Wang Duomin

just like what i ask in title, in paper author said "gradients can be propagated backward from the output features to the input features, but not to the input proposal...

about codebook

1

so in the whole architecture, no explicit codebook vecters are used right? only categorical logits as the input to the decoder when you train your dvae?

About uv_weight_mask

1

Hi, Thanks for your awsome work! I am now confused in the generation of uv_weight_mask. Did you generate it using uv_kpt_ind.txt? And is uv_kpt_in.txt generated from Model_keypoints.mat? How can i...

why some end second are negative value?

2

just like the picture shows, some end seconds are negative, is there something wrong? besides i found some annotations has the same start_sec and end_sec, maybe it's also a mistake

does the beta weighted wrong loss term of embedding loss which should be commitment loss in the vanilla va-vae?

about rasterization for T_uv

1

Hi, this is a good work! Can you explain the process of rasteriazation for T_uv in more detail?

为什么3D VAE的结果仔细看是由64*64的网格组成的？

4

 如图所示，直接使用3d vae重建sora的example，会发现结果是64*64的patch组成的，重建512*512的视频会有8*8个patch，1024*1024的视频会有16*16个patch。我找遍了code也没有发现哪里有patch的构建，64*64的patch对应到latent上应该是8*8个latent为一组进行处理，可代码中并没有这个操作。