empty.py - Githubissues

ChenWu98 / cycle-diffusion

[ICCV 2023] A latent space for stochastic diffusion models

Other

575 stars 35 forks source link

empty.py #33

Open beiqiqi opened 5 days ago

beiqiqi commented 5 days ago

"Could you provide the empty.py script of the translate_ffhq256_to_celeba256 project?"

ChenWu98 commented 5 days ago

Hi! Could you provide more details about what you’ve run and the error messages?

beiqiqi commented 5 days ago

Hello, thank you for your response! I want to use a latent diffusion model for unpaired image-to-image translation. I have the following two questions:

Do I need to train two separate latent diffusion models for the two types of images? If so, which project should I use for training? In the cycle-diffusion project, how do I input these two datasets? Should they be input in the empty.py script? How does cycle-diffusion distinguish between the two datasets?

ChenWu98 commented 5 days ago

Hi! Unfortunately, I cannot provide actionable suggestions because we didn't include this experiment in the paper, but here are some high-level ideas:

If you have two domains, yes you need to train two latent diffusion models. You may find https://github.com/CompVis/latent-diffusion helpful!
To use with the repo: the translate_ffhq256_to_celeba256 config is a good place to start! I think empty.py is shared for different scripts -- so if you set the config files correct, it should work with some little tweaks.

Disclaimer: I haven't tested this part as it's not included in our paper.

beiqiqi commented 5 days ago

Alright, no problem, I just wanted to try it out. So, how should I pass images from these two domains to cycle-diffusion? Just like the cat and dog experiment in the paper, where do I input the cat images and dog images? I understand that both empty.py and afhqcat256.py in the project are related to image input, but I’ve never quite understood how this works.