DDPM-IP icon indicating copy to clipboard operation
DDPM-IP copied to clipboard

[ICML 2023] official implementation for "Input Perturbation Reduces Exposure Bias in Diffusion Models"

Results 7 DDPM-IP issues
Sort by recently updated
recently updated
newest added

Thank you to the author for your work,but I'm curious how to train to get ADM-IP on the CiFar-10 dataset. The FID values I obtained during iterations were significantly higher...

Hi @forever208 , Great work! I like your observation of the inconsistency between training and sampling and propose a simple input perturbation to mitigate/align it. A question about your implementation...

Hello, may I ask how many steps you iterated to get the ADM_IP_015.pt you provided

I found an issue when I use ddim50 sampling aftrer training ddpm-ip on my own face dataset. The sampled images have much noise (however using ddpm50 is ok). I tried...

I run the cmd 'mpiexec -n 16 python scripts/image_train.py XXXXXX' while only has one gpu. and got some error. have anybody tried it with one gpu?

例如guided-diffusion给出的pt文件。 另外我想请问readme里面没有给出256的训练,请问是在256*256大小训练效果不是很好吗? 谢谢