search

Ldhlwh / DFMGAN

[AAAI 2023] Official implementation of "Few-Shot Defect Image Generation via Defect-Aware Feature Manipulation".
https://arxiv.org/abs/2303.02389
Other
90 stars 12 forks source link

MVTec Zipper Problem #51

Open xxg1413 opened 4 weeks ago

xxg1413 commented 4 weeks ago

fakes_init

I'm training the zipper dataset, and in the third step, the init images are inconsistent. Could this be the https://github.com/Ldhlwh/DFMGAN/issues/22 cause? Does changing single-channel images to RGB affect training and inference?

Below are the images generated from inference.

24_img

Do you have any other suggestions?

Ldhlwh commented 16 hours ago

in the third step, the init images are inconsistent

About the third step, do you mean stage 2? Did you run stage 2 when generating these images?

Could this be the #22 cause?

No. I think the images of MVTec AD are already in RGB format.