Reproduce “SqueezeMe: Mobile-Ready Distillation of Gaussian Full-Body Avatars” on OpenResearch

About this paper

Forrest Iandola, Stanislav Pidhorskyi, Igor Santesteban, Divam Gupta, Anuj Pahuja, Nemanja Bartolovic, Frank Yu, Emanuel Garbin and 2 more

Researchers from Meta Reality Labs developed SqueezeMe, a framework that distills high-fidelity 3D Gaussian full-body avatars for efficient real-time rendering on mobile-grade hardware. This approach allows three full-body avatars to animate and render simultaneously at 72 FPS on a Meta Quest 3, reducing decoder latency from an estimated 50 ms to 0.45 ms.

Key insights

• Pose-dependent corrective parameters for Gaussian avatars exhibit low-frequency characteristics, allowing them to be effectively shared among nearby Gaussians without substantial loss of visual quality.
• Complex non-linear neural networks driving avatar animation can be accurately approximated and distilled into much simpler linear models for efficient deployment on resource-constrained mobile NPUs.
• A two-stage approach—training a high-capacity model first, then distilling it—is crucial for achieving both high fidelity and stability in the simplified mobile-ready models.
• Strategic optimization for mobile hardware, including specialized processing units (NPUs) and low-level graphics APIs (Vulkan), is essential for enabling real-time performance of photorealistic avatars.

Summary by alphaXiv.