search

AUTOMATIC1111 / stable-diffusion-webui

Stable Diffusion web UI
GNU Affero General Public License v3.0
139.15k stars 26.41k forks source link

Performance regression in image2image #7

Closed DustyWantingMicrowave closed 2 years ago

DustyWantingMicrowave commented 2 years ago

Using the default image, same settings (I think?), and the following prompt: A fantasy landscape Steps: 50, Sampler: k-diffusion, CFG scale: 7, Seed: 2057480064

It takes 8 seconds to generate on my RTX 3060 Ti for this version of the script. before

But it takes several magnitudes longer with the latest version? after

AUTOMATIC1111 commented 2 years ago

This must be something on your side. I tested this locally and I get same speeds for both versions.

old: 100%|██████████| 50/50 [00:04<00:00, 10.70it/s]

new: 100%|██████████| 50/50 [00:04<00:00, 10.74it/s]

DustyWantingMicrowave commented 2 years ago

Any ideas on what it could be? I followed the guide here yesterday.

Swapping the script version back and forth reliably reproduces the same speed differences for me.

AUTOMATIC1111 commented 2 years ago

There really isn't anything relevant except image resizing modes, and the one selected by default works the same as in the version you linked. Are you sure you are reliably reproducing this with all same settings for image generation?

AUTOMATIC1111 commented 2 years ago

Actually, did you refresh the web page after restarting the script version of a different version?

DustyWantingMicrowave commented 2 years ago

Yup, refreshing the page and same settings that I can see.

I think I figured out the problem - disabling gfpgan seemed to fix the speed difference. Curiously, I don't have it turned on for image2image, and it seems to work fine when its on for text2image.

Probably installed it wrong, will reinstall it using this repo's method.

khalidmarescalchi commented 2 years ago

www.google.com