Automatic1111 vs ComfyUI (Which one should you use???)

in this video I'm going to be going over the difference between automatic 1111 and comy UI so to begin I'm going to do a simple text image in automatic 1111 so I'm going to be doing a mountain landscape then click generate and automatic 1111 has generated the actual image now to do the same in comu y all I need to do is just take out this prompt over here and then type in a mountain landscape and then then click generate or Q prompt and comi was able to generate the image as expected now that was relatively simple what happens if you want to do something a bit more complex like image to image so going back to automatic 1111 to do an image to image all I need to do is click on this button here to take the actual image and now I'm going to change the prompt to say mountain landscape on fire and now if I click generate we have a mountain landscape on fire now let's say I want to do this in comy y now this is where things start to become a bit more tricky because the default workflow that comy UI provides you is actually just a simple text to image workflow if you want the image to image workflow you actually have to build it yourself so to do that what I would do is load image get an encoder delete the empty latent image then drag the load image VA incode and then connect the two together then add the latent to the latent image and then the vae to the vae of the encoder lower the D noising strength to something like 0.7 and now if I change the prompt to say a mountain landscape fire and then click Q prompt I have successfully managed to create an image to image workflow now if you're a beginner you probably had no idea what I've just done and that is to be expected and it's for this reason why I don't recommend confi for beginners because when you start off using stable diffusion there's a lot of things that you actually have to learn regarding the actual image generation itself so stuff like what sampling steps are what are CFG scales what do things like seeds do whereas in comi not only do you have to learn that as a beginner but you also have to learn at the same time how confi workflows work as well and so that's just double work and it will make your stable diffusion Journey a lot harder than it needs to be now am I saying that complete beginners can't use comfi not necessarily because comi has a really nice feature where you can copy another person's workflow so for example take this image over here of Gandalf riding a motorcycle now a person online generated this image in confi so now if I go back into confi and actually drag the image into the workflow you can see that it's brought about the exact workflow that the person used to generate the image and if I actually click Q prompt I'll be able to generate the exact same image that the person had made and looking at the image that was generated you can see that it's an exact replica of the original image now this ability to drag and other people's workflows is a really handy feature in comu y because you will find many sorts of different workflows that people online have generated and it saves you a lot of time this is not to say we can't do something like this in automatic 1111 in fact we can so if I go into PNG info and then drag an image that was generated in automatic 1111 so for example this car over here and then if I click Text image we can see that it's filled in the prompts for us and now if I go click generate you can see that it generated the image of the C that we had and while this feature is useful in automatic 1111 it's not as important as it is in confu y because if a person doesn't provide me with the details which they use to create the image in automatic 1111 it's much easier to actually reverse engineer how they did it whereas in comu Why Not only would I be trying to figure out what prompts they did and what settings they used I'd actually be trying to work out what workflow they actually used as well which makes things a lot more harder now the ability to create custom workflows in comi is one of the reasons why it's so powerful so take for example this workflow over here now while it does look very messy what it's doing is doing a simple text to image and then an image to image so if I do Q prompt two images have been generated so we have the text to image image that was generated and we have the image to image that was generated as well now I just want to highlight that I did all of this using one button click now if I wanted to do the same in automatic 11 11 first I'd have to do text the image generate the image send this image to image to image change the prompt so for example say I want a yellow car and then click generate and now we've generated an image of a yellow car now all of that took around about three button clicks and rewriting a new prompt versus the one click that I had to do in comi now I know that might seem like a very small difference but when you're generating a th000 images com you y with its ability to make custom workflows will save you a lot of time so for example I could have text to image that goes to image to image that goes to an upscaler which then goes back to an image to image and all sorts of other things whereas if I wanted to do something similar in automatic 111 I would have to use each of the separate ones manually so I'd have to do text to image image to image and then try to find the upscaling options and then go back to image to image and as you can see that's much more tedious than the simple one button click that you'll be able to do with conf Y and while this ability to create custom workflows is conf y's biggest strength it's perhaps it's also biggest weakness from my point of view as well so let's say for example I did not want to use this control net portion of the image over here and for those who don't know basically what control net does is in a nutshell allow our image to image to more closely resemble the starting image over here so now let's say I don't want to use this over here here what would I need to do well I'd actually have to manually disconnect the nodes so first I have to remove this and then connect this to the positive lower the D noising strength into something suitable and then click Q prompt and comi generated the image and you'll be able to see that the initial image and the final image look slightly more different than when we weren't using the control net let's say that I really liked what control net did for the image and I want to actually use it in my workflow again so to do that I'd actually have to reconnect it to the workflow so I'd have to drag all the nodes back to where they were originally and then take the d noising strength to where it was originally and then click Q prompt and conf why was able to generate a stylized image that looks very similar to the original one now the point here I'm trying to make is that I had to take out time to actually disconnect the nodes and reconnect the nodes and while this was a very small example and was very quick let's say if I wanted to change this workflow to include another image to image or another text to image or I wanted to use additional lurus all of that requires disconnecting and reconnecting notes and while technically I could copy another person's workflow it still takes time for me to actually find the workflow and make any adjustments if need be all this means is that in my personal opinion experimenting and comu white is a really big hustle compared to something like automatic 1111 because with automatic 1111 I sort of have pre-built workflows for me at the click of a button and I don't have to worry about reconnecting or disconnecting nodes and going back to the control net example I gave you beforehand in confy UI if I scroll down in automatic 1111 provided that I had control net installed all I need to do is actually select it from the options menu enable it and then click generate and automatic 1111 would use control net to generate the image now if I didn't want to use that all I'd have to do is actually just take off the enabled button over here and just like that I'm no longer using control net no need to actually manually reconfigure nodes so when it comes to experimentation I find myself using automatic 1111 much more just because I have everything that I need to on hand and if I do need to reconfigure anything it's much more simpler than it is in comi where a simple change could involve me reconnecting conting and disconnecting multiple nodes and while some may point out that you can have certain extensions which disable or enable parts of the workflow depending on if you want to use them setting things up like that requires additional time and when I'm experimenting I don't want to be messing around with the actual workflows I just want to be generating images now the final point I want to talk about is performance and why comi is a lot better than automatic 1111 when it comes to this so to demonstrate this I'm going to be using SD XL model at the 1024x 1024 resolution and I'm just going to use a motorcycle as the prompt and for those who are unaware sdxl are sort of the newer high resolution models that you can use in stable diffusion and as a result there are a lot more performance demanding than the older models that worked at a lower resolution in stable diffusion now I'm going to click generate and I also have this task manager open here and the only met you really need to pay attention to is the dedicated GPU memory usage and notice how it spikes when we're generating the image so it's almost reaching 8GB and now as we're approaching the final parts of the image generation you can see that it's basically touching HGB worth of vram now that the image has finished generating you can clearly see when the image was stopped being generating and you could tell that when it was generating it was reaching 8 GB worth of vram you could tell that it was really pushing my PC to its limits in fact whenever I do try to generate an image using XD XL it actually ends up freezing because it's just that demanding on my PC now I'm going to generate an sdxl image in comu and compare the difference back in confi I've just made a simple workflow that will take an sdxl model and generate an image of a motorcycle and now I'm going to click Q prompt and I'm going to bring up the task manager again so we can mon to the performance now that the image has finished generating we can clearly see that comu y uses less vram than automatic 1111 and this is where one of the key Parts where comi shines because with this one on a 8GB vram card I'm able to generate sdxl images without freezing up my PC in comu y whereas in automatic 1111 anytime I triy to do that it would just basically freeze up my PC and so if you do want to generate higher resolution images or use more performance demanding models comfi is the way to go so to wrap this video up I think both comi and automatic 1111 are both amazing tools but they're both suited for different purposes if I'm a beginner or somebody who just wants to quickly experiment then automatic 1111 is the way to go however if I already have a custom workflow in mind or performance is very important to me then comu is a lot better in this regards so that's just my opinion if you have an opinion of of your own please feel free to leave it in the comments section and I'll see you in the next video

Channel: Neo Professor

Views: 3,849

Keywords:

Id: MA5hg9iHfJg

Length: 12min 4sec (724 seconds)

Published: Wed Apr 10 2024