search

Gal4way / TPD

This is the official repository for the paper "Texture-Preserving Diffusion Models for High-Fidelity Virtual Try-On". CVPR 2024
114 stars 21 forks source link

High resolution training #18

Open trqminh opened 3 weeks ago

trqminh commented 3 weeks ago

Hi @Gal4way, Thank you for publishing your work.

I am using this repository and trying to train on higher resolution than the default config. In specific, I changed the image size to 1024x768. I am using the same config setup, I only change the batch size per gpu to 2 to avoid the out-of-memory issue. The training results look quite bad (I added some examples). 00006_00_grid 00008_00_grid

I am not sure what could be the problem so I would like to ask if you have any idea what the problem could be?

Thank you for your consideration.

Spritan commented 3 weeks ago

@trqminh i am also trying to do the same, if you dont mind can you tell me, what is is the vram consumption at batch size = 2. for 1024x768

Gal4way commented 6 days ago

I noticed that the model's predicted masks are inaccurate and the generated try-on images don't look good. I think this might be because the training steps weren't enough, so the model hasn't fully converged yet. The reduced batch size could also be a factor. But I recommend training for more epochs first to rule out the possibility of insufficient training.