EDGAR: Efficient Diffusion-Driven Gaussian Avatar Refinement via Synthetic Failure
EDGAR uses corrected views to improve the underlying head avatar, so refinement need not run on every output frame.
The system trains on real and synthetic examples that pair degraded avatar renderings with ground-truth and reference images. Its one-step diffusion model uses identity features and a mixture-of-experts architecture to handle different artifacts and severity levels. Across three avatar reconstruction methods and three settings each, the authors report better perceptual and reconstruction scores than raw outputs and an SDEdit-style baseline. Tests also show improved cross-view consistency and robustness to large head turns and open-mouth expressions, within the limits of recoverable appearance artifacts.
HF Daily Papers' note
The system trains on real and synthetic examples that pair degraded avatar renderings with ground-truth and reference images. Its one-step diffusion model uses identity features and a mixture-of-experts architecture to handle different artifacts and severity levels. Across three avatar reconstruction methods and three settings each, the authors report better perceptual and reconstruction scores than raw outputs and an SDEdit-style baseline. Tests also show improved cross-view consistency and robustness to large head turns and open-mouth expressions, within the limits of recoverable appearance artifacts.
HF Daily Papers' note
score 4