Who Says You Can't Run SDXL 1.0 8GB VRAM? Automatic1111 & ComfyUi

Video Statistics and Information

Video
Captions Word Cloud
Reddit Comments
Captions
so here's a system info on automatic 1111 you see my 3060 TI we're running sdxl 1.0 the base model 30 steps Euler ancestral 1024 by 1024 this is in real time to show you that I can run this under 45 seconds only one image though and uh yeah I'm not sure what's going on with the rest of you running like 40 80s 3080s that are not being able to generate images I hear five six minutes although I do think maybe some of the extensions might be causing some issues so I've done a fresh install you see here we're done and this is the base model image we're gonna send it to image to image and use the refiner and one of the things though with automatic 1111 is that you know this isn't really the proper way to use the refiner sort of a workaround so we're going to bring the sampling steps down to 20 because really what we're doing is we're trying to finish off the base model and give it a bit more detail I did bring the denoising strength to 0.2 and as you see here it's running fine now even though Ron's okay here I encourage you to try comfy UI it's just a better interface to take advantage of the refiner so here we are in comfy UI and if this looks all confusing to you don't worry about it I'm going to cover it in another video but I wanted to show you a quick demo on using comfy UI specifically for how the refiner works so if I could just quickly show you the setup here we have our prompt area here we have our checkpoint models here's the base model the refiner model this area here is the clip encoders here's the latent image section and we're going to run a 1024 by 1024 image this area here is our sampler information one thing I wanted to point out to you quickly here since this is the base model we're only going to run it at 20 steps starting at zero but when we come up to the refiner portion of the process we have 30 steps however it's going to start at 20 steps this is the proper way to use the refiner which automatic 1111 can't do at the moment theoretically sdxl 1.0 is a two-step process sort of like an advanced imaged image where we develop the base model it creates the image leaving a bit of noise in it and then gets passed off to the refiner to finish it off as I generate the image you see that these areas are highlighted and you can really see the process of the whole pipeline so we've got it going through the checkpoint here it's going through the clip encoders now now it'll go through the latent image section and as I said before it's going to go for 20 steps and then pass it off to the refiner step above well after it gets vae decoded there we have the base image now it's going through the refiner process here and then it will take it at 20 steps and finish it off at 30 steps which will give us our final image so if we look at the base image that was processed at 20 steps you see it doesn't look too bad but it does require some fine tuning and now if we look at the refined image you see the details are more prominent it looks much nicer than the base model so that's the advantage of comfy UI where you can have that much control as opposed to automatic 1111 currently you're only able to do sort of a hack with image to image and it's really not the best way to take advantage of the refiner there's a few things I want you to try out if you really want to try it sdxl 1.0 first make sure automatic 11 11 that you've updated it to the latest 1.51 if you do the git pull method that's fine I encourage you though to do a fresh install because I have a feeling some of the extensions are causing some conflicts and if you're brand new to this and you want to start from scratch I highly recommend this installer from Empire media clients basically it's a standalone launcher and it pulls all the files from automatic 1111. now I will tell you comfy UI runs it very well on my 3060 ti so I highly encourage you to install this the other thing to consider other than doing a fresh install if you notice on my launcher here this is basically the same as the web UI user.bat file and in the startup options I've entered dash dash Med vram Dash Dash xformers and this line is what I think helped run it better in automatic 1111 for me which is dash dash no dash half dash vae one of the reasons why I like the launcher see here it says additional launch options I just entered it in there I've got the low vram option and the enable xformers options as check boxes so I don't really have to do anything except type this in here now those of you who've done a manual installation you just have to go into your stable diffusion web UI folder look for web UI user bat open it up with something like Notepad as you see here I have no half vae and you would also put in those other arguments dash dash xformers dash dash Med vram on this line but that's pretty much all I did so I'm not sure what's going on with the rest of you why you can't run it some people are even saying they can't run it on 3080 or even a 40 80 something's not right so I would recommend you to try with a fresh install your initial generation might take long but the rest of them should be fine hey if I was able to run it on my 3060 TI you should be able to too and if I have any viewers that are really schooled in this kind of stuff and know what the issues are let me know in the comments below perhaps you can help somebody else as well in the meantime I'll see you when I see you
Info
Channel: Monzon Media
Views: 20,152
Rating: undefined out of 5
Keywords: sdxl 1.0, stable diffusion xl 1.0, stable diffusion sdxl 1.0, sdxl, stable diffusion, automatic1111, comfyui, automatic 1111, automatic1111 sdxl, comfyui sdxl, sdxl rtx3060, sdxl 3060, stable diffusion sdxl 3060, stable diffusion xl 3060
Id: gguLtMM4g_Q
Channel Id: undefined
Length: 6min 38sec (398 seconds)
Published: Sun Jul 30 2023
Related Videos
Note
Please note that this website is currently a work in progress! Lots of interesting data and statistics to come.