RepoPicks Radar

Updated every day around 23:30 UTC

Guides

VideoChat

Henry-23/VideoChatReviewed by AI

Checked by RepoPicks Radar on

What is VideoChat?

A demo you run yourself that turns a recorded video of a person into a talking avatar you can speak with: it listens to your microphone, answers with a language model and replies out loud while the face lip-syncs, with the first audio in about 3 seconds. You can add your own avatar video and clone a voice from a 3 to 10 second sample. It is assembled from FunASR, Qwen, GLM-4-Voice, GPT-SoVITS, CosyVoice and MuseTalk, needs Ubuntu, Python 3.10, CUDA and an NVIDIA card with 8 to 20 GB of memory, and the main instructions are in Chinese.

How easy is it to install VideoChat?

D · terminal, several tools or a specific graphics card · Linux

Linux is the only documented platform and it is a D: the README asks for "ubuntu 22.04, python 3.10, CUDA 12.2", a conda environment, `pip install -r requirements.txt` with around a hundred pinned packages, a `git lfs` clone from ModelScope to get the model weights, and an API key pasted into line 14 of `app.py` before `python app.py` opens the local page. The graphics card requirement is roughly 8 GB for the cascade mode and 20 GB for the end-to-end mode, measured on an A100.

Coming soon: install in 1 click with the RepoPicks app.

VideoChat on GitHub

Stars
1,331on Nov 10, 2026
Downloads
—all releases
Growth
—no gain in the last reading

Stars over time

1,3311,3321,333 Nov 6Nov 7Nov 8Nov 9Nov 10 1,3311,3321,333 Nov 6Nov 8Nov 10

Facts

Runs on
Linux
You need to install
Python, NVIDIA graphics card, FFmpeg
To build from source
Python, NVIDIA graphics card, FFmpeg
Disk space
1.8 GB–6.7 GB · estimate
Accounts or keys
Alibaba Cloud DashScope (Qwen and CosyVoice)
Cost
Free, API costs extra · The code is MIT and free. By default the language model and the voice come from Alibaba Cloud's paid Qwen and CosyVoice APIs; to run everything locally instead you have to edit the source files and provide a much bigger graphics card.
License
MIT (Permissive)
Can replace
HeyGen, Synthesia, D-ID
Last change on GitHub
Dec 18, 2025

VideoChat risk signals (8)

  • AttentionDependencies with known vulnerabilities

    Pinned dependencies with known vulnerabilities (OSV): gradio: GHSA-39mp-8hj3-5c49, GHSA-5cpq-9538-jm2j, GHSA-6655-8ph2-63j3; torch: GHSA-3749-ghw9-m3mg, GHSA-53q9-r3pm-6pq6, GHSA-63cw-57p8-fm3p; diffusers: GHSA-7wx4-6vff-v64p, GHSA-98h9-4798-4q5v, PYSEC-2026-2446; accelerate: GHSA-4j2p-28q2-5m79, PYSEC-2026-3804; gdown: GHSA-76hw-p97h-883f, PYSEC-2026-2158; requests: GHSA-9hjg-9r4m-mvj7, GHSA-gc5v-m9x4-r6x2, PYSEC-2026-1872; pytorch-lightning: GHSA-75m9-98v2-hjpm, PYSEC-2026-3043, PYSEC-2026-3967; sentencepiece: GHSA-38vq-g6vr-w8wf, PYSEC-2026-1909; opencc: GHSA-7fqq-q52p-2jjg; modelscope: GHSA-fhhq-h4hg-549x, PYSEC-2026-2663; transformers: GHSA-29pf-2h5f-8g72, GHSA-37mw-44qp-f5jm, GHSA-4w7r-h757-3r74; deepspeed: GHSA-8cp5-3rf8-8gfh, PYSEC-2024-109; hydra-core: GHSA-2cp2-2r3c-7p7r, GHSA-c3wx-c55w-pxjq, PYSEC-2026-3850; lightning: GHSA-cgwc-qvrx-rf7f, GHSA-mr7h-w2qc-ffc2, GHSA-qqmf-gpg7-g8gw; protobuf: GHSA-7gcm-g887-7qv7, GHSA-8qvm-5x2c-j2w7, PYSEC-2026-1805; requests: GHSA-9hjg-9r4m-mvj7, GHSA-gc5v-m9x4-r6x2, PYSEC-2026-1872

    automatic check

  • Attentionvoice and face cloning can impersonate

    The project description promises "可自定义形象与音色,支持音色克隆" (custom appearance and voice, with voice cloning). Section 5.1 takes "录制好的数字人形象视频" — a recorded video of a person — as the avatar, and 5.2 clones a voice from a reference clip of "3-10s". Nothing in the repository asks whether the person in the video or the recording agreed, and there is no disclaimer.

    AI review

  • Attentiondevelopment slowed

    The last commit in the checked snapshot is 2025-12-18, 322 days before the check, with no releases and 25 open issues. The README itself warns "requirements中的版本仅供参考,最新依赖请查看对应模块的Repo说明" — the pinned versions are only a reference, check each component's own repository.

    AI review

  • Attentionconfiguration means editing source files

    The API key goes into `app.py` line 14 (`os.environ["DASHSCOPE_API_KEY"] = "INPUT YOUR API-KEY HERE"`), downloading GLM-4-Voice means pasting `snapshot_download(...)` calls into `app.py`, a new avatar means editing `avatar_list` in `src/thg.py`, and going key-free means deleting the `CosyVoice_API` parts of `src/tts.py`. There is no settings file.

    AI review

  • Infoweights and code from modelscope

    The recommended install is `git clone https://www.modelscope.cn/studios/AI-ModelScope/video_chat.git` — the code and the tracked weight files come from ModelScope, a Chinese model platform, not from this GitHub repository. Manual downloads point at MuseTalk, GPT-SoVITS and the ZhipuAI `glm-4-voice-*` models.

    AI review

  • Infoparts carry their own licences

    This repository is MIT, but it only glues together other projects: FunASR, Qwen, GLM-4-Voice, GPT-SoVITS, CosyVoice, edge-tts and MuseTalk. Each of those models and their weights have their own terms, which decide what you may do with the avatars and audio you produce.

    AI review

  • Infofree voice route uses a browser service

    Section 3.2.2 suggests, when you have no API key, "使用`Edge_TTS`调用Edge浏览器的免费TTS服务" — calling the free text-to-speech service built into the Edge browser through the `edge-tts` package, a service Microsoft publishes for its browser rather than for other programs.

    AI review

  • Infoknown rough edges

    The README's "已知问题" (known issues) section lists missing resource errors, model loading failures from incomplete weight downloads, and "右侧视频流播放卡顿:需等待Gradio优化Video Streaming效果" — the video stream stutters while waiting for Gradio to improve video streaming.

    AI review

Evidence is shown in its original language.

What Radar could not check about VideoChat

  • Whether the install still works, since the pinned versions are from 2024 and the README says to check each component itself
  • How the conversation really feels on a smaller card than the A100 used for the 3 and 7 second figures
  • What Alibaba Cloud charges for the Qwen and CosyVoice calls in the default setup
  • Real disk space taken by all the model weights
  • Whether it can be made to run on Windows or a Mac, which the instructions never cover
  • How good the lip-sync and the cloned voice are with your own video and audio

Similar projects

The RepoPicks app

Coming soon: install in 1 click with the RepoPicks app.

Before you download

See the release page on GitHub