search

NVIDIA / pix2pixHD

Synthesizing and manipulating 2048x1024 images with conditional GANs
https://tcwang0509.github.io/pix2pixHD/
Other
6.64k stars 1.39k forks source link

Strange patterns in the output image #95

Open cfanyyx opened 5 years ago

cfanyyx commented 5 years ago

In some of my output images, there are strange patterns occurred like the pics below: image image image I can't figure out why those things came out upon my output images. Do anybody know the reason? In addition, I think my loss curves are also strange: image image image

mmreza79 commented 5 years ago

@cfanyyx , It is not clear what happened with your output unless share more information like which params you are using? what is your ground truth? Are you using any label map or any instance map? It would be better if you share your training command and if possible to share (optional) a sample expected ground truth and original image. Thanks.

cfanyyx commented 5 years ago

@mmreza79 Hi, thanks for replying. I used the default params and trained without label map or instance map. The targets of my training are something like beautifying pictures, removing spots from pictures.

ChristianEschen commented 5 years ago

I have seen the same kind of artifacts. I try to translate one rgb image into another rgb image. (so no label maps and instance maps). If anyone can interpret these artifacts please let me know.

yxt132 commented 5 years ago

same problem here too. my input is 4-channel RGBA and output is also 4-channel RGBA. No label map nor instance map. I used pretty much the same setting as default but I turned off VGG loss. Also, normalization of image is set to False. I got some color spots (artifacts) that are sporadic. Any clue?

YeHaijia commented 4 years ago

same problem , how to solve?

kex243 commented 3 years ago

Im not a professional, but for AtoB translation... When you are training with not normalised data, like if most part of images are bright and some are dark, algorythm tends to use more light colors at all features. Like training with blank white background and testing on image with colored background- it will seek for the lightest spot on the image and will tend to bleach it. IDK how to solve it, maybe enlarging dataset, normalising it. IMHO it is not connected to network params. Maybe additional layers of discriminators can help, at the beggining it causes additional artifacts, but with time it forces generator to get rid of some bugs.