📛 Don't be so excited about SDXL, your 8-11 VRAM GPU will have a hard time!

[ZeroCool22]

You will need almost the double or even triple of time to generate an image that you do in a few seconds in 1.5, SDXL is designed to run well in high BUFFY GPU's.

If you have an 8-12 VRAM GPU or even a PASCAL one like 1080 TI, you will be waiting forever to the image finishing generating and not even talk if you use the refiner, which is necessary to not get shitty faces.

So i would advise you to stick to 1.5 until optimizations comes out (if that is possible) and could get almost or close times generations that with 1.5.

[Alcyon6]

What is the sweet spot? 12 GB? 16? 24? 32? 48??

[SansQuartier]

I tried it with comfyUI,, takes about 30-60 seconds to generate an image whereas Sd2.1 takes around 20s (on a RTX 2070 Max-Q, 8Gb VRAM)

Not a dealbreaker.

[reese555]

same not really seeing an issue the only thing i did was dial back resolution 768x768 works fine and generation speed is more than acceptable on my 8gig 3060ti. i mean sure it doesnt finish in 3 seconds but neither did sd 1.5 when it came out

[projecthorizon993]

I tried and it fast ten second with my rtx 4090

[projecthorizon993]

well I do some heavy stuff like 2048x2048 and run some game in background

[projecthorizon993]

It 10 second
lol

[jipsita]

OP: SDXL will be slow on old cards.
YOU: it's fast on my new card!

[2blackbar]

lol,test it again and turn that bg stuff off

[DragonZealot]

hes the only person saying anything relevant. could care less how it runs on low end hardware. I would like something that properly fully uses my 4090 kthx

[Davikar]

White it's not superfast, on SD Next it takes around 2 minutes on my GTX 1080ti for a 1024x1024 image, which is about the same time it takes to render an image with hires fix on SD 1.5.

[jipsita]

So it works on a GTX 1070 for example? I was afraid because the developers said that minimum was a GTX 20xx to run the model.

[fangtianfrance]

1080ti, I use this card to generate 1024 pictures in about 53 seconds under comfyui, which is much easier to use than automatic1111. It's just that I can't train sdxl with 1080ti.

[ihojun887]

I get out of memory error with my laptop 3070 vram 8GB

How come you don't face memory error while still using 1080ti?

[Wladastic]

I do not face that issue either.
Have you activated offloading?

[ihojun887]

I don't know what offloading is, exactly.
If offloading is some concept like one in the article, https://www.felixsanz.dev/articles/ultimate-guide-to-optimizing-stable-diffusion-xl, I think I didn't activate offloading.
It surely worth a try. Thx

[Rkkss]

Memory won't be huge issue with tiled vae.
Though SDXL use a completely different workflow, I don't know the current webui can even do it.

[mattieshadow]

My 10 VRAM gpu runs it just fine no issues, goes pretty fast too.

[MrKuenning]

How? my 10GB 3080 takes about 12 minutes to generate a 1024x1024.

[esprit112]

I tried SDXL on a 3060 12GB with no luck, might have to wait till the model is optimised.

[projecthorizon993]

good luck

[TomLucidor]

You borrowing this from here? https://stablediffusionxl.com/sdxl-system-requirements/

[neggles]

Overall it should be within the same ballpark as SD1.5 as long as you have enough VRAM.

[Centurion-Rome]

as long as you have enough VRAM.
main problem by many users with 8GB VRAM :(

[Appolonius001]

with sdxl .9 comfyui (i would prefere to use a1111) i'm running a rtx 2060 6gb vram laptop and it takes about 6-8m for a 1080x1080 image with 20 base steps & 15 refiner steps
edit: im using Olivio's first set up(no upscaler)

edit: after the first run i get a 1080x1080 image (including the refining) in Prompt executed in 240.34 seconds (4m)

[yamfun]

SDXL and 4060 Ti 16GB release on like, the same day, really makes you think.

[projecthorizon993]

very sad day

[lcmiracle]

Yeah, I'm just waiting to see if 60Ti 16G will be a good AI card, if not, I might just have shell out for a RTX 4080, sigh

[bjornlarssen]

I'm running it on Colab, T4/High RAM using this extension https://github.com/lifeisboringsoprogramming/sd-webui-xldemo-txt2img and it takes about 1 second/iteration. (High RAM is necessary, because the extension has massive RAM leakages, but it's more than fast enough for my needs.) Colab informs me I have 15GB VRAM, SDXL doesn't go above 9GB, same as 1.5 models. Don't see it running on my M2 8GB Mac Mini though…

Can't wait to use ControlNet with it.

[draxx318]

Did you try it now without the extension? I tryed to run it on t4 high ram but i get disconnected. Had to upgrade to a100 but its too expensive now.

[bjornlarssen]

I'm using 1.5.1 now with SDXL support built-in. And with a giant RAM leak, too. :/ I'm running on A100/High RAM and I have to restart (luckily not disconnect/delete, because that takes another five minutes) every 15-20 generations or so. Wish there was a "free RAM" extension…

[draxx318]

I am experiencing the same thing, i used the --medvram flag with a t4 and it works but there is a memory leak yeah. Even with ComfyUi the ram gets bloated after a couple of generation and it disconnects. This makes me think the model has a problem not the acual software you are running it on.

[jav12z]

I think what delays the proccess is the change of model from base to refiner. I don't know why the refiner model is so heavy.

[reese555]

noticing this too base goes fast but from switching off go refiner to completion is where it slows down.

[NicTanghe]

Wow i thought you`d need a super cluster to run it.

[DTown]

Well, I have a 3070ti with 8 gig vram. Worked well so far with all the other models. When I start SDXL it first starts pretty good and even fast, but right before the image is finished ... Cuda out of memory :-(

I think 8 gig should work, not sure if there is a good trick to free the vram before starting a1111. Maybe I should try comfyUI

[draxx318]

Today i found that you can execute the script with the flag

--medvram

And that should fix the problem. Will try later and let you know.

[draxx318]

On google colab it works using this flag. Its 15gb of vram tho.

[DTown]

It really worked with that, thx :-)

[GrindHouse66]

I can run SD XL - both base and refiner steps - using InvokeAI or Comfyui - without any issues. Takes around 34 seconds per 1024 x 1024 image on an 8GB 3060TI and 32 GB system ram. Automatic1111 won't even load the base SDXL model without crashing out from lack of VRAM. --medvram and --lowvram don't make any difference.

[draxx318]

Its strange. I have issues with ComfyUi where the Ram gets bloated after 3 or 4 generations, like there is a memory leak. Im on google colab with 25gb of ram.

[suonnon]

Idk but it took just 10-15seconds to generate 1024x1024 image for me. RTX 3060ti 8GB

[PaulFidika]

I have an RTX 4070 Laptop GPU in a top of the line, $4,000 gaming laptop, and SDXL is failing because it's running out of vRAM (I only have 8 GBs of vRAM apparently). I don't mind waiting a while for images to generate, but the memory requirements make SDXL unusable for myself at least. If anyone has suggestions I'd appreciate it.

[Frankenburger]

In my testing, Automatic 1111 isn't quite there yet when it comes to memory management. The issues I see with my 8GB laptop are non-existent in ComfyUI. So, with that said, these are your best options for now

1. Use --medvram
2. Use ComfyUI
3. Stick with SD1.5/2.1 for now and wait for Automatic1111 to be better optimized

[MrKuenning]

On my 3080 I have found that --medvram takes the SDXL times down to 4 minutes from 8 minutes.
But it has the negative side effect of making 1.5 images take 40 seconds instead of 4 seconds.

So SDXL is twice as fast, and SD1.5 takes 10x longer.

[SakuraJensen]

I use a Quadro P4000 8GB and I don't have any issues generating images with SDXL. When I upscale the images I'll go into the 18GB of VRAM territory but since webui-1.5.0 and the latest gaming drivers from nvidia I can use all of my shared memory without crashing. So it's not that big of a deal. Since it only takes 2-3 minutes even when bleeding over into system memory to do the upscale. In the initial txt2img generation I average about 6.7GB for a 1024x1024, 50 euler a steps, with --xformers --no-half-vae --medvram. ~30seconds-1m 30s per image.

[SakuraJensen]

Have a Quadro RTX 5000 16GB now and I still don't have any vram issues with XL now. Really don't need more than 12GB though even at 896x1216 and upscaling to 4K.

[creativojavier]

I use a GTX 1060 6Gb, with 64 Gb DDR4 RAM. Besides installing Manager and some custom nodes, like the ones from this video, I ran CounfyUI, loaded SDXL 1.0 base and refiner models, with max 40 steps each, and it worked, about 3-4 minutes/image. And GPU reaching 85ºC.
Right now I'm looking for update GPU, not sure if 4060 is worthy instead 3060. 4060 has less frame buffer and apparently matters.

[SakuraJensen]

IMO The 4060 Ti is reasonable. So are used RTX A series cards that have more VRAM.

[creativojavier]

Yup, after some research i decided to get the 4060 TI 16 Gb.

[MichaelMcKinnon]

It takes between 2-15 minutes on my RTX2060 6GB. Not sure why the time is so variable. The first run is usually the fastest. My 16GB RAM is almost all used as well. My SSD get hits hard periodically. I basically can't do anything while it's running at times.

[MichaelMcKinnon]

The GPU is 56ºC in a mini-ITX build, so it's entirely VRAM bottlenecked. The demand for VRAM is going to increase the amount on professional GPUs. Hopefully there's some trickledown to consumer GPUs as well. It doesn't look like the case for the next year or so. The mid range price/performance of PCs hasn't improved much since I built my mine. As much as I want to build a new PC, I should wait a couple of years until components are more optimized for AI workloads in consumer hardware. Or drop $4k on a 4090 build now. The high end price/performance is actually good now. Everything is backwards.

[MichaelMcKinnon]

I just upgraded from 16GB to 64GB RAM. It made a huge difference in performance and stability. I went from 5-15 minutes per image to 1-2

[MichaelMcKinnon]

That said I could probably have gotten away with 32GB

[tzenrick]

2060 Super 8GB, [email protected], 32GB@2866, 2TB NVME, 20TB rotating storage: It takes me 5-30 times longer, per step, than 1.5 models. While models are loading and being transferred, it'll spike my CPU and RAM usage, and start hitting the swap file. That only last about 30-60 seconds, for the large models. It then runs about 10-15% CPU usage, roughly 50-70% (allocated by python) RAM usage, and all of the GPU is used.

And that's with the Automatic1111 gui running, previews on, reddit, a google search, an article, and a youtube video, across 2 Firefox windows, on 2 screens; Discord, Messenger, Stardock Deskscapes, and just everything, running with full hardware acceleration.

Occasionally, my 1080p Youtube video stutters...

MSI Afterburner reports 2100MHz on the GPU, 7175MHz on VRAM, 1043mV, 50 degrees C, 58% fan speed.

Contrary to the title of the thread, my 8GB card works fine for SDXL. I'm probably not going to be training any models without some hardware upgrades.

[JJSPACE85]

I am researching a laptop to buy that I intend to use stable diffusion on and it brought me across this forum. You guys clearly are knowledgeable about SD and most the things that you are saying I don't even understand. I had been looking at a laptop with a RTX 4070 w/ 8gb vram. It seems that isn't quite ideal but people are getting it to work. In reading this forum it brought up another question for me. I am wondering what purpose you guys learned all this knowledge about SD? Do you just do it for fun to create cool images, or do you guys use these skills somehow for business/money earning endeavors? Would you have suggestions on ways to use SD for making income? Thanks in advance.

[JJSPACE85]

I am researching a laptop to buy that I intend to use stable diffusion on and it brought me across this forum. You guys clearly are knowledgeable about SD and most the things that you are saying I don't even understand. I had been looking at a laptop with a RTX 4070 w/ 8gb vram. It seems that isn't quite ideal but people are getting it to work. In reading this forum it brought up another question for me. I am wondering what purpose you guys learned all this knowledge about SD? Do you just do it for fun to create cool images, or do you guys use these skills somehow for business/money earning endeavors? Would you have suggestions on ways to use SD for making income? Thanks in advance.

[PositivPy]

IMO if you have the cash, you should go for a macbook. The M chip has shared RAM so if you get a 16Gb laptop, you get roughly a 10Gb GPU. If you get the 32, you'll be relatively future proof.

[Wladastic]

I use a RTX 4080 and it doesn't even generate anything for me.
I either get unfinished renderings or it just crashes with some weird UNAN exception.
Maybe it broke while downloading..

[Wladastic]

Fixed for me, forgot to update.

[mocheez13]

I'm using an RTX 3060 12 gb, on Automatic 1111 1.6.1, and I find SDXL surprisingly fast, even though it's a little slower. Still fast enough to be useable. The only reason I'm sticking to 1.5 is because I like the results better 😊

[ArmyOfPun1776]

This didn't age well... lol... 3050Ti 4GB Run SDXL in Comfy @ 896x1152 in about 30 seconds. 5 min if I upscale x2 with CR or Iterative Upscalers

[Wladastic]

yeah same for me, you can use upscaling later

[iocmet]

It generates fast on my RTX 2060 SUPER 8GB VRAM and 32GB RAM with --medvram but after some amount of generations crashes with out of memory (RAM, not VRAM)

[TomLucidor]

So it is a RAM issue and not a VRAM one? Kinda surprising! What are the requirements, and can most PC boost their RAM on the cheap?