kayamin / DR-GAN

121 stars 38 forks source link

G_loss can't converge however the generated images aren't bad while training #6

Open timho102003 opened 6 years ago

timho102003 commented 6 years ago

Loss Curve: https://imgur.com/a/UrLpP Train Parameters: Batch Size=64 Train Images=360,000 lr=default beta1=default beta2=default

Generated Image while training: https://imgur.com/a/IasJP

Specify pose code with one input images to Generate different poses (0-180 degree, Image on the left is the input and the first and third rows are generated images while second and fourth rows are groundtruth ): https://imgur.com/a/WNPFA Another thing to be mentioned is that I've tried to input 45 degree face(and even frontal face) to generate poses from 0-180. The outputs are almost the same as the one generated from the 90 degree face(as in the link). Note:Input images are from MPIE