podgorskiy/ALAE

question about w

zhang-zx opened this issue ยท 7 comments

zhang-zx commented
question about w
zhang-zx commented

Thanks for your excellent work. As stated in your paper, your encoder generates styles code in w space. However, in your code, I think you use z space instead. Maybe there's some setting I overlooked. I hope you can help me understand it better.

  • ๐Ÿ‘2
  • ๐Ÿš€1
  • ๐Ÿ‘€1
huangzh13 commented

I have the same question.

csxmli2016 commented

same question.

5agado commented

I am also very interested regarding this aspect. Looking at the make_figures scripts, seems indeed you are just passing z to the model.generate function.
At the beginning I thought that the low fidelity of projected images was because of w tiling, but it would be a total different story if your encoder works only to z.

florian-hoenicke commented

I have the same question. Since w contains more information, using w should improve the reconstructions.

youngAt19 commented

I think actually the author is using w space. The discriminator is implemented as a component in encoder. And the two returning terms of model.encode is w and d's output
ๅ›พ็‰‡

podgorskiy commented

@youngAt19 is right.
Naming is misleading, that was a result of trying many configurations and things got their names after the writing started. I'll try to fix whatever I can without breaking compatibility with existing models