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

EDGAR. Efficient Diffusion-Driven Gaussian Avatar Refinement via Synthetic Failure. By Kwan Yun and three others, from KAIST. Published on the fifth of November.

Gaussian splatting has made head avatar reconstruction both efficient and photorealistic. The trouble starts when the input is thin. Methods that build an avatar from a single monocular video, or from images, work with geometry and texture that are under-constrained, and the gaps show up as artifacts. Faces blur. Regions stretch. Frames flicker against one another, so the avatar is temporally unstable. None of these is a failure of the representation as such. They are what happens when the evidence does not pin the avatar down.

EDGAR, short for Efficient Diffusion-Driven Gaussian Avatar Refinement, treats those artifacts as something a conditional diffusion model can learn to correct. Its job is narrow: take a degraded rendering, remove the artifacts, and keep the person recognisably the same person. Identity fidelity is the constraint the whole design is organised around.

To learn that job, it needs examples of failure. The authors build EDGAR-SET, a hybrid triplet dataset mixing real-captured and synthetic-rendered samples. It simulates realistic Gaussian avatar failure cases: each degraded rendering is paired with a ground-truth image and a reference image. That is the synthetic failure in the title. Instead of waiting for broken avatars, the training data manufactures them.

Refinement runs through a one-step diffusion backbone. A fusion block injects identity features into it, and those features come from Sapiens, a self-supervised model. The single step is what keeps the refinement efficient, and the identity features are what keep it from drifting away from the subject. Because the damage varies in kind and in severity, the model also carries a Mixture-of-Experts architecture, meant to handle that diversity of degradation types and levels.

The last step changes what the refiner is for. Running a diffusion model on every frame an avatar ever renders would be costly. So EDGAR refines multi-view renderings once, and those refined images become pseudo-targets for optimizing a single underlying Gaussian avatar. The corrected avatar is then rendered directly, with no refinement applied at output time.

Testing spans three different Gaussian avatar reconstruction methods, each evaluated under three different settings. In each case, perceptual and reconstruction measures improve over the raw outputs, and over an SDEdit-style refinement baseline. Cross-view evaluation shows improved calibrated consistency between viewpoints. Two harder conditions probe robustness: large yaw, where the head turns well away from the camera, and open-mouth expressions. The results hold up under both.

The authors frame their results modestly, as indicating a useful operating regime for recoverable appearance artifacts. That phrasing marks the boundary of the claim. EDGAR is a repair step for damage that can be undone, trained on failures made to order, and then folded back into the avatar so the repair costs nothing at render time.