knazeri
color discrepancy
I tried your model and fine tune it on my won dataset (about 60k images for 20 epochs) It seems that the model suffers from color discrepancy.
input:
test on Places2:
test on fine-tune model:
The context of the mask is simple and clean, and there is no much edge need. However, the color discrepancy is very obvious. Do you have any idea why the model failed in such a simple case? Or do you have any suggestion about postprocessing?
Thanks
Do you think the batch normalization layers are harmful in such image generation task? Or it is ok for generation model to use such layers?
@1900zyh We were not using batch normalization unless you have changed the code! Our generators use instance normalization that does not change the statistics at inference time! And honestly, I don't think this has anything to do with normalization schemes!
What I see is that the mask is not a clean binary mask! As you can see in the zoomed version of your image. This normally happens with mask image being a .jpg file and not being entirely binary!
Hi @knazeri, I followed your model and code with instance normalization (I didn't modify it. ) As for the mask, it's a clean mask in fact, because I create it by using numpy not loading image. And I save the result as a jpg image (which you have seen above)
Do you have any other suggestions? Thanks!
@1900zyh Can you post a mask image and an actual image you are using here? Please make sure they are both .png images.