ProregssivaAvatars: Progressive Animatable 3D Gaussian Avatars

CVPR 2026

Kaiwen Song, Jinkai Cui, Juyong Zhang
University of Science and Technology of China

ProgressiveAvatars introduces a progressive, animatable 3D Gaussian avatar representation that stays usable at low transmission budgets and continuously refines as more bandwidth, memory, or compute becomes available.

Abstract

In practical real-time XR and telepresence applications, network and computing resources fluctuate frequently. Therefore, a progressive 3D representation is needed. To this end, we propose ProgressiveAvatars, a progressive avatar representation built on a hierarchy of 3D Gaussians grown by adaptive implicit subdivision on a template mesh. 3D Gaussians are defined in face-local coordinates to remain animatable under varying expressions and head motion across multiple detail levels. The hierarchy expands when screen-space signals indicate a lack of detail, allocating resources to important areas. Leveraging importance ranking, ProgressiveAvatars supports incremental loading and rendering, adding new Gaussians as they arrive while preserving previous content, thus achieving smooth quality improvements across varying bandwidths. ProgressiveAvatars enables progressive delivery and progressive rendering under fluctuating network bandwidth and varying compute and memory resources.

Highlight

01

Single Asset

ProgressiveAvatars replaces redundant discrete LOD copies with one unified, streamable 3D asset that adapts to different bandwidth and transmission budgets.

02

Continuous LoD

Newly arrived Gaussians are added on top of existing content, enabling smooth quality refinement instead of discrete model switching.

03

Adaptive Subdivision

Screen-space gradients drive refinement toward high-frequency regions, avoiding the waste of uniform subdivision and using transmission and compute more efficiently.

Method

Starting from a tracked FLAME mesh, ProgressiveAvatars builds a mesh-anchored hierarchy of 3D Gaussians through adaptive implicit subdivision. New levels are introduced only when the current representation lacks detail, while face-local parameterization keeps the avatar stable and animatable during progressive growth.

Progressive Rendering

ProgressiveAvatars is designed to deliver useful renderings before the full asset arrives. Importance ranking decides which Gaussians are transmitted first, while adaptive growth reveals how the hierarchy expands toward detail-rich facial regions over training.

Importance Ranking

After training, ProgressiveAvatars assigns each face an importance score from its accumulated rendering contribution across views. In practice, faces whose bound Gaussians carry higher opacity and transmittance contributions receive higher priority, so transmission starts from the most visually influential regions. This importance-first ordering makes early partial renderings stay close to the final avatar instead of wasting bandwidth on low-impact details first.

Visualization of Training

Comparison

BibTeX

@inproceedings{song2026progressiveavatars,
  title     = {ProgressiveAvatars: Progressive Animatable 3D Gaussian Avatars},
  author    = {Song, Kaiwen and Cui, Jinkai and Zhang, Juyong},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2026}
}