LivePortrait: a two-mode PyTorch portrait animator where Animals mode drags in extra dependencies
Bring portraits to life!
At a glance
- What is it?
- LivePortrait is the official PyTorch implementation of a Kuaishou paper on portrait animation with stitching and retargeting control. Humans mode is the light path; Animals mode needs X-Pose, and the license is not machine-detected.
- Who is it for?
- Use LivePortrait if you want to run an official portrait-animation implementation locally and can set up conda, FFmpeg and, for Animals mode on Linux or Windows, X-Pose with a matching CUDA. Skip it if you need a tagged release to pin or a clearly detected license, because the repository has neither a GitHub release nor a license GitHub recognises.
- Can I use it commercially?
- Check first. The repository uses a licence we do not classify automatically, so read its LICENSE file before any commercial use.
- Is it still maintained?
- Yes. The repository last received commits 167 days ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- KlingAIResearch/LivePortrait19,491 ★ · +408
| Repository | 2026-09-21 | 2026-11-15 |
|---|---|---|
| KlingAIResearch/LivePortrait | 19.1k | 19.5k |
408 Stars in the last 55 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
Animals mode needs X-Pose on Linux and Windows, Humans mode does not
LivePortrait ships two modes, and they do not cost the same to install. The README says X-Pose, required by Animals mode, is a dependency that must be installed on Linux and Windows. The step titled Check your CUDA versions is optional if you only want to run Humans mode.
The repository mirrors this split in its scripts. There are inference.py and app.py for humans, and inference_animals.py and app_animals.py for animals. That separation has a consequence: a working Humans setup tells you nothing about Animals, and the second may fail on a CUDA mismatch that the first never touched.
The project's own history shows the animals side moved quickly. An Animals model was released on 2024/08/02, and an updated version with more data followed on 2025/01/01. If you downloaded weights before then, you may be on the older one.
Three tools have to exist before the first command runs
The setup begins with a note: make sure your system has git, conda and FFmpeg installed. FFmpeg gets its own guide in the repository, assets/docs/how-to-install-ffmpeg.md. Then the README gives the environment commands:
git clone https://github.com/KlingTeam/LivePortrait
cd LivePortrait
# create env using conda
conda create -n LivePortrait python=3.10
conda activate LivePortraitThe clone address uses the KlingTeam organisation, while this repository is listed under KlingAIResearch, so check you are cloning the one you mean to. The environment is pinned to Python 3.10, which matters if you were hoping to reuse an existing newer interpreter. For Linux and Windows the README then opens a collapsible step that asks you to check your current CUDA version with nvcc -V, giving 11.1, 11.8 and 12.1 as example versions. That check exists because the GPU build of onnxruntime and X-Pose both depend on the CUDA installed on the machine, so skipping it can leave you debugging a library mismatch after the long downloads have already finished. On a machine with no NVIDIA GPU, this route offers no fallback, and the Apple Silicon path is the documented alternative.
Pinned onnxruntime-gpu and transformers decide which machines can run it
The requirements.txt reads: a reference to requirements_base.txt, then onnxruntime-gpu==1.18.0 and transformers==4.38.0. A separate requirements_macOS.txt exists beside requirements_base.txt.
The GPU build of onnxruntime in the main requirements file explains why the Linux and Windows route asks about CUDA. The macOS file is the answer for Apple Silicon, which gained support on 2024/07/17 in a change adapted from a community pull request.
The pins also carry a history. On 2024/10/18 the project updated the transformers and gradio library versions to avoid security vulnerabilities. If you pull these pins into a larger environment, expect conflicts with anything that wants a newer transformers, since the file fixes it at one exact version.
Windows users get a zip from Hugging Face, not a release from GitHub
The repository has no GitHub releases. The Windows package lives elsewhere: a one-click installer zip hosted on Hugging Face under the cleardusk account, with a file named LivePortrait-Windows-v20240829.zip linked from the README. Per the update notes, you unzip it and double-click run_windows.bat. The 2024/08/29 version added auto-update support, and earlier versions of the zip covered Humans mode and Animals mode.
This arrangement means the thing most Windows users will actually run is not versioned in this repository. You are trusting a separate Hugging Face repository, and its changelog is kept there too. If you need an audit trail, build from source with conda instead and compare against the pinned requirements.
Gradio apps and scripts cover different editing jobs
Besides the command-line scripts, the project has Gradio interfaces in app.py and app_animals.py, and a speed.py file at the top level. The update log lists what the interface and modes gained over time: portrait video editing, called v2v, on 2024/07/19; audio and video concatenating, driving video auto-cropping and template making to protect privacy on 2024/07/10; pose editing for source portraits on 2024/07/24; precise portrait editing on 2024/08/06; and image driven mode with regional control on 2024/08/19.
The detection threshold was also lowered by default on 2024/07/24 to increase recall. The trade is the usual one: a lower threshold finds more faces and may also find things that are not faces.
A hosted Hugging Face Space exists for trying it without installing anything, released on 2024/07/09 with help from the Hugging Face team and Gradio.
The license field is NOASSERTION and there are no releases to pin
GitHub reports the license as NOASSERTION, meaning it did not detect a recognised license, even though the repository has a LICENSE file. The practical rule is simple: open that file and read it yourself before using the code or the pretrained weights in pretrained_weights/ in anything beyond personal experiments.
There are no GitHub releases either, so the only fixed points are commits and the dated entries in the README's update list. The most recent dated entry there is 2025/06/01, in which the authors say LivePortrait has been adopted by video platforms such as Kuaishou, Douyin, Jianying and WeChat Channels. That is the authors' own statement about adoption. The last push to the repository was on 2026-06-01, and the repository is not archived.
Where the paper, the demo and the Chinese README live
The code is described as the official PyTorch implementation of the paper titled LivePortrait: Efficient Portrait Animation with Stitching and Retargeting Control, posted on arXiv as 2407.03168. The authors are listed from Kuaishou Technology, the University of Science and Technology of China and Fudan University, with Jianzhu Guo marked as project lead.
The project homepage is liveportrait.github.io, and the README offers a Simplified Chinese version in readme_zh_cn.md. Changelog details for individual releases of features sit in assets/docs/changelog/, one file per date.
If you want to compare it with other animation tools, this repository gives you no side-by-side data, so any comparison has to come from running both on your own images.
Editorial conclusion
Use LivePortrait if you want to run an official portrait-animation implementation locally and can set up conda, FFmpeg and, for Animals mode on Linux or Windows, X-Pose with a matching CUDA. Skip it if you need a tagged release to pin or a clearly detected license, because the repository has neither a GitHub release nor a license GitHub recognises. Before relying on it, read the LICENSE file in the repository and decide whether the Windows zip hosted on Hugging Face is a source you trust.
Frequently asked questions
Is LivePortrait open source?
The repository contains the official PyTorch implementation and a LICENSE file, but GitHub reports the license as NOASSERTION rather than a recognised one. Read the LICENSE file for the actual terms.
How do I install LivePortrait?
You need git, conda and FFmpeg. Clone https://github.com/KlingTeam/LivePortrait, create a conda environment with python=3.10, and on Linux or Windows install X-Pose if you want Animals mode. Windows users can instead unzip the one-click installer from Hugging Face and run run_windows.bat.
What is LivePortrait?
It is the official PyTorch implementation of the paper LivePortrait: Efficient Portrait Animation with Stitching and Retargeting Control, with a Humans mode and an Animals mode.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/klingairesearch-liveportrait)Star history chart for your README
[](https://hysenlabs.com/projects/klingairesearch-liveportrait)Was this useful?