Ultralight Digital Human
anliyuan/Ultralight-Digital-HumanReviewed by AI
Checked by RepoPicks Radar on
What is Ultralight Digital Human?
Python training and inference scripts for a very small talking-head model. You film one person for three to five minutes with their whole face visible and clean sound, run the preprocessing, train a model for that person on an NVIDIA graphics card, and afterwards you can feed in any recording and get a video where that person's mouth follows the words. The author says it is light enough to run in real time on phones. There is no program to install, no ready model and no licence file, and the author now points to a newer project, FeatherTalk.
Ultralight Digital Human is free and can replace HeyGen, Synthesia, and D-ID.
How easy is it to install Ultralight Digital Human?
D · terminal, several tools or a specific graphics card · Windows
- Windows D
- Linux D
Both Windows and Linux are a D and need the same work: a conda environment with Python 3.10 and PyTorch with CUDA, a file downloaded by hand from Google Drive into data_utils/, then separate commands for preprocessing, training and inference plus an ffmpeg command to put the sound back. Macs are not covered, since the setup is built around an NVIDIA card.
Coming soon: install in 1 click with the RepoPicks app.
Ultralight Digital Human on GitHub
- Stars
- 2,670on Nov 10, 2026
- Downloads
- —all releases
- Growth
- +4stars in 6 days
Stars over time
Facts
- Runs on
- Windows, Linux
- You need to install
- Python, NVIDIA graphics card, FFmpeg
- To build from source
- FFmpeg
- Disk space
- not determined
- Accounts or keys
- None
- Cost
- Free · Nothing is charged, but no licence file was found, so the terms for using the code and anything you make with it are unclear. You pay in computer time for training.
- License
- No license (No license)
- Last change on GitHub
- Jul 22, 2026
Ultralight Digital Human risk signals (7)
- AttentionDownload links to file-sharing sites
README links to https://drive.google.com/file/d/1e4Z9zS053JEWl6Mj3W9Lbc9GDtzHIg6b/view?usp=drive_link
automatic check
- Attentionmakes a real person say anything
The whole purpose is to train on one real person's video and then drive that face with any audio file (`python inference.py --asr hubert --dataset ... --audio_feat your_test_audio_hu.npy`). The README even says the demo video is a broadcast by a Chinese television presenter taken without permission: "我会提供一个demo视频,来自康辉老师的口播,侵删" (I will provide a demo video, from Mr Kang Hui's broadcast, remove it if it infringes). Making a real person appear to say words they never said can break likeness rights and impersonation laws, and there is nothing in the project that asks for the person's consent.
AI review
- Attentionlicense badge without license file
The README shows a badge reading `license-Apache 2` linking to `./LICENSE`, but the repository's top-level files are only .gitignore, README.md, data_utils/, datasetsss.py, demo/, dihuman_run.py, face_utils.py, inference.py, pth2onnx.py, train.py and unet.py — there is no LICENSE file, so the badge promises terms the repository does not contain.
AI review
- InfoNo license detected
No license detected by GitHub, the license file, the LICENSES folder or the winget manifest
automatic check
- Infoauthor points to a successor project
The README now opens with "If you want a newer and more deployment-friendly version of this project, please check out FeatherTalk. It is a cleaned-up successor to Ultralight Digital Human". The last commit in the snapshot is from 2026-07-22, 105 days before the check, and 119 issues are open.
AI review
- Inforesearch code not finished
The README says the earlier code had several bugs that were only fixed recently, that the streaming code "代码还没有完善" (is not polished yet), that for real-time use "重构代码是必须的" (rewriting the code is necessary), and that a faster method the author has is kept closed because "有人在商用,暂时不做开源" (someone is using it commercially, so it is not open-sourced for now).
AI review
- Infostrict requirements for the input video
README: the video must be 3 to 5 minutes with the full face visible in every frame, exactly 20 fps for the wenet encoder or 25 fps for hubert, the test audio must be 16000 Hz, and the author warns that noise, echo from an empty room or unclear speech will probably ruin the result, recommending an external microphone.
AI review
Evidence is shown in its original language.
What Radar could not check about Ultralight Digital Human
- How good the resulting face and mouth look with your own video
- How long training takes and how much graphics memory it needs
- Whether the file on Google Drive is still available and what exactly it contains
- Whether it really runs in real time on a phone, since no mobile code is published here
- What you are allowed to do with the code and the videos, given the missing licence