EDGAR: Efficient Diffusion-Driven Gaussian Avatar Refinement via Synthetic Failure

Researchers present EDGAR, a diffusion-driven method for refining Gaussian-splatting head avatars built from monocular video or images, targeting artifacts such as blurring, stretching and temporal instability caused by under-constrained geometry and texture. The work addresses the limits of recent Gaussian splatting approaches that have made photorealistic head avatar reconstruction efficient but not robust to sparse input constraints.

Recent advances in Gaussian splatting have enabled efficient and photorealistic head avatar reconstruction, yet monocular video-based or image-based methods still suffer from under-constrained geometry and texture, leading to artifacts such as blurring, stretching, and temporal instability. We prese