MultiTalk
MeiGen-AI/MultiTalkReviewed by AI
Checked by RepoPicks Radar on
What is MultiTalk?
MultiTalk is research code (NeurIPS 2025) that turns a reference image, one or more voice tracks and a written prompt into a 480p video of people holding a conversation, singing or interacting, with lips matching the audio. It can also read text aloud itself with a built-in voice model (Kokoro). Clips are short, around 15 seconds. You run it from the terminal or from a simple web page you start yourself, after setting up Python on Linux, PyTorch for NVIDIA CUDA, FFmpeg and four separate model downloads. An NVIDIA graphics card is needed.
MultiTalk is free and open source (Apache-2.0) and can replace HeyGen, Synthesia, and D-ID.
How easy is it to install MultiTalk?
D · terminal, several tools or a specific graphics card · Linux
Linux is the only documented platform: a conda environment, PyTorch and xformers built for CUDA 12.1, flash-attn compiled with ninja, FFmpeg, four model downloads from Hugging Face, and then moving and symlinking files inside the weights folder before the first command. Every example assumes an NVIDIA graphics card.
Coming soon: install in 1 click with the RepoPicks app.
MultiTalk on GitHub
- Stars
- 3,033on Nov 10, 2026
- Downloads
- —all releases
- Growth
- +4stars in 6 days
Stars over time
Facts
- Runs on
- Linux
- You need to install
- Python, FFmpeg, NVIDIA graphics card
- To build from source
- Python, FFmpeg
- Disk space
- 210 MB–570 MB · estimate
- Accounts or keys
- None
- Cost
- Free
- License
- Apache-2.0 (Permissive)
- Last change on GitHub
- May 22, 2026
MultiTalk risk signals (5)
- Attentionrealistic talking video of real people
README.md promises "Realistic Conversations" generated from "a multi-stream audio input, a reference image and a prompt" with "consistent lip motions aligned with the audio". The license section adds that your use "must not involve sharing any content that violates applicable laws, causes harm to individuals or groups, disseminates personal information intended for harm, spreads misinformation". It can make real people appear to say things they never said.
AI review
- Infolarge model downloads not in repo
The README asks you to download Wan2.1-I2V-14B-480P, chinese-wav2vec2-base, Kokoro-82M and MeiGen-MultiTalk with huggingface-cli into ./weights. These files are many gigabytes and are not part of the repository.
AI review
- Infoinstall steps use sudo symlinks
Model preparation tells you to run "sudo ln -s {Absolute path}/weights/MeiGen-MultiTalk/multitalk.safetensors weights/Wan2.1-I2V-14B-480P/" and to rename a file inside the downloaded base model, so the setup changes files with administrator rights.
AI review
- Infocloud api sdk in requirements
requirements.txt includes "dashscope", the client library for Alibaba Cloud's AI service, although the documented commands run locally.
AI review
- Infoauthors point to newer models
README.md news lists InfiniteTalk (Aug 2025) and LongCat-Video-Avatar-1.5 (May 21, 2026) as newer and "upgraded" frameworks from the same team; the last commit here is 2026-05-22, and 720p and ComfyUI are still unfinished items on the Todo list.
AI review
Evidence is shown in its original language.
What Radar could not check about MultiTalk
- Total disk space after downloading the Wan2.1 14B base model and the other three model sets
- How much graphics-card memory is really needed and how long a 15-second clip takes
- Whether the pinned install steps (CUDA 12.1 wheels, flash-attn 2.7.4) still work today
- Whether it can run on Windows or macOS, which are not documented
- The real quality and the colour drift the README mentions in longer videos