Question Regarding the Separate VAE Decoder for the 0.5B Model

[wafer9]

Hi,

I've noticed that while the VAE decoders for the 1.5B and 0.5B models share the same architecture, their parameters (weights) are different.

This leads to a few questions:

• Why was a new VAE decoder trained for the 0.5B model instead of reusing the one from the 1.5B model?

• Is this decision related to the semantic representation in the latent space? In other words, was the VAE retrained specifically to better align with the 0.5B model's semantic understanding?

• If it was indeed retrained, could you please provide some details on the changes made during the retraining process (e.g., differences in training data, configuration, or objectives) compared to the original VAE?

Thanks for your clarification!