Curiosity about upscaling GAN generated results

thunil / TecoGAN

This repo contains source code and materials for the TEmporally COherent GAN SIGGRAPH project.

Apache License 2.0

6k stars 1.14k forks source link

Curiosity about upscaling GAN generated results #54

Open 5agado opened 4 years ago

5agado commented 4 years ago

I generate or edit some works of mine via GAN or other generative models, but as they generally work on small resolution I wanted to rely on secondary models for the upscaling part. However I obtain very poor results.

See for example how this 256x256 image

Results in this poor 1024x1024 version

output_1

Curious to hear any explanations or speculation about this, and if there is any way around it.

caryknoop commented 4 years ago

It looks like it is chroma subsampling related. I would make sure your generated images from the video are sampled from 4:2:0 through 4:4:4 to RGB correctly.

rpolyano commented 4 years ago

I'm seeing similar results, and I think I'm sampling it correctly: ffmpeg -i 'video.avi' -vf format=rgb24 out_%05d.png anything obvious wrong with those args?