search

csmliu / STGAN

STGAN: A Unified Selective Transfer Network for Arbitrary Image Attribute Editing
MIT License
437 stars 84 forks source link

How to Improve the effect of High Resolution Image #16

Closed zhoushiwei closed 5 years ago

zhoushiwei commented 5 years ago

Hi, csmliu At present, the results of training on high-resolution images are not good. Is there any other way to improve the conversion results of high-resolution images, such as 512 or 1024 images?If I add a data set of high resolution images, will it improve the current effect at high resolution?

csmliu commented 5 years ago

Sorry I didn't try images with higher-resolution, but the model size may be a problem, and I have to reduce the number of channels when training the 384x384 model. BTW, are you using low-resolution images to train the model? Using extra images to train the model should lead to a better result, but I can't say how much they can help.

zhoushiwei commented 5 years ago

yean ,I also think that adding data and restrict the pose of the face should enhance the effect . But for the size of 512*512 image or larger, it seems difficult to train a robust model. If you make modifications at the algorithm level, do your have any good suggestions?

csmliu commented 5 years ago

Maybe you can refer to BigGAN or StyleGAN. Their architectures are designed for HR images, maybe you could get inspired by their work.

zhoushiwei commented 5 years ago

ok, Thanks,by the way,Is there a schedule for the pytorch multi-card version?

zhoushiwei commented 5 years ago

If you deepen the training network, do you think the effect will be better, or will it be more difficult to converge?

csmliu commented 5 years ago

If you train the model from stratch, deepen the network should result in a harder training. Yet you can initialize the model with the pretrained model, maybe this could help.