Problem occurres during reprojection.

[gsygsy96]

Hi @tetratrio, I tried to reproject LSUN images using run_projector.py. But I failed. Here are the input and the results. Using default parameters during reprojection. Could you help me figure out the problem? Thanks!

The input:

The results at step 200, step 800 and step 1600:

[adriansahlman]

Hey! What network weights are you using? "church-config-f"?
It does seem like its having a difficult time projecting back the input image. I have only tried the projection on FFHQ (faces) where it worked really well. I do think it will be harder to project other types of images as they are less structured than the faces (the faces are aligned where every image has the eyes in the exact same position e.t.c.). Can you try with some other pretrained weights and another input image?

The projection of real images will never be as good as that of generated images (generating an images and then running projection with the generated image as input will give very accurate results).

You can see how it can not recreate the real images, just create something that looks a bit similar.

The quality of the church dataset is also lower than both the cars and ffhq dataset as far as I remember.

If you reduce the weight for noise regularization you may get more accurate recreation of your input image but the latent vector projected will not be the reason for that, instead the noise will be the main factor for how the output looks.

Tell me if you get it to work better with some other model and input data or by changing some settings