Real time interactive streaming digital human, realize audio video synchronous dialogue. It can basically achieve commercial effects.
- Supports various digital human models: ernerf, musetalk, wav2lip
- Supports voice cloning
- Supports digital humans being interrupted while speaking
- Supports full-body video stitching
- Supports RTMP and WebRTC
- Supports video editing: plays custom videos when not speaking
Tested on Ubuntu 20.04, Python3.10, Pytorch 1.12 and CUDA 11.3
conda create -n nerfstream python=3.10
conda activate nerfstream
conda install pytorch==1.12.1 torchvision==0.13.1 cudatoolkit=11.3 -c pytorch
pip install -r requirements.txt
#If only using the musetalk or wav2lip models, there is no need to install the following libraries.
pip install "git+https://github.com/facebookresearch/pytorch3d.git"
pip install tensorflow-gpu==2.8.0
pip install --upgrade "protobuf<=3.20.1"
By default, the ernerf model is used, and WebRTC is used to stream to SRS.
export CANDIDATE='<Server network ip>'
docker run --rm --env CANDIDATE=$CANDIDATE \
-p 1935:1935 -p 8080:8080 -p 1985:1985 -p 8000:8000/udp \
registry.cn-hangzhou.aliyuncs.com/ossrs/srs:5 \
objs/srs -c conf/rtc.conf
python run.py
Open http://serverip:8010/rtcpushapi.html in a browser, enter any text in the textbox, and submit. The digital human will broadcast the entered text. Note: The server needs to open the port tcp:8000,8010,1985; udp:8000
Currently,where the LLM model supports ChatGPT, Qwen, and GeminiPro. You need to enter your own api_key in run.py.
Open with browser:http://serverip:8010/rtcpushchat.html
You can choose from the following two services, with gpt-sovits recommended
Service deployment referencegpt-sovits
run
python run.py --tts gpt-sovits --TTS_SERVER http://127.0.0.1:9880 --REF_FILE data/ref.wav --REF_TEXT xxx
The REF_TEXT is the voice content in REF_FILE, and the duration should not be too long.
Run the xtts service
docker run --gpus=all -e COQUI_TOS_AGREED=1 --rm -p 9000:80 ghcr.io/coqui-ai/xtts-streaming-server:latest
Then run it, where ref.wav is the voice file that needs to be cloned
python run.py --tts xtts --REF_FILE data/ref.wav --TTS_SERVER http://localhost:9000
If HuBERT is used to extract audio features during model training, use the following command to start the digital human
python run.py --asr_model facebook/hubert-large-ls960-ft
python run.py --bg_img bc.jpg
ffmpeg -i fullbody.mp4 -vf crop="400:400:100:5" train.mp4
Train the model with train.mp4
ffmpeg -i fullbody.mp4 -vf fps=25 -qmin 1 -q:v 1 -start_number 0 data/fullbody/img/%d.jpg
python run.py --fullbody --fullbody_img data/fullbody/img --fullbody_offset_x 100 --fullbody_offset_y 5 --fullbody_width 580 --fullbody_height 1080 --W 400 --H 400
- --fullbody_width, --fullbody_height are the width and height of the full-body video
- --W, --H are the width and height of the training video
- For the third step of ernerf training, the torso, if not trained well, seams may runear at the joints. You can add --torso_imgs data/xxx/torso_imgs to the command above. For the torso, instead of model inference, use the torso images directly from the training dataset. This method may result in some artificial marks at the neck and head.
- Extract images from a custom video
ffmpeg -i silence.mp4 -vf fps=25 -qmin 1 -q:v 1 -start_number 0 data/customvideo/img/%d.png
- Run the digital human
python run.py --customvideo --customvideo_img data/customvideo/img --customvideo_imgnum 100
This mode does not require srs
python run.py --transport webrtc
The server needs to open the ports tcp:8010 and udp:50000~60000. Open http://serverip:8010/webrtcapi.html in a browser.
-
Install the rtmpstream library
-
run srs
docker run --rm -it -p 1935:1935 -p 1985:1985 -p 8080:8080 registry.cn-hangzhou.aliyuncs.com/ossrs/srs:5
- Run the digital human
python run.py --transport rtmp --push_url 'rtmp://localhost/live/livestream'
Open http://serverip:8010/echoapi.html in a browser.
RTMP push is not supported for now
- Install the dependency libraries
conda install ffmpeg
pip install --no-cache-dir -U openmim
mim install mmengine
mim install "mmcv>=2.0.1"
mim install "mmdet>=3.1.0"
mim install "mmpose>=1.1.0"
- Download the model Download the model required to run MuseTalk from the following link: https://caiyun.139.com/m/i?2eAjs2nXXnRgr Access code: qdg2 After extraction, copy the files under the models directory to the models directory of this project. Download the digital human model from this link: https://caiyun.139.com/m/i?2eAjs8optksop Access code: 3mkt, and after extraction, copy the entire folder to the data/avatars directory of this project.
- run
python run.py --model musetalk --transport webrtc
Open http://serverip:8010/webrtcapi.html in a browser You can set --batch_size to improve GPU utilization, and set --avatar_id to run different digital humans
git clone https://github.com/TMElyralab/MuseTalk.git
cd MuseTalk
Modify configs/inference/realtime.yaml and set preparation to True
python -m scripts.realtime_inference --inference_config configs/inference/realtime.yaml
After running, copy the files from results/avatars to the data/avatars directory of this project.
Method 2:
Execute
cd musetalk
python musetalk.py --avatar_id 4 --file yourpath
Supports video and image generation, which will be automatically generated in the avatars directory under data
rtmp push is not supported for now
- Download the model
Download the model required to run wav2lip from the following link: https://pan.baidu.com/s/1yOsQ06-RIDTJd3HFCw4wtA Password: ltua Copy s3fd.pth to wav2lip/face_detection/detection/sfd/s3fd.pth in this project, and copy wav2lip.pth to the models directory in this project The digital human model file wav2lip_avatar1.tar.gz, after extraction, should be copied as a whole folder to the data/avatars directory in this project - run
python run.py --transport webrtc --model wav2lip --avatar_id wav2lip_avatar1
Open http://serverip:8010/webrtcapi.html in a browser You can set --batch_size to improve GPU utilization and set --avatar_id to run different digital humans
cd wav2lip
python genavatar.py --video_path xxx.mp4
After running, copy the files from results/avatars to the data/avatars directory in this project
can be replaced with models you have trained yourself(https://github.com/Fictionarry/ER-NeRF)
.
├── data
│ ├── data_kf.json
│ ├── au.csv
│ ├── pretrained
│ └── └── ngp_kf.pth
No need for previous installation, run directly.
docker run --gpus all -it --network=host --rm registry.cn-beijing.aliyuncs.com/codewithgpu2/lipku-metahuman-stream:vjo1Y6NJ3N
- GPU: Tesla T4
- Frame Rate: Approximately 18 FPS. Disabling audio and video encoding and streaming increases frame rate to about 20 FPS.
- Upgraded GPU: RTX 4090
- Achieved Frame Rate: Over 40 FPS
- Approach: Implement a new thread dedicated to audio and video encoding and streaming.
- Benefits: Reduces bottlenecks associated with single-threaded execution, enhancing frame rate.
- Duration: Around 3 seconds
- Issue: The edgetts component processes the entire sentence before outputting, causing significant delay.
- Optimization: Transition to a streaming input method for TTS, allowing real-time text processing as it is received.
- Issue: Requires buffering 18 frames of audio before computation.
- Optimization: Explore reducing buffer size or optimizing the model to process fewer frames without losing accuracy.
- Issue: Default buffering settings of the Simple Realtime Server (SRS) introduce latency.
- Optimization: Adjust the SRS server settings to minimize buffering delays, such as the
gop_cache
orqueue_length
.
1.Implemented ChatGPT for interactive dialogues with digital humans
2.Integrated voice cloning technology
3.Enabled video replacement for digital humans when muted
4.Implemented MuseTalk feature
5.Added Wav2Lip synchronization
6.Pending SyncTalk implementation
7.Enhance real-time interaction capabilities
8.Develop multi-language support for global accessibility
9.Incorporate advanced facial expression recognition
10.Introduce adaptive learning algorithms for better user personalization