Questions about the architecture of latent transparency decoder

Love your work, thanks for sharing the code.

I got some questions about the decoder part, could you kindly provide me with some hints or guidance on this aspect?

Q1: In Appendix B page 22, “Then the decoder goes through 64 × 64 × 512 →128 × 128 × 512 → 256 × 256 × 256 → 512 × 512 × 128 → 512 × 512 × 3”. should the last output be 512 × 512 × 4?

Q2: The input of decoder is $(x_a, \hat{I})$, how about just input a $x_a$, since it has all the decoded information. If it did not work, then why.

Q3: Is the U-net decoder a must? Is it because that it is too hard to reconstruct $\hat{I}_c$, where the backgrounds information in the premultiplied $I$ is discarded but also required to reconstruct.

Start with Appendix B, page 22, and compare its decoder dimensions and inputs with the implementation; the issue names no source file or test. Determine whether the three architecture questions are answered by the existing code and document the confirmed behavior and rationale for the decoder design.