TencentARC/VQFR

How to train FFHQ on 1024x1024 resolution?

Opened this issue · 0 comments

luser350 commented

Hi @guyuchao the current VQFR model is trained on 512x512 FFHQ but I want to train it on 1024x1024. There are some questions I would like to ask before proceeding with training since it is computationally expensive.

  1. Does the following changes to yml file are enough:
    a. out_size: 1024 at line 18 (datasets)
    b. in_dim:1024 at line 88 (network_sr, quantizer_opt)
    c. in_dim: 1024 at line 109 (network_g,quantizer_opt)
    d. out_size: 1024 at line 119 (network_d_global)
    e. out_size: 1024 at line 131 (network_d_main_global)

  2. Can I still use the 512x512 pre-trained models (paths in yml)

  3. If there are any other changes that I have missed, please help.