VideoReTalking
OpenTalker/video-retalkingReviewed by AI
Checked by RepoPicks Radar on
What is VideoReTalking?
VideoReTalking takes an existing video of a person talking and reworks the face so the mouth matches a different audio track, with an option to change the expression (neutral, smile, surprise, angry). It comes from a SIGGRAPH Asia 2022 paper by Xidian University, Tencent AI Lab and Tsinghua University. To run it you set up Python 3.8 and an NVIDIA card on CUDA 11.1, install ffmpeg, download the model files from a Google Drive folder and run one command. There is also a Colab notebook and a paid page run by other people.
VideoReTalking is free and open source (Apache-2.0) and can replace HeyGen, D-ID, and Rask AI.
How easy is it to install VideoReTalking?
B · installer plus one simple step · Web
- Windows D
- Linux D
- Web B
The gentlest way in is the web at grade B: the README offers a Colab notebook (quick_demo.ipynb) and a Replicate page, where you open the page and press run instead of installing anything. Installing it is D on Linux and Windows: `conda create -n video_retalking python=3.8`, `conda install ffmpeg`, then a graphics-card-specific PyTorch line the README itself limits ("This installation command only works on CUDA 11.1"), `pip install -r requirements.txt`, and a manual download of the pre-trained models into `./checkpoints` before `python3 inference.py` will run.
VideoReTalking on GitHub
- Stars
- 7,300on Nov 10, 2026
- Downloads
- 461,156945 this week
- Growth
- +1stars in 6 days
Stars over time
Facts
- Runs on
- Windows, Linux, Web
- You need to install
- Python, NVIDIA graphics card, FFmpeg
- To build from source
- Python, FFmpeg
- Disk space
- 100 MB–300 MB · estimate
- Accounts or keys
- None
- Cost
- Free · The code is free under Apache-2.0. The README also links a Replicate page run by someone else (cjwbw/video-retalking); services like that charge for the machine time, which the README does not explain.
- License
- Apache-2.0 (Permissive)
- Last change on GitHub
- Nov 4, 2023
VideoReTalking risk signals (9)
- Attentionstale releases
Last release on 2023-03-15 (1334 days ago)
automatic check
- AttentionDownload links to file-sharing sites
README links to https://drive.google.com/drive/folders/18rhjMpxK8LVVxf7PI6XwOidt8Vouv_H0?usp=share_link
automatic check
- AttentionDependencies with known vulnerabilities
Pinned dependencies with known vulnerabilities (OSV): basicsr: GHSA-86w8-vhw6-q9qq, PYSEC-2026-1215
automatic check
- Attentionmakes a real person appear to say other words
README: "a new system to edit the faces of a real-world talking head video according to input audio, producing a high-quality and lip-syncing output video even with a different emotion", with a "face enhancement network" to "improve the photo-realism". The project's own Disclaimer forbids "fraud, deception, infringement of others' portrait rights, reputation rights", but nothing in the code checks who is in the video or marks the result.
AI review
- Attentionthe model files come from a shared google drive
README Pretrained Models: "Please download our pre-trained models and put them in `./checkpoints`", pointing at a Google Drive folder (drive.google.com/drive/folders/18rhjMpxK8LVVxf7PI6XwOidt8Vouv_H0). Nothing runs without those files, the folder is outside GitHub, and no size, checksum or signature is given for them.
AI review
- Attentioncode frozen on an old software stack
The last commit is 2023-11-04 (1100 days before this check) and the instructions pin Python 3.8, `torch==1.9.0+cu111` and `dlib==19.24.0`, with 215 open issues against 7298 stars. Those versions are from 2021, so a current graphics driver or Python may not accept them and nobody has updated the recipe.
AI review
- Infonot an official tencent product
Disclaimer: "This is not an official product of Tencent", and item 5 adds that "This open-source code does not have the ability to directly provide services to end-users". Several authors are listed under Tencent AI Lab, but the repository is research code, not a supported product.
AI review
- Infobuilt on six other research projects
There is a `third_part/` folder and the Acknowledgement thanks Wav2Lip, PIRenderer, GFP-GAN, GPEN, ganimation_replicate and STIT "for sharing their code". Those pieces carry their own licenses and conditions, which the Apache-2.0 file in this repository does not cover.
AI review
- Infolimits stated by the authors
README: "You can test any talking face videos without manual alignment. But it is worth noting that DNet cannot handle extreme poses." The quality figures in the teaser come from the authors' own paper (arXiv:2211.14758).
AI review
Evidence is shown in its original language.
What Radar could not check about VideoReTalking
- Whether the recipe still installs today, given the CUDA 11.1 and Python 3.8 pins and three years without commits
- What exactly is inside the Google Drive checkpoints, since no checksum or file list is published
- Real disk use, as the checks estimate 100 to 300 MB from a heuristic and the model files are not counted
- How long one video takes on your graphics card, which the README never states
- How convincing the result looks on your own footage, especially with side views or fast movement