Performance difference

Question

Performance difference

Closed this issue 2 months ago · 5 comments

Thanks for your great work!

I tried to use your code to edit images on the PIE-Bench_v1 dataset, I found the reconstruction quality is much higher than that reported in the paper, but the edit performance is lower.

By the way, I use the run_SPDInv_P2P.py code to edit the Cat, but it is very different from your paper. Are there any environmental errors?

Answer 1 · 2024-05-09T03:58:36.000Z

Thanks for your attention.

However, The code seems to work fine with me. For this image, you can try to decrease the learning rate or reduce the ddim_step to 20 for better editing output.

Answer 2 · 2024-05-09T08:15:45.000Z

Thanks for your reply. Is the performance in Tab.1 based on the specific parameters of each image?

Answer 3 · 2024-05-12T07:32:21.000Z

We have updated the code and some typos, the performace in Tab. 1 is based the hyper parameters mentioned in our paper.

Answer 4 · 2024-05-14T10:08:18.000Z

Thanks for your reply, based on the updated codes, I still cannot obtain the same results reported in the paper.

Answer 5 · 2024-05-16T07:38:49.000Z

Note that the performance may be different due to the compression of jpg and the effects of random states in pytorch.