DreamPolisher icon indicating copy to clipboard operation
DreamPolisher copied to clipboard

Published 20 hours ago • verified publisher icon yuanze-lin

Camera Encoder and Geo. Loss

Open violag opened this issue 10 months ago • 0 comments

Hi, thank you for the amazing work!

I had a couple follow-up questions.

  1. Can you share some details about the camera encoder architecture?
  2. In your stage 2 (Appearance Refinement stage), the paper mentions that the geometric loss is computed on top of the views generated by the ControlNet refiner. Does this mean that you're doing an L2 loss between the outputs of the ControlNet refiner and the stage 1 renders? Or are you computing ISM loss with the ControlNet 1.1 tile model, similar to how you compute ISM loss through the SD 2.1 model in stage 1?

violag avatar Apr 01 '24 01:04 violag