MUA: Mobile Ultra-detailed Animatable Avatars

2026-04-20 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Graphics & Vision, Gaming & Interactive Media · Depth: Expert, quick

Summary

Heming Zhu, Guoxing Sun, and Marc Habermann introduce MUA (Mobile Ultra-detailed Animatable Avatars), a novel representation and distillation pipeline designed to create photorealistic, animatable full-body digital humans for resource-constrained platforms. Existing methods either achieve high fidelity with substantial server-class GPU computation or are lightweight but lack detail and suffer from artifacts. MUA bridges this gap by employing Wavelet-guided Multi-level Spatial Factorized Blendshapes, which transfers motion-aware clothing dynamics and fine-grained appearance from a high-quality teacher model into a compact, efficient representation. This approach achieves up to 2000X lower computational cost and a 10X smaller model size compared to the teacher model, while maintaining visually plausible dynamics and appearance. MUA demonstrates over 180 FPS on a desktop PC and 24 FPS natively on a Meta Quest 3.

Key takeaway

For developers building immersive applications requiring high-fidelity digital humans on mobile VR/AR platforms, MUA offers a significant advancement. You can now deploy visually rich, animatable avatars with real-time performance on devices like the Meta Quest 3, overcoming previous computational and size constraints. This enables more engaging and realistic user experiences in your next-generation applications.

Key insights

MUA enables high-fidelity, animatable avatars on mobile devices by distilling complex dynamics into an efficient, wavelet-guided representation.

Principles

• Combine multi-level wavelet decomposition with low-rank factorization.
• Distill high-quality avatar details into compact representations.

Method

The method uses Wavelet-guided Multi-level Spatial Factorized Blendshapes, coupled with a distillation pipeline, to transfer motion-aware clothing dynamics and fine-grained appearance details from a pre-trained ultra-high-quality avatar model.

In practice

• Achieves 2000X lower computational cost.
• Enables 24 FPS on Meta Quest 3.
• Reduces model size by 10X.

Topics

• Animatable Avatars
• Digital Humans
• Wavelet-guided Blendshapes
• Model Distillation
• Computational Efficiency

Best for: Research Scientist, AI Scientist, Computer Vision Engineer, AI Hardware Engineer

Related on AIssential

• 🌗Mobile Ultra-detailed Avatars🌗 👉Given skeletal poses and a virtual camera as inputs,... — MUA generates photorealistic, animatable, and hyper-detailed 3D clothed human avatars from skeletal poses and camera inputs.
• This INSANE Free NSFW Anime AI Model Is SO GOOD! 6GB VRAM! — Anima is a small, efficient anime AI model offering versatile image generation and rapid LoRA training with minimal VRAM.
• One-shot Compositional 3D Head Avatars with Deformable Hair — Decoupling hair from the face in 3D avatar generation significantly improves realism and dynamic behavior.
• Archon: A Unified Multimodal Model for Holistic Digital Human Generation — Archon unifies seven modalities for holistic digital human generation, addressing token explosion and enhancing control.
• GenLCA: 3D Diffusion for Full-Body Avatars from In-the-Wild Videos — GenLCA generates photorealistic 3D avatars from partial 2D video data using a novel visibility-aware diffusion training strategy.
• Google Vids gets AI upgrade with Veo and Lyria models, directable AI avatars — Google Vids integrates Veo 3.1 and Lyria AI models, plus directable avatars, for enhanced video and music creation.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.