Gameleon
Gaussianizing Point Clouds for On-Demand 4D Performance

SIGGRAPH Asia 2026 Conference Papers

Zehong Li1, Jinyang Li2, Yueyu Hu3, Dandan Ding1,†, Zhan Ma2

1Hangzhou Normal University    2Nanjing University    3New York University

† Corresponding author

Paper Code

TL;DR: Gameleon directly decodes compressed 3D data into render-ready Gaussian primitives for real-time on-demand streaming, without reconstructing intermediate point clouds.

Gameleon system overview from 4D input through geometry and Gaussian decoding to free-viewpoint rendering

Overview

Comparison

Rendering visualization comparison: 8iVFB red_and_black (top) and THuman (bottom) across Ground Truth, G-PCC, Unicorn, B2P, and Gameleon.
Rendering Visualization Comparison: (a) Top row: 8iVFB Redandblack sample. (b) Bottom row: THuman sample. (c) Columns from left to right: Ground Truth, G-PCC (OpenGL/Gaussian), Unicorn (OpenGL/Gaussian), B2P (G-PCC), and Gameleon (Ours).

Visual Demonstrations

Static Rendering Showcase

The area below presents 4 frames of Gaussian PLY files decoded by Gameleon, encompassing datasets such as 8iVFB, THuman, and DanceNet3D. Static rendering is performed using SuperSplat. The viewpoint can be freely adjusted for observation.

THuman 0501
Open in a separate window
Basketball Player
Open in a separate window
Model
Open in a separate window

Dynamic Rendering Showcase

This subsection presents the outcomes of Gameleon's On-Demand 4D Performance, as executed by the prototype system we developed. Owing to limitations in web resources, the real-time processing workflow is demonstrated via recorded videos.