High-level instruction on how to use the LatentTransparencyOffsetEncoder model

Hi,
Thanks again for this awesome tech.
I see in several issues that Encoder support will come in the near future. Thank you for that.
In the mean time, if I want to use LatentTransparencyOffsetEncoder and test a few things out, what's the expected input?
From reading the decoder,
1. seems like the input should be alpha then rgb. Is this correct?
2. Are the input values [-1,1] or [0,1]
3. From my simple testing of autoencoding, e.g. LatentTransparencyOffsetEncoder(alpha, RGB) + sdvae.encode( masked_rgb ) -> decode, it seems like not adding the offset performs better. Is this expected?
Thanks again.

Start by reading the decoder and the LatentTransparencyOffsetEncoder entry point, then trace the sdvae.encode(...)->decode path described in the issue. Document the expected channel order, value range, and offset behavior, including a minimal usage example once these questions are verified.

Written by the indexing model from the issue text.