Generate AI Images Without Paying A Dime | Stable Diffusion Tutorial

[00:00] Generative AI has opened the door to

[00:02] allowing anyone to create incredible

[00:04] images just using a computer and a few

[00:07] text prompts. Now, the platform that

[00:08] gets the least amount of attention is

[00:10] Stable Diffusion, but it might just be

[00:12] the hidden gem you're looking for.

[00:14] Unlike other major generative AI

[00:16] platforms like Adobe's Firefly or

[00:18] OpenAI's Dali or even Midjourney, Stable

[00:20] Diffusion is completely open source,

[00:23] which in short is free. So, I'd want

[00:25] Padme to wear her iconic [music] white

[00:27] turtle attire. See that image is coming

[00:29] in. It's looking great. There's a few

[00:31] issues going on in the lower parts of

[00:32] the frame. In this particular section,

[00:34] I'm changing how the body structure

[00:36] looks. Again, let's increase that to

[00:38] four images. Okay, we're going to pause

[00:40] it right there. Remember, there are no

[00:42] limitations in the software, which means

[00:44] the responsibility is on you.

[00:49] Stable diffusion lets you run on your

[00:51] own GPU instead of on the cloud, which

[00:53] allows you to tweak every minute setting

[00:56] that you can think of. Now, if that's

[00:57] not enough to convince you to learn this

[00:59] software, it's also the only platform

[01:01] with zero limitations, but it also comes

[01:04] with its own drawbacks. It's tough to

[01:06] get started with, but that's where we

[01:08] come in. Today, we're going to help you

[01:09] with everything from getting started,

[01:11] installation to even more complicated

[01:13] things like perfecting your prompts and

[01:15] getting the best possible image. Now,

[01:17] today's episode is a little bit

[01:19] different. It's just you, me, and a

[01:21] computer. And we're going to get through

[01:22] this together. It's also the first time

[01:24] that we haven't had someone behind the

[01:26] camera, so expect things to go wrong.

[01:30] First off, what is stable diffusion? In

[01:32] short, it's an image generation model

[01:34] created by stable AI. It turns text

[01:37] prompts into images. What makes stable

[01:40] diffusion really special is that it's

[01:42] open source, which in short means that

[01:44] it's free, but it also means that you

[01:46] can download it, bring it into your own

[01:48] computer, and customize it to your own

[01:50] art style.

[01:53] Now, in a majority of this video, we

[01:54] will work with a software application

[01:56] called Fucus. This is a front-end

[01:59] interface that allows us to interact

[02:01] with the stable diffusion model. Now,

[02:03] don't let any of those words scare you

[02:05] off. Think of stable diffusion as the

[02:07] engine, the code that's doing all the AI

[02:09] image processing. Focus is like the car

[02:12] that's built around the engine. It's

[02:14] what has all the buttons and the

[02:16] controls and the interface that allows

[02:18] you to interact with the engine without

[02:20] really messing around with the tech.

[02:22] Okay, enough talking. Let's jump into

[02:24] the application. The first thing you

[02:25] want to do is go to the GitHub

[02:26] repository for Focus. We'll leave a link

[02:28] in the description. After you download

[02:30] the file, you want to extract the file

[02:31] and you'll see and run.bat. You want to

[02:33] click on the run.bat file and start the

[02:36] installation process. Remember, the

[02:37] download file will be around 50 GB. So,

[02:40] make sure you have that capacity before

[02:42] you get started. By default, you'll get

[02:44] the standard model, but you'll also get

[02:45] the option to run the animate and

[02:47] realistic models as well. For right now,

[02:49] we're not going to change any of the

[02:50] default behavior of the software, which

[02:52] means as it launches, it'll look for new

[02:54] models, keep the software updated, which

[02:56] is generally what you want. But if you

[02:58] want to change that, you can use these

[03:00] two command lines. Let's do realistic

[03:02] right now. And now we just let it do its

[03:03] thing. Another quick tip that I like to

[03:05] use is keeping task manager open. As

[03:08] long as I see activity on the GPU, I

[03:10] know that the software is working with

[03:12] the GPU to get it collected.

[03:15] Awesome. And we are finally ready to

[03:16] start generating images. You'll notice a

[03:18] few things. The image generation window

[03:20] at the top, the text prompt window at

[03:21] the bottom where we can type things like

[03:23] cats on a window lid. And at the bottom,

[03:25] you can see input image. We'll go into

[03:27] that in detail. It's something that I

[03:29] use in great detail. Pain enhance, which

[03:31] we won't talk about too much.

[03:32] essentially post-processing steps that

[03:34] you can take on your final image to

[03:36] increase the resolution and detail. As

[03:38] we hit generate, you can see that there

[03:40] is an immediate spike in your GPU. So,

[03:42] you'll notice the image comes in a

[03:44] little fuzzy, then it cleans up over

[03:46] time. Okay, with those images in place,

[03:47] we can look at them in detail by

[03:49] clicking on each individual image. The

[03:50] second image is a really good example of

[03:52] the kind of problems that this kind of

[03:54] model has. You can see that the eye on

[03:55] the left is perfect, but the eye on the

[03:57] right just doesn't have the detail we

[03:58] need. So, there's two things that we can

[04:00] think about at this point. We can click

[04:01] on advanced and we can first look at the

[04:03] presets. We are using the realistic

[04:05] model which is what generates such fine

[04:07] detail but there's lots of other

[04:08] variations that you can try out here.

[04:10] There's also performance depending on if

[04:12] you want to take a qualitative or

[04:13] quantitative approach. Right now we're

[04:16] prioritizing speed of delivery, but we

[04:18] can also change that to quality to

[04:20] generate higher quality images. Now a

[04:22] feature that I use all the time is

[04:24] changing the number of images, but for

[04:26] right now let's do four images.

[04:29] >> [music]

[04:31] >> That's pretty incredible. The first

[04:32] three images have come in. I love the

[04:34] detail here. There's lots of detail in

[04:36] the hair and in the eyes. Both eyes look

[04:38] perfect. The second image is not quite

[04:40] as good. It has really great

[04:41] perspective. Lots of detail in the

[04:43] bricks and the window, but not quite as

[04:45] much detail on the cat itself. You can

[04:47] see that the eye on the right isn't

[04:48] perfect. Okay, this is great perspective

[04:50] cuz there's a beautiful window. It's a

[04:52] low angle shot. Again, not quite as much

[04:54] detail on the cat itself. And the final

[04:56] image has also come through. Oh, this is

[04:57] amazing. And you can see the cat looking

[04:59] through the window. Lots of detailing on

[05:01] the glass and the light and the texture

[05:03] on the skin. We're going to build a

[05:04] crazy Star Wars poster no one has ever

[05:06] seen of Padme fighting off Anakin

[05:09] Skywalker. But first, let's build a

[05:11] cyberpunk city with crazy [music]

[05:12] detail.

[05:14] So, you can notice the kind of detail

[05:16] that I'm putting in here. I'm I'm

[05:17] writing out very specifically what I

[05:19] want to see. So, a cyberpunk city at

[05:21] night. So, I'm specifying what the

[05:23] lighting looks like. Glowing neon

[05:25] lights. someone have that look where

[05:26] there's lots of light from the buildings

[05:27] themselves. And I'm specifically calling

[05:29] for cinematic lighting. With AI models,

[05:32] the more specific details you can give,

[05:34] the more likely you are to get a

[05:35] successful final image. Okay, so those

[05:38] two images have come through. They're

[05:39] both fairly similar, which is something

[05:40] I don't really like. Ideally, I want

[05:42] variation on shot. Also, something I

[05:44] like to do is watch the model as the

[05:46] image generates just to see if it's

[05:48] something that's in alignment with what

[05:50] I want. If it's completely off, I can

[05:51] choose to skip through that image. Okay.

[05:54] And that's pretty great. Everything's

[05:55] generally looking okay, but nothing is

[05:57] specifically looking good. That's also

[05:59] another problem with the AI generation

[06:00] model for wide and detailed shots. It's

[06:03] harder to get fine details correct when

[06:05] there's lots of details spread across

[06:07] the image. Let's say we love this image

[06:08] in general, but we want to enhance

[06:10] certain details. With that in mind,

[06:12] let's refine our model some more. Okay,

[06:14] now we've got two new images. Let's look

[06:16] at both of them. I love the detailing on

[06:17] this. Generally, everything seems okay.

[06:19] There's lots of lost details in the

[06:21] buildings in the background, but I think

[06:22] I can generally fix those as we go. The

[06:24] second image is a lot softer and there

[06:27] is a lot of billboards and written

[06:29] detail which [music] I probably won't be

[06:31] able to fix. What if we have a great

[06:32] image, but we want to refine that image.

[06:34] That's where image input comes in.

[06:36] First, if you want to take this image

[06:38] and we just want to scale it up, bring

[06:39] in new resolution, we can use image

[06:41] upscale. So, you just drag that image in

[06:43] here. First, we look at the bottom row.

[06:45] Upscaling by 1.5, upscaling by 2x. that

[06:48] essentially just expands the resolution,

[06:51] maintaining as much of the image as

[06:52] possible. And then you have upscaling

[06:54] fast 2x, which is the same thing, but it

[06:56] processes it a little faster with a

[06:58] little less accuracy. So, first let's do

[06:59] a 1.5 upscale. Now, as that image comes

[07:02] in, we can see we already have a lot

[07:04] more detail. Okay, the first image is

[07:06] in. Let's take a look at what that looks

[07:07] like. So, you're already seeing so much

[07:09] more detail in the building in the

[07:10] foreground. All the problems of the

[07:12] background have now been resolved. All

[07:14] the lines are straight. All the windows

[07:15] are visible. Generally, everything's

[07:17] looking good. There is still a little

[07:19] bit of specific problems I'm seeing,

[07:21] especially on this billboard on the

[07:22] right, parking lot down below, but the

[07:24] cars don't look perfect. We can work on

[07:26] those specifically. Okay, let's look at

[07:28] this image with some detail. That's not

[07:30] going to work cuz it's so close to the

[07:32] foreground element, which is this

[07:33] building. So, to refine this, we're

[07:34] going to use impaint. This allows you to

[07:36] refine and work on specific details of

[07:38] your image with great detail. There's a

[07:40] brush, and anything you brush over will

[07:42] then be changed. Anything that's outside

[07:45] the brush radius will not be affected.

[07:47] So, we've got three options here.

[07:49] Impaint, improve, and modify. Impaint is

[07:51] a great technique if you want to change

[07:53] something of your image with subtle

[07:55] variation and keep the general look of

[07:56] the image the same. The first thing

[07:58] you'll notice is that the GPU is using

[07:59] all of its processing power on just that

[08:01] one zone of the image, which means

[08:03] you'll get a lot of resolution in just

[08:05] that one area. Has a little motel

[08:07] looking building with a swimming pool in

[08:09] front. It's lots of details in terms of

[08:11] cars and people, which may not be what

[08:13] we want because it's going to attract a

[08:14] lot of attention. And then the second is

[08:18] a black building with a few windows. You

[08:20] can see that I can actually drag the

[08:22] image from my image generation window

[08:24] into my impaint window. This is a pretty

[08:27] common technique of refining your image

[08:29] as you move it back and forth within the

[08:31] software itself. And this is a good time

[08:33] to talk about the other two impend

[08:35] features as well. Improve detail is used

[08:37] very often. This is when you want to

[08:39] increase the resolution of something in

[08:41] the background, right? We can see far

[08:42] more detail in that image. You can see

[08:44] the specific floors of the building. You

[08:47] can see through some of the windows. You

[08:48] can see some trees and shbery brought in

[08:50] front of the building. Really quickly,

[08:52] let's look at modify content as well.

[08:54] This is a powerful tool when you want to

[08:55] make a dramatic shift of your image and

[08:58] then later go back and refine it using

[09:01] the refineer tool. So, let's look at one

[09:02] of these. And that's what that car path

[09:04] looks like. Again, here you can see a

[09:05] lot of imperfections in the car, in the

[09:07] floor, in the building next to the car

[09:08] park. All of which will need to be

[09:10] refined if you want to use it in your

[09:12] final image. But for right now, we're

[09:14] just going to revert back to the image

[09:15] that we had and look through our final

[09:17] settings. At this point, you probably

[09:19] get the idea of image generation, but

[09:21] let's dive into something a little bit

[09:22] more obscure that uses more advanced

[09:25] features of the software.

[09:29] So, [music] let's think of something

[09:30] that we can't find on the internet.

[09:32] though. Anakin Skywalker and Padme

[09:35] lightsaber dual Star Wars franchise high

[09:37] detail cinematic lighting at night.

[09:39] [music] All right, let's see what this

[09:41] looks like. Again, let's increase that

[09:42] to four images. Okay, we're going to

[09:45] pause it right there. So, this is where

[09:47] you need to be careful as you generate

[09:49] images. You need to make sure that you

[09:50] don't generate any images that's not

[09:52] safe for work. And this is a good time

[09:54] as any to talk about the power of these

[09:56] creative tools. Remember, there are no

[09:58] limitations in these software, which

[10:00] means the responsibility is on you. Make

[10:02] sure you don't create any content that's

[10:03] harmful, misleading, or violates

[10:06] privacy. Another tool that we can use to

[10:07] safeguard our content is negative

[10:09] prompts. Now, these are things that you

[10:11] don't want in your image. So, by

[10:13] default, I'm getting things like

[10:14] unrealistic, saturated, big nose,

[10:16] painting, drawing, sketch. These are all

[10:19] prompts that have come in default by the

[10:21] software. But now, I'm going to also

[10:22] include not safe for work as a prompt.

[10:25] And you can expand on that list. Okay.

[10:27] And those images have come through. They

[10:29] both have their own challenges. So over

[10:31] here, both Anakin and Padme are sharing

[10:33] a lightsaber, which doesn't really make

[10:34] the most sense. In the second image, I

[10:36] kind of have their hands crossed over.

[10:38] Both which we can fix. And we can do

[10:40] that by bringing this image, dragging it

[10:42] in using impaint and correcting for. But

[10:45] we're not going to do that right now

[10:46] because there is a holistic problem in

[10:48] the image. And that's the fact that I

[10:50] don't like the perspective that we're

[10:52] getting. Ideally, I'd want to see

[10:53] something similar to the Star Wars

[10:54] poster. It's a low angle shot with lava

[10:57] in the background and each character

[10:58] fighting aggressively against each

[11:00] other. Now I could give this in the form

[11:03] of a prompt, but I can also use image

[11:06] prompt. Image prompt essentially allows

[11:08] you to input an image and generate a

[11:11] similar image. Now, but for right now,

[11:13] let's turn off all of our advanced

[11:15] features. You can go over to image

[11:17] prompt and you can drag your image in

[11:19] and you have a few features that will

[11:20] show up. Now, if you don't have this bar

[11:22] at the bottom, you just want to scroll

[11:23] down to the bottom and click advanced.

[11:26] That'll give you four options. Now,

[11:27] image prompt will generally scan the

[11:29] image very similar to a textbased

[11:31] prompt. It'll attempt to understand

[11:33] what's happening in the image and it'll

[11:34] generally use that as a suggestion for

[11:36] generation. You can see that it's

[11:37] generally taking the idea of the image

[11:39] and generating something similar. But in

[11:41] this case, what we really need is

[11:43] something that's visually similar. So,

[11:45] we have two real options. Pyrochni and

[11:48] CPDS. Pyrochani essentially maps the

[11:50] position of characters and takes it into

[11:53] your new image which would be ideal for

[11:56] this particular image and CPDS

[11:58] essentially uses contrast, color and

[12:00] saturation to generate a similar image.

[12:03] So in this particular case, Pyrocheni

[12:05] would work perfect. Okay, we've got four

[12:07] images in. Let's let the remaining

[12:08] images come in. Let's look at the first

[12:10] one. So this first image, the first

[12:11] thing you'll notice is that it's fairly

[12:13] low resolution. There's not a lot of

[12:14] detail in the face structure. The

[12:16] lightsaber's broken. I don't like how

[12:18] the sand looks relative to the mountain.

[12:20] The second image is much better. And

[12:21] then it's converted Obi-Wan into some

[12:24] version of Anakin. So, that's really

[12:26] great in terms of perspective, in terms

[12:27] of each lightsaber. Lightsabers are red,

[12:30] but I'm assuming I can make those

[12:31] changes. Third image. I don't like it as

[12:34] much. The resolution is not quite there.

[12:36] The perspective isn't great. Anakin's

[12:38] looking a lot bigger than Padme. It's

[12:40] probably one we're going to avoid. Image

[12:41] four. That's looking much better. I love

[12:44] the perspective of this. This next image

[12:46] is completely broken. perspectives are

[12:48] wrong. Seems to be a dual lightsaber

[12:49] which Padme is holding from the wrong

[12:51] side. So, we're not going to use that

[12:53] image. Final image looks like two Padme

[12:55] fighting each other. Again, it's less of

[12:57] a problem because I know I can change

[12:59] one of those characters, [music] but I'm

[13:00] not a big fan of the perspective. I'm

[13:02] not a big fan of the background. So,

[13:04] with all of that in mind, I think I'm

[13:06] going to go with this image as the image

[13:08] I'm going to refine to final.

[13:11] Okay. Okay. So, the first thing I'll do

[13:12] is drag that image into impaint and

[13:15] start erase my previous painting and

[13:17] start working on specific sections of

[13:20] the image. So, I'd want Padme to wear

[13:22] her iconic white battle attire. So,

[13:24] these images are already more in

[13:26] direction of where it needs to be. This

[13:28] time, I'm only going to change up the

[13:29] outfit itself. Just highlighting very

[13:31] specifically. So, our first section of

[13:33] images are coming in and they're all

[13:34] looking pretty great. So, I love the

[13:36] position here. Doesn't look perfect in

[13:38] terms of outfit. Second outfit looks

[13:40] much better, but there's lots of

[13:41] problems with the hand structure. The

[13:42] hand seems a lot smaller than it needs

[13:44] to be. The foot structure isn't great.

[13:46] This third image is better in terms of

[13:48] hand structure, but there is a broken

[13:49] limb right here, and there's a few

[13:51] issues going on in the lower parts of

[13:53] the frame. So, here's a pro tip. In this

[13:54] particular section, I'm changing how the

[13:56] body structure looks, but I don't really

[13:59] want to change the general position. I

[14:01] find that it's helpful to highlight only

[14:03] parts of the image. So you'll notice

[14:04] here I've left the foot open and the

[14:07] back of the ankle open as well along

[14:08] with the waist of the character. That

[14:09] way I can change just the middle section

[14:11] of the image without affecting the

[14:12] overall look of the Now I'm liking the

[14:14] overall structure here. I just want to

[14:16] fix this broken limb. I'd like to change

[14:18] the shoes the character is wearing.

[14:19] Maybe reduce some of this armor. And

[14:21] that's what I'm going to do now. Now you

[14:22] can see that after a certain point I'm

[14:24] going to start running parallel

[14:25] processes. In this section I've

[14:27] highlighted Padme's hair very

[14:28] specifically. So I'm trying to get her

[14:30] hair tied up and it's proving to be a

[14:32] little bit difficult. But the important

[14:33] point here is that you need to keep a

[14:35] reference ready. I'm trying to match

[14:37] those references. Having a reference

[14:39] will always help get the details right

[14:41] and anchor your character in reality.

[14:43] So, one thing we'll definitely need to

[14:45] do is work on character features,

[14:48] especially in the face. You want to have

[14:49] as much detail as possible in the face.

[14:51] So, even if things aren't perfect, most

[14:54] people won't notice. And you can see I'm

[14:57] already painting in the second image

[14:58] while that first image is generated. I'm

[15:00] going to need to work on that next. So,

[15:02] Anakin, Star Wars, detailed face, it's a

[15:04] lot better. We're not going to perfect

[15:07] anything. We're going to do the best we

[15:08] can and really just go through all of

[15:10] the different tools and functions. If

[15:12] you'd want to see us explore that in

[15:14] another video where we deep dive into

[15:16] creating hyper realistic images, let us

[15:18] know in the comments below. Now, in this

[15:20] next section, I want to show you how to

[15:21] expand an image. Now, you've probably

[15:22] seen something similar if you've used

[15:24] Adobe's Photoshop with content of

[15:26] airfield, but it's quite powerful in

[15:28] focus because you can really control and

[15:30] refine how the expansion work prompt you

[15:32] put in here is very specifically the

[15:34] information that you want in the

[15:35] background and in the expanded areas.

[15:38] But ideally, I'd like to have this in

[15:40] landscape format. So, what I'm going to

[15:41] do is I'm going to take that image, put

[15:43] it back into content expansion, and just

[15:46] generate the left and right side of the

[15:48] image. Okay, that's looking really

[15:50] great. I love how that looks. Now, I

[15:51] don't really like how the characters are

[15:53] standing. I'm seeing a lot of

[15:55] imperfections in the hand structure,

[15:57] mannequin's leg structure, but generally

[15:59] this is a great starting base

[16:00] considering we generated this image in

[16:02] under an hour. That's a really great

[16:04] starting point. Now, there's a lot more

[16:05] you can do here using the refiner tools.

[16:07] You can bring in things like smoke and

[16:09] fog. You can work on the character

[16:10] outfits, bring in detail, so things like

[16:12] the shoes and the hair. I'll also look

[16:14] back at my references and really

[16:16] understand where I missed the ball. And

[16:17] that brings us to the end of this

[16:19] episode. If you like this video, hit the

[16:21] like and subscribe buttons. Now, there

[16:23] is a lot more we can do to perfect this

[16:25] photo.