TL;DR of the paper, generated automatically.

A diffusion model that cleans up artifacts in 3D head avatars.

• Method: one-step diffusion with identity features from Sapiens, plus Mixture-of-Experts for degradation types.
• Use: its cleaned renders train a single avatar, so EDGAR doesn't run on every frame.
• Results: the paper says it improves on raw outputs and an SDEdit baseline across three avatar methods.