lower quality image than pix2pix in lower-resolution(256*256) dataset

[zoeyxu59]

i trained both pix2pix and pix2pixhd with my own dataset(default param and networks)，image size is 256*256, pix2pixhd generates better image on train dataset but strange and blur result on test dataset. should I modify the G_net and D_net structure for lower resolution input image?

[songyn95]

@zoeyxu59 @aaronp24 @lukeyeager @javl @elezar
I encounter the same problem. The result of my training is very good, and the test result is very blur. *There are 320 training sets in total, and the data resolution is 25601440.** The effects of 200, 300 and 400 iterations are tested respectively. The results are very blur. How can I solve them? Is it because my scenario is relatively simple and there is less data? Looking forward to your reply！！

[zoeyxu59]

@songyn95 in my case ,accuracy improved with bigger training sets (with data cleaning ,sufficient pics in each category ) .Try coarse-to-fine training, it works when the dataset is large enough

[songyn95]

@zoeyxu59 How much training data do you have? Can you provide your training and test commands?