Clarification on Architecture Components in mPLUG-DOCOWL2

[FeliceSchena]

Hi, thanks for your effort in making the new mPLUG-DOCOWL2 open source. I'm currently trying to understand how it works. Referring to the image below, Section 3 states that there is a low-resolution vision encoder and an H-reducer. However, I can't find either of those components in the image, or they aren't depicted clearly. Could someone clarify the overall architecture? Apologies if this is a basic question!

[HAWLYQ]

Hi, @FeliceSchena , the low-resolution vision encoder and the HReducer are included in the High-resolution Visual Encoding module in this figure~