Is the CNF model trained on W or W+ space?

[datbu178]

Thank you for the great project. However I am a bit confused. In the paper, you write that CNF is trained to map from space Z to space W. However when looking at your code, I see that CNF maps from space Z to space W+. And can also use the inverse function of CNF to map from W+ to Z. So is it trained on W or W + space? Thank you.

[RameenAbdal]

We found that the CNF are easier to train on the W space as compared to the W+ space. This is because of the different projecting methods having variable projection errors and rendering the latents far off in the distribution. However, in the paper, we also test on the W+ space (e.g projected real images). Since these W+ latents suffer from the same problems, we modify the ranges and degrees of the edits to get the results.

[datbu178]

@RameenAbdal Thank you for the answer. So the code provided in this project is trained on W or W+ space? Can I use this code to edit real images (projected on W+ space)? Thank you.