KAIST · VISUAL MEDIA LAB

EDGAR

Kwan Yun · MinHyeok Bang · Colette Suhjung Koo · Junyong Noh

ACM Transactions on Graphics

PaperCode

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

Repair a Gaussian avatar with identity-guided diffusion. Keep the improvements in 3D—and the original rendering pipeline.

Explore before & after How it works
THE SAME AVATAR. REFINED.01 / RGBAvatar
Original avatar → EDGAR-refined avatarSame motion. Same camera.

01 / SEE THE REFINEMENT

Same avatar.
Sharper details.

Compare the original and refined avatars. Pause, scrub, or move the divider to inspect the details.

Before → refined 3D avatar

Loading comparisons…

All EDGAR results are rendered from the refined 3D avatar.

02 / FROM IMAGES BACK TO 3D

The refinement stays
with the avatar.

Diffusion supplies the supervision. A shared Gaussian representation carries the result into new views and expressions.

01

Learn the failures.

Learn to repair avatar artifacts using paired synthetic failures and identity references.

02

Refine the views.

Refine rendered views with identity-guided diffusion.

03

Update one avatar.

Transfer the improvements into a shared Gaussian avatar, ready to render without per-frame diffusion.

THE SHARED-AVATAR UPDATERefined views → refined 3D avatar
The image refiner

Identity conditioning and expression supervision guide the image refiner.

Architecture of the identity-conditioned EDGAR image refiner.

03 / RESULTS

Better appearance.
A familiar rendering cost.

Reconstruction quality and rendering speed, before and after refinement.

Reconstruction quality

BeforeEDGARSelected settings · full results in the paper

THE COST IS PAID ONCE

Refinement runs once.
Rendering stays fast.

The refined avatar keeps the original renderer, with nearly unchanged rendering speed.

One-time refinement and final-avatar rendering
PipelineRefine onceBefore FPSEDGAR FPS
CAP4D5.04 min50.349.4
RGBAvatar7.96 min813.8826.3
GaussianAvatars16.60 min153.4156.1

Measured on an RTX 3090. See the paper for evaluation details.

Limitations

Geometry

Missing regions and severe structural errors remain challenging.

Expressions

Extreme poses and mouth interiors can retain artifacts.

Trade-offs

Improvements vary by metric and reconstruction pipeline.

THE PAPER

EDGAR.

Efficient Diffusion-Driven Gaussian Avatar
Refinement via Synthetic Failure

PaperCode
BIBTEX
@article{yun_edgar,
  title = {EDGAR: Efficient Diffusion-Driven
    Gaussian Avatar Refinement via Synthetic Failure},
  author = {Yun, Kwan and Bang, MinHyeok and
    Koo, Colette Suhjung and Noh, Junyong},
  journal = {ACM Transactions on Graphics}
}

Reference