LAM is a PyTorch-based framework for generating ultra-realistic, animatable 3D avatars from a single image in seconds. It targets researchers and developers building interactive 3D applications, offering fast cross-platform animation and rendering with a low-latency SDK for real-time chatting avatars.

How It Works

LAM leverages Gaussian Splatting for high-fidelity 3D avatar representation. The core innovation lies in its "Large Avatar Model" architecture, enabling one-shot creation and efficient animation. This approach allows for rapid generation and real-time performance, making it suitable for interactive experiences.

Quick Start & Requirements

• Installation:
  • Linux: git clone https://github.com/aigc3d/LAM.git && cd LAM && sh ./scripts/install/install_cu121.sh (or install_cu118.sh for CUDA 11.8).
  • Windows: Refer to the Windows Install Guide.
• Prerequisites: CUDA 11.8 or 12.1/12.8.
• Model Weights & Assets: Download via Hugging Face or ModelScope.
• Demo: Run python app_lam.py.
• Docs: https://github.com/aigc3d/LAM

Highlighted Details

• Reconstruction time: 1.4 seconds.
• Performance: 562.9 FPS on A100, 110+ FPS on Xiaomi 14 phone (Animating & Rendering).
• Features: Audio2Expression model for audio-driven animation, WebGL SDK for cross-platform interactive chatting avatars.
• Integration: Supports LLM, ASR, TTS for full conversational avatar experiences via OpenAvatarChat.

Maintenance & Community

The project is from Tongyi Lab, Alibaba Group. It has active development with recent releases of export features, WebGL SDK, and Audio2Expression. A roadmap is partially outlined with planned releases for larger models and cross-platform rendering.

Licensing & Compatibility

The repository does not explicitly state a license in the README. This requires clarification for commercial use or integration into closed-source projects.

Limitations & Caveats

The project is presented as a SIGGRAPH 2025 submission, implying it may be research-oriented and potentially subject to rapid changes or experimental stability. Specific model weights (LAM-large) are still pending release.