image, 10%, 1%, or 99% of the image? The value is right down here, and that's going to be with the overall number of steps, divided by that value—or actually, more so, that value divided by the total number of steps

12:30That will then give you your denoise strength

12:33 For example, if you have 10 steps with a total of 20 steps overall, that's the same as a denoise strength of 0.5

12:40Also, starting at 5 steps with a total of 20 steps is the same as 0.75 denoise

12:44Denoise, and then alternatively, starting at 15 steps out of 20 steps, is the same as 0.25 denoise

12:53The last little bit down here contains all the variables that you can change to create different kinds of images

12:58We have the base checkpoint and the refiner checkpoint, which are right here

13:02The base checkpoint and the refiner checkpoint are listed right in these spots

13:06As we mentioned before, you can actually see how all the different models show up here because we changed that configuration file earlier

13:13If I go back down here, we have the upscale model over here

13:16We do actually have the ability to change the upscale model right then

13:20Then we have the positive prompt and the negative prompt, which is pretty simple and goes without saying

13:23Prompt, negative prompt

13:26Right there, it says, uh, we have the Laura strength over here

13:29You can choose your Laura and your Laura strength

13:31Right now, I have the Laura strength at 0.5, so it's about half strength

13:36Then we have our steps, CFG, and our sampler, and right here we can see we've got our steps, our CFG, and our sampler, and we've got those for not just the base but the refiner image as well, so we've got it for both of them

13:51Then you have your seed, which obviously is going to be randomized

13:55If you want it to stay fixed, you just go in here and click fixed

13:58Lastly, initial size, upscale size—you can actually choose that

14:02Right now, it's just four times, and then sharpen amount

14:10If you want to change your sharpen amount, you go in here and..

14:12You can change your sharpen radius, your Sigma, and your Alpha right now

14:14They're both set to 0.5, and then your blend amount

14:17What it does is it creates a sharpened image, and then it has the original image that’s upscaled, and it takes these two images and blends them together

14:24You can tell how much you blend it, and right now, they're blended at 0.5 percent, or a half

14:29Then they are doing a normal blend as opposed to a multiply, screen, overlay, or light blend

14:34So, that's how those are working

14:36Okay, well, as you can see, we set it to text-to-image earlier, and because we set it to text-to-image and we have this prompt: "a beautiful, futuristic, shiny metal AI robot," it created a beautiful AI robot

14:45And this right here is the refined image, and then… Over here, we even have a really sharpened image, which actually looks much worse

14:53So, in this case, when things are small like that, sometimes the sharpening does not look great, which is why we have the different options available

15:01So, what we're going to do is take this refined image right here, and we're going to put it over here in our image-to-image section

15:06We're just going to drag and drop it there

15:09Now that it's there, I'm going to go ahead and switch this back to image-to-image by just dragging this latent over to the latent space, simple as that—one click to do it

15:16Then, what I do is go into the prompt, and I'm going to change the prompt now

15:23Let's say, okay, I'm going to put a photo of Elvis Presley

15:24Standing at Madison Square Garden for a live concert, realistic

15:29I'm not going to go into anymore detail

15:31For the negative prompt, we've got "ugly" and "deformed," and of course, we've got our Laura set up

15:34We're going to keep that at 0.5

15:37We're going to keep all this other stuff pretty much the same

15:41I don't think there's any need to change anything right now

15:43Let's say the start steps—should we keep this at 0.5? One thing I didn't mention is, as you can see, it says "end at step 20." We actually have 25 steps in total, and the reason is that wherever you're going to put the end step, like for example 20 steps, are going to be done on the base model, and the last five will be done on the refiner model

16:06So if I change this step right here to a higher value..

16:10Let’s say 30 or 40, and then I keep this at 20

16:12Well, if it's 40, then there would be 20 for the base model and 20 for the refined model

16:18Alternatively, if I set this to 21, then there would be 20 on the base model and just one on the refined model, regarding the steps

16:29And obviously, where we start the steps is going to determine how closely it will resemble our original image

16:35If we wait a long time—if we wait until step 10, 12, 13, 14, 15, 16—if we wait that long, the base image is going to look a lot like our original image

16:47If we chop it up or start changing it right at the beginning, it will change exactly the other way

16:52So right now, I’m going to keep this at 25, and I’m going to… To set this start step value, I'm actually going to change this down here to 4

17:03I'm going to give it a lot more room, so as you can see right here, if there's, as far as the base model is concerned, on the base model we have 20 steps, even though up here it says 25 steps

17:14The base model is only doing 20 steps

17:16So, for the base models, as far as it's concerned, we're going to be starting four steps in

17:21That gives us 16 steps where we're not even using the original image, or we're just going to be using our freedom

17:28Right, we have to think about it like that

17:30So, the more steps you give it, with freedom, 16 steps of freedom

17:34Right now, we're going to go ahead and see a different image

17:36So, I'm going to go ahead and click Q prompt, and let's see what this makes for us

17:39Right there, we go

17:41So, as you can see, we've got the image-to-image working really, really well

17:47The original with, right, came from a text-to-image here in SDXL, and then right over there, we have the output for the original base

17:53Above that, the refined version, which gets a little bit more bedazzled, and then over here, a little bit sharpened

17:59Of course, if you want to do more touch-ups, you could just do some inpainting there on the face to make it even more high definition

18:06If you feel, I did one other version with this guy right here, and it turned out to be like that

18:10So I actually wanted to make it a little different, but I still actually kind of like the way this one is, even stylized a bit, even though the face isn't there at all

18:18This looks very, very good

18:26Cool, and there's just a lot that can be done with this, so hopefully I can inspire you guys to start playing with it, make even better workflows, and then share them back with me in the comments so that I can learn from them and use them myself

18:33I'm going to continue to make more and more of these, so stay tuned to the channel

18:37Also, be sure to support if you can; I’ll be posting more tutorial videos and special files there for anyone who wants to help support the channel

18:46And lastly, I think some of the coolest things are these little AI robots, especially how well they look with some of the simplest prompts I’ve been putting in

18:58I just wanted to share some of these images with you

19:01What this is right here is actually without the Laura, and then when I added the Laura to it, it changed more like this

19:07So this is no Laura, and then Laura

19:09This right here is like 50 Laura, and then I started changing the samplers around, and you can see just how completely different they really are

19:18With these different ones, I had so much fun seeing the various intricacies and how much they could change based on the denoise level

19:29So, that really is the thing you're going to want to experiment with a lot

19:32When I say the denoise level, I’m talking about the start steps value right here

19:35You know, I usually keep it the same myself

19:36Right around the 4 to 10 value, uh, you know, if I want to make it very similar to the same thing, I'll put it up to 15, but I usually don't go past 4 to 15; even that's kind of the range that I use

19:49But I'm going to go ahead and keep it at eight, just for results like this, where you can see an image that stays

19:55You've got the same feet placement for the most part; the hand completely changed from where it was

19:59But for the most part, you've got the same structure of the scene, even though there is no background in there

20:05But I did make sure in the prompt to say "concert." So, yeah, that's just one of those things we're going to go ahead and keep learning right here

20:11But as I said, as you guys continue to learn and find out more things, please feel free to share

20:14Leave your comments with me, and I'm going to go ahead and keep updating you guys as much as I can

20:19Oh yeah, I also did the UFO one, where I made UFOs last time

20:23I took one image and did that—did a whole bunch of my stylization

20:26The quality is pretty good on these things, almost too good, which makes it take a long time to load

20:30There we go

20:32All right, so I'm going to go ahead and end the video right there, and I hope you guys got some value from this

20:37I'm going to keep playing with this and learning what I can

20:41Please feel free to send me any information that can help me continue my growth as well

20:44Hope you guys have a wonderful day, a wonderful night, and I'll talk to you soon

20:50Bye.