HunyuanVideo-Avatar: High-Fidelity Audio-Driven Human Animation

Introduction

HunyuanVideo-Avatar is an advanced open-source project by Tencent-Hunyuan that enables high-fidelity, audio-driven human animation for multiple characters. Built upon a multimodal diffusion transformer (MM-DiT), this model excels at generating dynamic, emotion-controllable, and multi-character dialogue videos. It tackles key challenges in the field by ensuring strong character consistency, precise emotion alignment between characters and audio, and facilitating multi-character animation.

The project introduces three core innovations:

Character Image Injection Module: Replaces conventional addition-based conditioning to eliminate condition mismatch, ensuring dynamic motion and strong character consistency.
Audio Emotion Module (AEM): Extracts and transfers emotional cues from a reference image to the generated video, enabling fine-grained emotion style control.
Face-Aware Audio Adapter (FAA): Isolates audio-driven characters with a latent-level face mask, allowing independent audio injection via cross-attention for multi-character scenarios.

These innovations allow HunyuanVideo-Avatar to produce realistic avatars in dynamic, immersive scenarios, surpassing state-of-the-art methods.

Installation

To get started with HunyuanVideo-Avatar, follow these installation steps, primarily for Linux environments. CUDA versions 12.4 or 11.8 are recommended.

First, clone the repository:

git clone https://github.com/Tencent-Hunyuan/HunyuanVideo-Avatar.git
cd HunyuanVideo-Avatar

Then, set up the Conda environment and install dependencies:

# 1. Create conda environment
conda create -n HunyuanVideo-Avatar python==3.10.9

# 2. Activate the environment
conda activate HunyuanVideo-Avatar

# 3. Install PyTorch and other dependencies using conda
# For CUDA 11.8
conda install pytorch==2.4.0 torchvision==0.19.0 torchaudio==2.4.0 pytorch-cuda=11.8 -c pytorch -c nvidia
# For CUDA 12.4
conda install pytorch==2.4.0 torchvision==0.19.0 torchaudio==2.4.0 pytorch-cuda=12.4 -c pytorch -c nvidia

# 4. Install pip dependencies
python -m pip install -r requirements.txt
# 5. Install flash attention v2 for acceleration (requires CUDA 11.8 or above)
python -m pip install ninja
python -m pip install git+https://github.com/Dao-AILab/flash-attention.git@v2.6.3

For specific GPU types encountering float point exceptions, refer to the repository's README for additional solutions, including CUDA 12.4/CUBLAS/CUDNN updates or explicit CUDA 11.8 compiled PyTorch installations.

Alternatively, you can use the provided Docker images:

# For CUDA 12.4
docker pull hunyuanvideo/hunyuanvideo:cuda_12
docker run -itd --gpus all --init --net=host --uts=host --ipc=host --name hunyuanvideo --security-opt=seccomp=unconfined --ulimit=stack=67108864 --ulimit=memlock=-1 --privileged hunyuanvideo/hunyuanvideo:cuda_12
pip install gradio==3.39.0 diffusers==0.33.0 transformers==4.41.2

# For CUDA 11.8
docker pull hunyuanvideo/hunyuanvideo:cuda_11
docker run -itd --gpus all --init --net=host --uts=host --ipc=host --name hunyuanvideo --security-opt=seccomp=unconfined --ulimit=stack=67108864 --ulimit=memlock=-1 --privileged hunyuanvideo/hunyuanvideo:cuda_11
pip install gradio==3.39.0 diffusers==0.33.0 transformers==4.41.2

Pretrained models can be downloaded as detailed in the weights/README.md file within the repository.

Examples

HunyuanVideo-Avatar supports both parallel inference on multiple GPUs and single-GPU inference, including options for very low VRAM environments.

Parallel Inference on Multiple GPUs (e.g., 8 GPUs):

cd HunyuanVideo-Avatar

JOBS_DIR=$(dirname $(dirname "$0"))
export PYTHONPATH=./
export MODEL_BASE="./weights"
checkpoint_path=${MODEL_BASE}/ckpts/hunyuan-video-t2v-720p/transformers/mp_rank_00_model_states.pt

torchrun --nnodes=1 --nproc_per_node=8 --master_port 29605 hymm_sp/sample_batch.py \
    --input 'assets/test.csv' \
    --ckpt ${checkpoint_path} \
    --sample-n-frames 129 \
    --seed 128 \
    --image-size 704 \
    --cfg-scale 7.5 \
    --infer-steps 50 \
    --use-deepcache 1 \
    --flow-shift-eval-video 5.0 \
    --save-path ${OUTPUT_BASEPATH}

Single-GPU Inference:

cd HunyuanVideo-Avatar

JOBS_DIR=$(dirname $(dirname "$0"))
export PYTHONPATH=./

export MODEL_BASE=./weights
OUTPUT_BASEPATH=./results-single
checkpoint_path=${MODEL_BASE}/ckpts/hunyuan-video-t2v-720p/transformers/mp_rank_00_model_states_fp8.pt

export DISABLE_SP=1 
CUDA_VISIBLE_DEVICES=0 python3 hymm_sp/sample_gpu_poor.py \
    --input 'assets/test.csv' \
    --ckpt ${checkpoint_path} \
    --sample-n-frames 129 \
    --seed 128 \
    --image-size 704 \
    --cfg-scale 7.5 \
    --infer-steps 50 \
    --use-deepcache 1 \
    --flow-shift-eval-video 5.0 \
    --save-path ${OUTPUT_BASEPATH} \
    --use-fp8 \
    --infer-min

Run with very low VRAM (e.g., 10GB VRAM with TeaCache):

cd HunyuanVideo-Avatar

JOBS_DIR=$(dirname $(dirname "$0"))
export PYTHONPATH=./

export MODEL_BASE=./weights
OUTPUT_BASEPATH=./results-poor

checkpoint_path=${MODEL_BASE}/ckpts/hunyuan-video-t2v-720p/transformers/mp_rank_00_model_states_fp8.pt

export CPU_OFFLOAD=1
CUDA_VISIBLE_DEVICES=0 python3 hymm_sp/sample_gpu_poor.py \
    --input 'assets/test.csv' \
    --ckpt ${checkpoint_path} \
    --sample-n-frames 129 \
    --seed 128 \
    --image-size 704 \
    --cfg-scale 7.5 \
    --infer-steps 50 \
    --use-deepcache 1 \
    --flow-shift-eval-video 5.0 \
    --save-path ${OUTPUT_BASEPATH} \
    --use-fp8 \
    --cpu-offload \
    --infer-min

Run a Gradio Server:

cd HunyuanVideo-Avatar

bash ./scripts/run_gradio.sh

Why Use HunyuanVideo-Avatar?

HunyuanVideo-Avatar offers a robust solution for generating high-quality, dynamic human animations from audio input. Its key advantages include:

High-Fidelity and Dynamic Video Generation: Produces realistic and natural videos with high-dynamic foreground and background, preserving strong character consistency.
Emotion-Controllable Animation: Allows precise control over facial emotions, driven by input audio and emotion reference images.
Multi-Character Support: Capable of animating multiple characters simultaneously, making it ideal for dialogue videos and complex scene creation.
Versatile Avatar Styles: Supports a wide range of avatar images, including photorealistic, cartoon, 3D-rendered, and anthropomorphic characters, at arbitrary scales and resolutions (portrait, upper-body, full-body).
Optimized for Performance: Includes features like DeepCache and TeaCache for efficient inference, even on single GPUs with limited VRAM (as low as 10GB).
Broad Applications: Suitable for various downstream tasks such as e-commerce, online streaming, social media video production, and general video content creation and editing.