Open-source research framework for real-time, audio-driven avatar video generation with streaming, infinite-length output built on a 14B diffusion model.
LiveAvatar generates talking avatar videos from an audio track and a reference image, with an optional text prompt. It is designed for streaming use: frames are produced continuously rather than as a single fixed-length clip, using a block-wise autoregressive approach that the project says supports videos over 10,000 seconds long.
The system is built on the Wan2.2-S2V-14B base model with a LiveAvatar LoRA, and uses 4-step sampling via distribution-matching distillation. The project reports about 45 FPS on multi-card H800 GPUs. Inference scripts cover multi-GPU real-time streaming, single-GPU offline generation on an 80GB card, and Gradio web UIs. Optional FP8 quantization targets 48GB GPUs, and compilation and cuDNN attention options are available for speed.
The project is released under Apache-2.0 and is described by its authors as a research preview. Model checkpoints are hosted on Hugging Face, and the repository lists pending work such as training code and TTS integration.
Create and activate a conda environment
conda create -n liveavatar python=3.10 -y && conda activate liveavatarInstall PyTorch and FlashAttention, then Python requirements
pip install torch==2.8.0 torchvision==0.23.0 --index-url https://download.pytorch.org/whl/cu128 && pip install flash-attn==2.8.3 --no-build-isolation && pip install -r requirements.txtDownload the Wan2.2-S2V-14B base model and LiveAvatar LoRA into ./ckpt/
huggingface-cli download Wan-AI/Wan2.2-S2V-14B --local-dir ./ckpt/Wan2.2-S2V-14B && huggingface-cli download Quark-Vision/Live-Avatar --local-dir ./ckpt/LiveAvatarRun single-GPU CLI inference (needs 80GB VRAM)
bash infinite_inference_single_gpu.shSummary generated by AI from the project's README. GitHub stats refresh daily.
Open-source large-scale video generative models for text-to-video, image-to-video, speech-to-video and character animation.
Open-source interactive game video generation model that turns a reference image and keyboard or mouse inputs into controllable gameplay-style video.
A node-graph GUI, API and backend for running diffusion models locally to generate images, video, audio and 3D content.
A serving framework that extends vLLM to run omni-modality models producing text, image, audio, video and action outputs.
Open-source interactive world model that renders playable game scenes from low-poly frames using streaming neural rendering.
A Gradio-based web UI for running Stable Diffusion and newer image/video models, continuing the Forge fork with added model support and optimizations.