Abstract

We propose a pipeline that leverages Stable Diffusion to improve inpaintingresults in the context of defurnishing -- the removal of furniture items fromindoor panorama images. Specifically, we illustrate how increased context,domain-specific model fine-tuning, and improved image blending can producehigh-fidelity inpaints that are geometrically plausible without needing to relyon room layout estimation. We demonstrate qualitative and quantitativeimprovements over other furniture removal techniques.

 

Quick Read (beta)