a question about 3d-vae encoding

[qiwang1996]

HunyuanVideo is really a good open-source video generation model.
I have a question to ask and it maybe too easy for you to answer but it is troubling me :)

if i have five sequential frames in time: f1, f2, f3, f4, f5
i put them into your used 3d-vae to encode and then i get e1, e2 Later i put f1, only to encode and i get ee1,
does ee1 = e1 ?

[ttxxr]

I have the same issue. Since HunyuanVideo uses a 3D causal VAE, theoretically, ee1 should be equal to e1. However, I found in the tests that ee1 is not equal to e1.

[qiwang1996]

yes, i also conduct a test in comfyui and find e1 is not equal to ee1

[yzy-thu]

Because of GroupNorm