Gaussian splatting solved the “photorealistic head avatar from ordinary video” problem faster than almost anyone expected a few years ago — but monocular, single-camera capture still leaves the underlying geometry and texture under-constrained. Turn a Gaussian-splat avatar’s head too far, or open its mouth too wide, and the gaps in what the camera actually saw start to show as blur, stretching, and frame-to-frame instability. EDGAR (Efficient Diffusion-Driven Gaussian Avatar Refinement), accepted to ACM Transactions on Graphics, tackles that failure mode directly — not by capturing better data, but by training a model specifically on the ways these avatars already break.
Training on purpose-built failure
The team’s key move is EDGAR-SET, a hybrid dataset of real-captured and synthetic-rendered triplets that deliberately pairs degraded avatar renderings with their ground-truth and reference counterparts — in other words, a training set built by intentionally simulating the artifacts a Gaussian avatar produces under stress, rather than hoping enough real failure examples turn up naturally. EDGAR then refines degraded frames through a fusion block that injects identity features from the self-supervised Sapiens model into a one-step diffusion backbone, correcting artifacts while explicitly preserving the subject’s actual likeness — a one-step process chosen specifically to keep the correction cheap enough to use in practice rather than turning refinement into its own expensive render pass.
One model, many kinds of broken
Because avatars fail in different ways depending on pose, lighting, and capture quality, EDGAR uses a Mixture-of-Experts architecture to handle different degradation types and severities rather than a single generic cleanup pass. Tested across three different Gaussian avatar reconstruction methods under three separate settings each, the refined outputs beat both the raw renders and an SDEdit-style refinement baseline on perceptual and reconstruction measures, with specific gains in cross-view consistency and robustness under large yaw rotation and wide-open-mouth expressions — exactly the extreme poses that break monocular capture hardest.
The efficiency trick: refine the avatar, not every frame
Rather than running EDGAR on every single output frame forever, the refined multi-view renderings are used as pseudo-targets to optimize the underlying Gaussian avatar itself once. That means the correction gets baked into the avatar model directly — after one refinement pass, the avatar renders cleanly on its own, without needing the diffusion model in the loop at render time. For anyone building avatar pipelines for games, virtual production, or telepresence, that’s the detail that matters: a fix applied once during asset creation rather than a cost paid on every single frame rendered afterward.