Question about resamplers in V2

[LinZhuoChen]

Thank you for your awesome work!

I want to know which type of resample is used in mogev2?

['pixel_shuffle', 'nearest', 'bilinear', 'conv_transpose', 'pixel_unshuffle', 'avg_pool', 'max_pool']

Looking for your reply

[pkqbajng]

You can find this by loading their pretrained checkpoint.

[EasternJournalist]

We use conv_transpose for the first three upsamplers and bilinear for the last one, as specified in the model configuration from the pretrained checkpoint:

"resamplers": ["conv_transpose", "conv_transpose", "conv_transpose", "bilinear"]

Note that these choices are not critical to the model’s performance. We experimented with different upsampling methods and attempted to search for the optimal combination, but the results showed only minor differences in accuracy and convergence. Therefore, the current configuration is selected mainly for its simplicity and consistency with MoGe-1 to reduce the confusion.