search

taki0112 / StyleGAN-Tensorflow

Simple & Intuitive Tensorflow implementation of StyleGAN (CVPR 2019 Oral)
MIT License
211 stars 60 forks source link

draw_uncurated_result_figure return different colors #6

Open nbl7 opened 5 years ago

nbl7 commented 5 years ago

Hello, and thanks for sharing this amazing repo.

I am having an issue while I try to generate images with draw_uncurated_result_figure

post_process_generator_output return images that seems overexposed.

Do you know why this is happening?

During the training the image generated are fine. It gets the correct color. But when I draw the colors are all incorrect.

Thanks a lot

test_fake_img_128_99_589

nbl7 commented 5 years ago

I should mention that I stopped the training and I am trying to draw on an intermediate checkpoint. Could that be the issue?

Thank you

nbl7 commented 5 years ago

Ok, yes. It seems the "problem" is that the checkpoint is at 64px while the final training should be a 128px.

If I change

images = self.sess.run(self.generator(latents, alpha=alpha, target_img_size=self.img_size, is_training=False))

to

images = self.sess.run(self.generator(latents, alpha=alpha, target_img_size=64, is_training=False))

it works.

Is there a way to know at what resolution is the checkpoint loaded?

Thanks!

rrquizon1 commented 3 years ago

Was this ever solved? THanks man!

I thought when the output written in current res is the current output of the images but it seems that it isnt. When will it reach the target resolution?