LLaVA Vision Model's Feature Resolution and Dense Features Availability

[pengsongyou]

Hello LLaVA Community,

Firstly, thanks for the incredible series of LLaVA works! While I have yet to experiment with it personally, I found it highly relevant to my current project.

I have a specific query regarding the LLaVA visual encoder. Assuming I input an image with a resolution of 256x256, could you clarify what the resolution of the language embedding tokens H_v would be (as shown in the architecture below)? Is it typically reduced by factors like 4x or 8x compared to the original image resolution, or does the model generate a single global embedding?

Furthermore, I am curious if any models within the LLaVA family are capable of providing such dense feature representations?

Thank you so much in advance for your help!

Best regards,
Songyou