Higher and higher resolution

[wanxiayushaonian]

Hello ,

Thank for your great work on LFM! It's very impressive.

I have a question regarding the pre-trained models. In your paper/repository, you mentioned that the model was pre-trained on ImageNet at 224x224 resolution. I was wondering if you have also released (or plan to release) a version that was pre-trained on ImageNet at a higher resolution, specifically 512x512?

My use case requires higher input resolution, and having such a pre-trained model would be very helpful.

Thank you for your time and contribution!

Best regards,

[hao-pt]

Hi wanxiayushaonian,

In this repo, we mainly use 256x256 as default resolution on most datasets and only use 512x512 for CelebA data. Besides, just for clarification, on ImageNet, we resize all images to 256x256 which is a standard resolution for all generative models these days. Unfortunately, due to compute limit, we do not have ImageNet on 512x512 resolution. That said, we have a follow-up work on distilling StableDiffusion to few-step generation that works on 512x512 resolution (https://arxiv.org/abs/2412.16906). We will release code soon. Stay tune!

Best.

[wanxiayushaonian]

Thanks again for your guidance,I will pay attention.