Extremely slow SDXL image generation

[Renaldas111]

I am using RTX3070Ti Laptop GPU with 8G of VRAM. The speed of image generation is about 10 s/it (10241024 batch size 1), refiner works faster up to 1+ s/it when refining at the same 10241024 resolution. Why so slow?
In comfyUI the speed was approx 2-3 it/s for 1024*1024 image.
Does it mean 8G VRAM is too little in A1111? Anybody able to run SDXL on 8G VRAM GPU in A1111 at "normal" generation speeds?

[edgartaor]

i think there is a bug right now. #12267
I'm using this commit 68f336b and getting ~1.5 it/s with a 8GB VRAM RTX2070
Consider downgrade for now.

Another possible cause are the NVIDIA drivers #11062

[je-skyne]

Same problem here but I can confirm you that your solution worked for me. I've got an RTX 3060 12Gb and everything was super slow. with SDXL models, but the refiner option is not available by default.

[dhwz]

8GB VRAM is absolutely ok and working good but using --medvram is mandatory

@edgartaor Thats odd I'm always testing latest dev version and I don't have any issue on my 2070S 8GB, generation times are ~30sec for 1024x1024 Euler A 25 steps (with or without refiner in use)

Note you need a lot of RAM actually, my WSL2 VM has 48GB RAM assigned

Can you please share your sysinfo file?

[MD-AZMAL]

i am using --medvram, i have 6GB RTX 3060 with 32GB RAM, this is my bat file:

@echo off

set PYTHON=F:\Pyhon\python.exe
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=--xformers --medvram --no-half-vae 

call webui.bat

even with medvram i am getting 6-7min to generate 1024x1024 images, also i am using TAESD, still i've got little to no improvement. Any idea how to fix it

[usigg3]

Use ComfyUI instead. I have the same problem and nobody here seems interested in fixing it. With Comfy I can generate a 1152 by 892 in 30 steps in just over a minute with controlnet!

[MD-AZMAL]

yepp, i'm using both, until automatic1111 fixes it, as the plugin ecosystem is much better here. I still dont understand why --medvram has no effect for me

[Renaldas111]

So, it is normal to get 10s/it without --medvram on 8G VRAM GPU?
Ok, must admit, --medvram gives me 1.5 it/s generating 1024*1024 images with Euler a sampler.

[dhwz]

Yes thats absolutely normal as you run out of memory without --medvram which makes it extremely slow then.

[AbstractEyes]

I'm getting slow generations with my 4090 24gig.

[dhwz]

How slow? RAM size? Which live preview mode is configured? Do you use xformers?

[AbstractEyes]

64 gigs normal ram, live preview mode is high quality, xformers enabled. My computer's normal ram isn't capped.

[dhwz]

Disable the full live preview, try TAESD or the cheap versions in settings. Full decode is way too to slow.
That's probably your reason for slow generation.

[zheng0002]

does anyone had slow slow loading for SDXL models.... and the console freeze when tasked queue..... render doesnt output but freezes at 98%... even console show 100% completed..... on webui 1.51 with SDXL???

[dhwz]

You're running out of RAM? Did you check your pagefile settings?
And please use fp16 fixed VAE of you stuck at 98% you're running out of VRAM on VAE decode.
Or try Tiled VAE extension.

[shyt47]

3080fe 10gb 32gb ram . Ultra slow image generation with sdxl

[dhwz]

Use --medvram
Don't use --no-half or --no-half-vae or --precicion full
Change your windows pagefile size
Don't use live preview mode "full", use TAESD or one of the cheap live preview settings
Use the fp16 fixed VAE (https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/tree/main)

[yaronelh]

How do you install the fp16 fix, is it just placing a file?

[Mahdialemohammadi]

I have 4070 8G on my laptop but have the same problem it takes around two or three min to generate a single pic. what is the solution??

[dilectiogames]

I've posted here a possible explanation on what is going on. It worked for me so maybe you people get lucky also

#11418 (comment)