This paper gates how much a model trusts its own geometry guess, based on how consistent several guesses are.
The problem: Single-image to multi-view diffusion models mix geometry and texture, so bad geometry can leak into the RGB output.
The method: Several geometry predictions are sampled for the same views, and a parameter-free gate (SCG) scores how much they agree to build a reliability map.
The injection: An RGI module uses that map to scale how strongly geometry residuals feed the RGB model. The authors say it works with different geometry generators.
Status: Accepted to ACCV 2026. The abstract claims better geometric quality, and code is linked.
Sign in to comment.
TL;DR of the paper, generated automatically.
This paper gates how much a model trusts its own geometry guess, based on how consistent several guesses are.