Abnormal FID metrics of diffusion-based models

[LidongWang-26]

I notice that diffusion-based models perform much worse than GAN-based models on FID metric in the task of Day-Night conversion, and it seems that the visual results of InstuctPix2Pix have some problem of brightness, could you please explain this abnormal phenomenon? Thanks!

[yassineAlouini]

Could you share some metrics you are getting and some generated results? For how long do you train the model?