Generate g higher resolution images about consingan

[tohinz]

Hi, I suggest increasing the number of --trained_stages (default is 6) in general. This should already improve results for higher resolutions.
If this is not enough you can also increase the --lr_scale (default is 0.1 - you could try 0.5).
Finally, it might be necessary to train for more iterations on higher stages/resolutions and/or adapt other hyperparameters (receptive field of the discriminator/generator, learning rate, etc).

[yangyvhang666]

Thanks for your reply,I tried increase the --trained_stages and the lr_scale,but it might did't work,I want to know the --max_size and the --min_size should be setted to have some kind of proportional relationship?

[tohinz]

--min_size should be left at some small value, so I wouldn't change it or at least not increase it. --min_size and --max_size do not need to have a proportional relationship, rather --max_size and --trained_stages should be somewhat proportional, i.e. the larger --max_size the larger --trained_stages should be.
How high did you set --trained_stages to be? What exactly is the problem with the higher-resolution images? Is the global structure wrong or are there artifacts in the images?