search

menyifang / ADGAN

The Implementation of paper "Controllable Person Image Synthesis with Attribute-Decomposed GAN" CVPR 2020 (Oral); Pose and Appearance Attributes Transfer;
475 stars 90 forks source link

[Q] Performance of #11

Open nitthilan opened 4 years ago

nitthilan commented 4 years ago

More of a question than an issue. Have you tested using attributes from images that do not belong to the dataset? I tried using the below picture for attributes bts_people_1

And the output I got was not proper. bts_output

I am not sure why this is happening. Have you experimented with something like this?

Thanking you.

Regards, K. J. Nitthilan

menyifang commented 4 years ago

Hi @nitthilan, we didn't try this way. The model relies on the training data and constructs a manifold space for it. I think valid synthesis results are actually are the mixtures of seen ones via the interpolation operation. If the provided images extremely deviate from the training data, the results will be not ideal.

geenie97 commented 3 years ago

Hi @nitthilan

Can I ask something?

I want to get Pose-transfer result like you

but when I using test.py I just got Style transfer I think...How Can I make pose-transfer output?