PDillis
commented
7 months ago Since I haven't used PTI, I can tell you where that error comes from and how to find where the code fails: in StyleGAN1/2, the mapping network $f$ or G.mapping will take a random latent z ($z\in\mathbb{R}^{512}$) and will output a disentangled latent w ($w\in\mathbb{R}^{1\times n\times512}$); for unconditional models, you simply do w = G.mapping(z, None). The disentangled latent w is the one you wish to find to do the editing with DragGAN (using either simple inversion or PTI), whose dimension $n$ will depend on the image resolution of your dataset/size of images that will be generated.
Concretely, StyleGAN expects two sections of the disentangled latent per block resolution in the synthesis network $g$ or G.synthesis (which starts from 4 and goes up by powers of 2 up until your final output resolution; more info in the StyelGAN architecture). So, from the AssertionError you posted above, it seems like PTI is giving you a disentangled latent of shape [1, 18, 512] whereas the network you are training is expecting a disentangled latent of shape [1, 12, 512]. In other words, PTI has hard-coded an image resolution of 1024 ($n=18$) whereas your StyleGAN2 model has a resolution of 128 ($n=12$).
I could be wrong and be the other way, so it's always helpful to tell us which code you ran and which line gave the AssertionError above, otherwise all we can do is guess.
Yurains
Thank you for your outstanding work I am training my own model using StyleGAN2 ada pytorch and importing other photos with PTI but I encountered an "AssertionError: Wrong size for dimension 1: got 18, expected 12" issue This seems to be a dimension-related problem, but I'm not sure how to resolve it Is there a way to make the necessary changes?