RepoMIT (models non-commercial)Video models7 min read

LivePortrait: make any portrait wink, smile and talk

Repo · picked by ReelMe Academy

Open-source portrait animation from Kuaishou: a still picture copies the face in a driving video, every blink, smile and head turn. Works on paintings, 3D characters, cats and dogs.

A fluffy white cat with bright cyan eyes sits on a wet, lamp-lit evening street: an example image from the LivePortrait repo
Get it

Run one command.

Paste it in your terminal. That's it.

pip install -r requirements.txt
The full write-up

LivePortrait: make any portrait wink, smile and talk, in detail

Give LivePortrait one still portrait and a short video of a real face, and the portrait moves like that face: it blinks, smiles, talks and turns its head. It comes from Kuaishou Technology, and its README says it has become a portrait-animation solution for humans, cats and dogs adopted by major video platforms: Kuaishou, Douyin, Jianying and WeChat Channels. The code is free and open; the paper is on arXiv.

The showcase at the top of the LivePortrait README: three classic paintings that glance, smile and talk.

How it works

You give it two things. The source is the portrait you want to animate: an image, or a video. No portrait yet? Make one in ReelMe Image. The driving video is a clip of a real face. LivePortrait reads the motion of the driving face and applies it to the source, so the result keeps the source's look and takes on the driver's expressions and head moves.

The LivePortrait Gradio app with source upload, driving video upload, cropping options and the Animate button

The LivePortrait app (python app.py): source image or video on the left, driving video on the right, then Animate.

WhatFrom the README and docs
InputA source image or video, plus a driving video, a driving image, or a saved .pkl motion template
FacesHumans mode, and an Animals mode trained on about 230K frames of animals, mostly cats and dogs
SpeedRTX 4090 with torch.compile: 7.59 ms per frame for the generator, the largest module. In the app, --flag_do_torch_compile makes later runs 20-30% faster, but it is not supported on Windows and macOS
Runs onLinux and Windows with an NVIDIA GPU; macOS with Apple Silicon for Humans mode, maybe 20x slower than an RTX 4090
No installAn online demo on HuggingFace and a Windows one-click installer

One face moves any portrait

These clips come from the LivePortrait project page. In each one, a single driving video moves every picture in the frame at once: oil paintings, statues, photos, 3D characters and pets.

Paintings start to talk

A seated lady, an angel, a woman among yellow roses and a marble Madonna follow one face.Full-length paintings: a dancer by the sea, a girl with a flower basket, and three young girls with a tambourine laugh and squint while the brushwork stays still.

Art styles and 3D characters

Paintings and 3D cartoon characters copy the same driving face.Marble angels, an illustrated couple, a painted child and an anime girl laugh with the driver.

Cats, dogs and a panda

Animals mode: cats, a shiba, an alpaca and a panda all follow one driving face (cropped out here).A white cat.
A corgi.A British shorthair.

Video in, video out

Portrait video editing (v2v): the dance stays, the face takes the driving expressions. The driver's eyes are blacked out in the original.A baby video, re-acted: new blinks and mouth moves come from the driving clip.

More results, including side-by-side comparisons with other methods, are on the project page.

Animals mode for cats and dogs

A separate Animals model is fine-tuned on about 230K frames of animals, mostly cats and dogs. The v1.1 update in January 2025 added more data: dogs got the bigger gain, and the model is now better at recognizing their mouths instead of mistaking them for noses. Animals mode uses X-Pose to find keypoints, and it is only tested on Linux and Windows with an NVIDIA GPU.

The README's Animals demo: the cat photo, and the same cat winking.

The README's command for it (the wink comes from a saved motion template):
python inference_animals.py -s assets/examples/source/s39.jpg -d assets/examples/driving/wink.pkl --driving_multiplier 1.75 --no_flag_stitching

The animation options and results of the LivePortrait Animals mode app: a ragdoll cat winking

The Animals mode app (python app_animals.py): stitching off, driving multiplier 1.75, and the cat winking in the results.

Fix the eyes and lips: retargeting

Retargeting is in the paper's title: *Efficient Portrait Animation with Stitching and Retargeting Control*. In the app's Retargeting Video panel you set a target lip-open ratio for a source video, and 0 closes the mouth. A keeping the lip silent option gives the lips a neutral expression; the authors note it may cause jitter between frames.

The Retargeting Video panel of the LivePortrait app with lip-open ratio settings and three result videos

Retargeting Video in the app: the source video, the retargeted result and the pasted-back result side by side.

Edit a still face with sliders

You do not always need a driving video. In the app, the relative pitch, relative yaw and relative roll sliders turn the head of the source portrait. Since August 2024 the app also has precise portrait editing for expressions: with each adjustment of the slider, the edited image updates in real-time.

The LivePortrait Retargeting panel: eyes-open ratio 0 and yaw -25 make a cartoon alpaca close its eyes and turn its head

The Retargeting panel: target eyes-open ratio 0 and relative yaw -25 close the alpaca's eyes and turn its head.

Drive with a photo, or only the lips

The driving signal can also be a single image. If both source and driving inputs are images, the output is an image too. Regional control picks what gets driven: exp, pose, lip, eyes or all.

Image-driven portrait animation in the LivePortrait app: a shouting woman's photo drives a painting of a girl with a book

Image-driven animation in the app: the regional options (exp, pose, lip, eyes, all), then one photo of a shouting face driving a painted girl.

Drive only the lip region, from the release notes:
python inference.py -s assets/examples/source/s5.jpg -d assets/examples/driving/d0.mp4 --animation_region lip

Film a driving video that works

The result is only as good as the face that drives it. The README's advice for your own driving video, plus a note on sound from the July 2024 release notes:

  • Crop it to 1:1: for example 512x512 or 256x256, or let LivePortrait do it with --flag_crop_driving_video.
  • Frame the head: like the example videos, and keep shoulder movement small.
  • Start neutral: the first frame should be a frontal face with a neutral expression.
  • Keep the sound: if the driving video has audio, it is added to the result, at the driving video's FPS.

d0.mp4, the README's default driving video: frontal, neutral at the start, head and shoulders only.

Want to share a motion without sharing a face? Each driving video is saved as a .pkl motion template that only contains the motions. The authors write that it is theoretically impossible to reconstruct the original face from it, and you can pass it with -d instead of the video.

Quick start

First get the code and a Python 3.10 environment: git clone https://github.com/KlingTeam/LivePortrait, then cd LivePortrait, conda create -n LivePortrait python=3.10 and conda activate LivePortrait. Then run the command under Get it, above. On macOS with Apple Silicon use pip install -r requirements_macOS.txt instead, and run with PYTORCH_ENABLE_MPS_FALLBACK=1 python inference.py. On Windows you can skip all of this with the one-click installer: unzip it and double-click run_windows.bat. Just want to try it? Use the online demo.

Download the weights, then animate the built-in example or open the app:
huggingface-cli download KlingTeam/LivePortrait --local-dir pretrained_weights --exclude "*.git*" "README.md" "docs"
python inference.py
python app.py # humans mode

Pick your own source with -s and driving video with -d:
python inference.py -s assets/examples/source/s9.jpg -d assets/examples/driving/d0.mp4

Requirements and license

  • git, conda and FFmpeg: installed before you start.
  • An NVIDIA GPU: on Linux or Windows. Animals mode is only tested there, and needs one extra op built first.
  • Or a Mac with Apple Silicon: Humans mode works, Animals mode does not, and it may be 20x slower than an RTX 4090.
  • The model weights: from HuggingFace (or Google Drive / Baidu Yun), placed in pretrained_weights.

The LivePortrait code is under the MIT License. Two parts it relies on are not free for business: the InsightFace models are for non-commercial research purposes only, and X-Pose, used by Animals mode, is only for non-commercial scientific research. For commercial use, the license file says to remove and replace InsightFace's detection models, and the Animals notes say the same for X-Pose. The authors also warn that portrait animation can be misused to create deepfakes, and they take no legal responsibility for what you generate.

No GPU? Animate your character in ReelMe

LivePortrait runs on your own machine and works on faces. For whole-body moves in the browser, open Motion Control in ReelMe Video (reelme.ai/video?tab=motion): add a motion video and a character image, and it copies the person's moves onto your character. To bring one picture to life from a text prompt instead, use Image to Video.

Credits

Media: LivePortrait by Jianzhu Guo, Dingyun Zhang, Xiaoqiang Liu, Zhizhou Zhong, Yuan Zhang, Pengfei Wan and Di Zhang (Kuaishou Technology), MIT-licensed code. The showcase GIF, demo GIF, example images and app screenshots come from the LivePortrait repo; the result clips come from the project page. Clips are re-encoded, shortened, silent and some are cropped to the result row; GIFs are converted to video.

Nov 13, 2026

How it works

3 steps

  1. Install LivePortrait

    Clone the repo, pip install the requirements, download the weights.

  2. Pick two inputs

    A portrait to animate, and a frontal, neutral-start video of a face.

  3. Animate and tune

    Press Animate and tune. Whole body? Use ReelMe Motion Control: reelme.ai/video?tab=motion

  • Video models · Repo · picked by ReelMe Academy

    Tencent Hunyuan Community License

    Tencent's open video model with over 13 billion parameters: write a prompt, get a cinematic 720p clip. It runs on a big NVIDIA GPU (45 to 80 GB) under Tencent's own community license.

    1. Clone and install: Linux, conda, PyTorch for CUDA 12.6 or 11.8, then the requirements.
    2. Download the weights: Follow ckpts/README.md. You need a GPU with 45 GB or more.
    3. Prompt and render: Run sample_video.py. No 80 GB GPU? Try text-to-video in ReelMe: reelme.ai/video?mode=t2v
  • Video models · Repo · picked by ReelMe Academy

    Apache-2.0

    Lightricks' open video model (v0.9.x): animate a still or a prompt, extend clips forward or backward, and make a new video follow another clip's depth, pose or edges.

    1. Start from a still: Pick one image to start from. No image? Describe it in ReelMe: reelme.ai/video?mode=t2v
    2. Write a shot list: One paragraph under 200 words: action, look, setting, camera, light.
    3. Add a guide clip: Optional: a depth, pose or edge video makes the result move the same way.
  • Video models · Repo · picked by ReelMe Academy

    Apache-2.0

    Open-source image-to-video that runs on your own NVIDIA GPU: one picture and a short prompt become a minute-long video, and you watch the frames arrive as it renders.

    1. Install FramePack: Windows: one-click package, then update.bat and run.bat. Linux: pip install.
    2. Add image and prompt: Upload one picture and say who moves and how, e.g. dances gracefully.
    3. Watch it grow: Watch frames arrive. No GPU? Try image-to-video in ReelMe: reelme.ai/video?mode=i2v

Want it step by step? From Zero to AI Creator