search

Darkbblue / diffusion-feature

Use pretrained Stable Diffusion model to extract features for various vision tasks
3 stars 0 forks source link

Test on image classification tasks #1

Open Yeez-lee opened 7 months ago

Yeez-lee commented 7 months ago

Hi, have you ever tried any evaluations on image classification tasks? How about the results?

Darkbblue commented 2 months ago

Hi, have you ever tried any evaluations on image classification tasks? How about the results?

Sorry for the late reply. Didn't expect anyone to see this repo... I've tested it on CIFAR10 and the result is ~94, not very comparable with true SOTAs but I'm usinga a very simple downstream model. I have to say it's not a good idea to use diffusion feature for image-level tasks. You change the original image input of 3 x h x w to 3000+ x h/8 x w/8 for some small improvement, which is not worthy. I'd recommend focusing on pixel-level tasks such as semantic segmentation instead.

hritam-98 commented 2 months ago

Hi, did you try it on image segmentation?