Question about Decoder Architecture

First of all, I appreciate your gratitude for sharing our impressive project publicly.

I have a question about a specific aspect of the decoder architecture, as mentioned in the paper, where image embeddings and tokens are used as separate queries for cross-attention. This structure seems to be unconventional, and I haven't seen it before. Could you explain the benefits of using this approach?

The issue asks for a conceptual explanation of the decoder architecture described in the paper, but names no implementation file, test, or concrete change. Start by reading the paper's decoder section and inspecting the repository's model code and example notebooks; done would be a documented explanation of why image embeddings and tokens are separate cross-attention queries.

Written by the indexing model from the issue text.