Tech Craft: Fooocus vs ComfyUI for Intel Arc GPUs

Video Statistics and Information

Video
Captions Word Cloud
Reddit Comments
Captions
Hey it's Bob Duffy from Intel here and once again we're doing stable diffusion on Intel Arc gpus but here we're going to battle it out between two WebUIs the fooocus UI and the comfyUI now these are both very powerful WebUIs in their own right but they couldn't be more different they are at the extreme ends of the spectrum for user experience so we're going to kick it off with fooocus first because stable diffusion has been criticized for being not very consumer friendly. Really a UI developed for developers rather than consumers well Fooocus is here to challenge that thinking. It's a super simple interface and then we're going to jump over to ComfyUI see how that works and how different it is and I guarantee you one of these uis is the right UI for you. And also look in the description below because I will have installation instructions or at least links to installation instructions for you guys to figure out how to get this to work on your own Intel Arc GPU. So let's start with Fooocus. See you there. Okay so here we are in the Fooocus WebUI for stable diffusion and like I said a super simple interface you have this large image preview area here that's there for as when you generate images or you can click through any images within your gallery that get generated. Other options are pretty simple. You've got an area for your positive text prompt and then you have a button to generate. And so that's the bulk of the the UI. It's it's super simple to use a couple of other options um you have an"image" input option if you want to do image to image or inpainting and I'll show you that and then you have "advanced", which is really a settings area. A lot of things are just set up by default um so we're going to just kind of keep it as is, and I'm just going to go ahead and click generate um so we'll get an image here, and I'll walk you through what's going on. So fooocus is using the Juggernaut XL model and that is by default you don't have to go to CivitAI or Hugging face to download anything it's already included so it's super simple, and so you can see it's working here but it had some overhead up front and you'll notice this as we do things there's a little bit of an overhead so there's extra time that you'll see in Fooocus doing things like Refiners. So what it does is it tries to do the things that would imagine that you'd want to do in order to create a good image and it just automatically does them for you. And so you can see here. This is just gorgeous work here. I really love Fooocus because it really takes a lot of the work out of getting it right. Things just look uh really good most of the time. So if you want to do a little bit more you could jump into the advanced features. But what I want to do first is show you kind of the feature that I love most and it's this um out painting area so I'd already done it here so let let me show you what I did um you can take an image that you get generated um select the input image and then drag that down below into this area here. And then you've got some options here for in paint and out paint I'm going to show you this very first one here and demonstrate it here but let me talk you through some of the others you could do improve detail which that allows you to then go ahead and paint an area and then describe what you want to see in that area and that's usually what it's going to do is improve that so it's going to use the information that's already there and try to make it better and then you have the other area which is modify content again you'd paint an area and then you describe what you'd want there but it's going to try to differ the information from what's already there. But what I'm going to show you guys is the out painting feature. The feature that I really like So we we've got this set to out painting and I'm going to select the right and bottom area. So what it's going to do is going to make a picture where it expands it out to the right into the bottom and it's going to imagine what could have been there, or what should be there. So let's go ahead and click generate and it'll run through that. But to prove this is all running on Arc you can see that will hit my GPU here and also you can see that it is doing some of that kind of work it's it's taking the model moving it over to the GPU. Calculating the refiners and then it's going to go ahead and start working on it. And here we go so we're starting to see this image get created here, down below and off to the right. Now if you're going to do out painting what I suggest is you make sure that you're on the "speed" option. So if you go to advance you'll see some of these preferences. There's "speed", "quality" and "extreme speed". Speed is set to default and that's going to be doing sdxl at 30 steps, which is usually what you do in automatic 1111. "Quality" is going to bump that up to 60 steps. And extreme speed is going to do an LCM fast LoRA, down to eight steps. But as you can see here it did a really good job in this out painting. The areas in the background. It kept that out of focus where it matched the depth of field that was happening in the image. And then the images or the stuff that's in the foreground. It kept that clean and crisp And it matched her garments there so a really powerful feature. And if I wanted to I could take this image, I could drag it down below and I could do more to the right or the bottom, or I could do top left and bottom. I could you know try to imagine or see other features that are not in the picture and generate that So very very powerful, but what I want to show you is the extreme speed So let's go ahead and jump to that and let's go ahead and do three images there. We'll go ahead and turn off uh image input and turn off the Advanced,Wait a minute and let me make sure that I'm on random. There we go, if I wanted to I could add a negative prompt I didn't show you that but that is a feature that's available. Tt's hidden in the advanced area there because mostly what they figure you're just going to type hit the button and there you go. But let me go ahead and click generate and then it's going to do the LCM now one of the things that you notice that all I did is I went over to Advanced and I clicked the extreme speed. If you were going to do LCM on automatic 1111 or comfy UI what you're used to doing is changing the number of steps say down from 30 to 4 or 8, um setting the CFG to something like 1 to 2, and then you also be changing the the scheduler. So you'd be changing it from like Euler A or something like that to LCM. And this is doing all of those things that you would normally do just in one button, for you, so it's super convenient and great and the work the stuff that it does is really good. What I notice is when I'm doing A1111 or ComfyUI and I'm doing fast Laura, to me it's like a-cut-above what I would typically get if I was doing 512 x 512 at say 20 or 30 steps but this does it really well and you can see that it it ran through each of these and when it did those when it was actually generating and was was doing the diffusion it took about 3 seconds per image. But again it has this overhead thatFooocus usually has in order to do the things that it thinks that you should be doing. So there you have it that is the Fooocus UI. Super simple interface. It takes all the thinking out of it. Takes all the tweaking configuration out of it. But it allows you to go in and do some of that stuff if you really wanted to. Anyways we'll contrast that with ComfyUI so we're going to jump over to ComfyUI. See what that's like. And we'll first do LCM fast LoRA and you can compare the difference. see you there. Okay here we are in The Comfy UI interface and we're already generating images. I thought I'd just get this going and then I can talk through what's going on. So right now we're generating images using LCM fast LoRA and I'm using RealisticVision 1.5 so this is SDXL. This is fast. but we're generating images pretty fast here let go ahead and change and look at this here. Yeah in less than a second we're generating each image. It's not loading a model or moving them over to the GPU for every time that it's running. There's no refiners running. It's just straight happening and it's happening very fast let me go ahead and turn off this auto queue and I can talk through what's going on in ComfyUI So you'll notice that the node is uh that the interface is very different than Fooocus. It's using this node-based system. So if you're used to Blender or UnReal Engine or other node-based systems here, this is going to be pretty familiar to you. It's essentially it's like a visual programming method. Essentially programming without code, but just connecting the outputs from um one function to the inputs of another. So the very first thing that you're going to have is something like checkpoint Checkpoint is that thing where you'd be in automatic 1111 or SDNext and essentially you're seting which model that you're using. Some of these models could come from Hugging Face. They could come from CivitAI, and you can choose from all sorts of models. So I've gone ahead and downloaded a number of models from CivitAI that I like to use and I've made them available here and I've selected the RealisticVision Model. Typically clip is going to connect into, if you follow these yellow lines, into your text prompt and you have a positive prompt that's going into your sampler and then you have a negative prompt that's going into your sampler. You also have latent image as part of the sampler and that's where you're going to specify your resolution. And then most of the other features that you would have typically within stable diffusion, like the number of steps the CFG the Sampler Name, all that kind of stuff is available here in this sampler node. But we were doing LCM fast LoRA and that's available through this one note here that is essentially intersecting between the checkpoint in the sampler, and my text prompts to the sampler. So that's what it allows you to do. It allows you to establish what's the model so I take the model here and I plug it in and I take the clip information and I plug it in here. And then the model goes back out to the sampler. And the Clips connect to your text prompts which then go into the Sampler here. But I do have an option for 1.5 or XL. So let's go ahead and give that a shot. I'm going to go and change this over to the JuggernautXL model that we were using in Fooocus, and then let me go ahead and change this to SDXL and then I'm going to go ahead and change my image to 1024 and 1024, Okay and that should be all that I need to do in order to change this to an SDXLmodel using Juggernaut okay now because I'm doing LCM fast Laura I don't need to really change anything else here that is all set so let me go ahead and Queue prompt. Now the first time it runs this it's going to go ahead and load that model and then it's going to move over to the sampler and get that done. So that's happening pretty quickly and as you can see we've got an SDXL image here that looks really nice and you can tell I'll go ahead and queue that up as it's going on and this is faster than we were seeing withFooocus so Fooocus has this additional overhead um and we don't see that here these are happening in about three to three and a half seconds which is about what it took on Fooocus as well but Fooocus was doing some refiner work as well as moving things over to the GPU um before it went ahead and worked with the sampler this is just happening licky split. Things are going really fast but I noticed that you know the quality maybe not as good as I was getting in Fooocus but it's certainly really nice um and if I wanted to I could increase the steps or I could you know change the resolution and hope that I would get that better or what I could do is just not do LCM fast Laura and essentially bypass that. So let's go ahead and do that so what I'm going to do is I'm going to disconnect the clip here and disconnect this clip here from the LoRA and I'm going to connect directly from the checkpoint node and then I'm going to do the same here for the mod model disconnect it and connect directly into the model there and then I think uh this information terms of the latent image is fine but I do have to make some changes because I'm doing standard um SDXL so I'm going to change this to 30 steps and I'm going to change this to 7 and I'm going to change my sampler from LCM to Euler A and then this to Kerras and that's typically what you'd see these are the standard things no matter what you'd be doing if you're doing SDXL um and you're doing it in on let's say automatic 1111, this is what you'd have so I think everything's good so I'm going to go ahead and hit Queue prompt now this will take longer to do because we're obviously we're now doing 30 steps we are doing LCM fast low at eight steps so this is going to take more I think typically this takes usually around 20 seconds that's my experience um 15 to 20 seconds depending if you're doing say uh 15 or 20 steps uh so it's going to go ahead and get done and we will see that image over here on the right yeah there we go another beautiful Alchemist image ccoming out of this. And I go ahead and kick that off again but um that demonstrates the the possibilities with ComfyUI is you can do ControlNet you know you can do inpainting. You can do out-painting. You can do image to image. There's a lot that you can do here and there's even stuff that you can do with custom nodes. And that's the next thing that I'd love to show you guys is let's step this up a notch. Let's add in the custom nodes and see what comfyUI is really all about Okay here we are back in stable diffusion and I'm using one of my other favorite applications and that is Blender. Blender is being used with a custom node called ToyXYZ which will capture information from blender and pipe it directly into ComfyUI and that's using ControlNet in order to influence the outcome of that image. So as you can see I've got fast LCM LoRA going on here and we're generating a lot of images really quickly based um that Blender imag,e but it's happening real time. So I can move this around and I'm going to go ahead and then use that in order to get new information rendering So you can see the power of this is quite incredible you can imagine companies like Blender and Autodesk and how they might actually use artificial intelligence use tools like stable diffusion as another tool in their toolbox for doing things like fast or realtime rendering. So we had Fooocus which is super simple to use it takes a lot of the extra working configuration out and gives you high quality images and then we have ComfyUI that allows you to really adjust all the knobs bells and whistles and switches to really influence the outcome with some really cool custom nodes from the community, So I hope you found this enjoyable, so especially if you did, please like this video and subscribe to this YouTube channel. Many more videos like this to come. Until next time I'm your host Bob Duffy and we'll catch you later. [Music]
Info
Channel: Intel Technology
Views: 4,043
Rating: undefined out of 5
Keywords:
Id: L0RKVvdIpFA
Channel Id: undefined
Length: 15min 55sec (955 seconds)
Published: Wed Jan 24 2024
Related Videos
Note
Please note that this website is currently a work in progress! Lots of interesting data and statistics to come.