# SoulX-FlashTalk — card for AI agents

14B open-weights model for real-time, audio-driven talking-avatar video (image to video).

Category: AI presenters and avatars · maker Soul-AILab · checked 2026-10-27

## Price
- Price: Free
- Free tier: Open weights, free to download; self-hosted.

## License
Apache-2.0 — https://huggingface.co/Soul-AILab/SoulX-FlashTalk-14B/blob/main/README.md

- commercial: **yes** (Commercial use OK)
  > license: apache-2.0
  (https://huggingface.co/Soul-AILab/SoulX-FlashTalk-14B/blob/main/README.md)
- in app: **yes** (Can ship inside your app)
  > license: apache-2.0
  (https://huggingface.co/Soul-AILab/SoulX-FlashTalk-14B/blob/main/README.md)

## Rules to follow
- Keep the Apache-2.0 license and notices when you redistribute the weights.
- Do not clone a real person without their consent; the model card gives no output license, so check the repo.

## How to connect
- Download only
- Repository: https://github.com/Soul-AILab/SoulX-FlashTalk
- Website: https://huggingface.co/Soul-AILab/SoulX-FlashTalk-14B

## Notes
Model card says the inference code and weights were released on 2026.01.08.

## Numbers (AgentGid data)
- Gid Score: 13
- GitHub stars: 1,569

## Lines to paste into your prompt or AGENTS.md

```
- SoulX-FlashTalk (AI presenters and avatars). 14B open-weights model for real-time, audio-driven talking-avatar video (image to video). Price: free; commercial use OK.
  Connect: download from https://huggingface.co/Soul-AILab/SoulX-FlashTalk-14B (no API).
  Rule: Keep the Apache-2.0 license and notices when you redistribute the weights.
  Rule: Do not clone a real person without their consent; the model card gives no output license, so check the repo.
  License: https://huggingface.co/Soul-AILab/SoulX-FlashTalk-14B/blob/main/README.md (checked 2026-10-27; details: https://agentgid.com/assets/soulx-flashtalk/)
```

## Links
- AgentGid page: https://agentgid.com/assets/soulx-flashtalk/
- Is it free for commercial use? https://agentgid.com/assets/soulx-flashtalk/commercial-use/
- Source: https://huggingface.co/Soul-AILab/SoulX-FlashTalk-14B
- Source: https://github.com/Soul-AILab/SoulX-FlashTalk


## SoulX-FlashTalk: a talking video from one photo and audio

SoulX-FlashTalk is a free, open AI model that turns one portrait picture and a voice recording into a video of that face talking in sync with the audio. It is built for live, endless streams, but it is very large and needs data-center graphics cards. The model and code are under the Apache 2.0 license, which allows paid use.

### Where to use it
- A live talking avatar: Put a speaking face on a voice assistant or chat, so the picture moves its lips while the reply plays, with little delay.
- Live streams and video podcasts: Run a virtual host for a stream or a video podcast that talks for as long as the audio goes on.
- Long videos without cuts: Make a long talking-head video, such as a lecture or a narrated story, where the face stays the same from start to end.
- Cartoon and animal characters: Make a drawn character, a painted portrait or an animal picture sing or speak, for a fun clip or a mascot.
- Presenter for course videos: Pair a recorded or generated voice-over with a presenter photo you have rights to, for lessons or product explainers. (idea, not from the official pages)
- Research on digital humans: Study or compare talking-avatar methods with an open model and its technical report, instead of a closed paid service. (idea, not from the official pages)

### How to start
- Make a Python 3.10 environment and install PyTorch 2.7.1 for CUDA 12.8, then the other requirements, flash-attention and FFmpeg.

```
conda create -n flashtalk python=3.10
conda activate flashtalk
pip install torch==2.7.1 torchvision==0.22.1 --index-url https://download.pytorch.org/whl/cu128
pip install -r requirements.txt
```

- Download the 14B model and the audio encoder it needs from Hugging Face.

```
pip install "huggingface_hub[cli]"
huggingface-cli download Soul-AILab/SoulX-FlashTalk-14B --local-dir ./models/SoulX-FlashTalk-14B
huggingface-cli download TencentGameMate/chinese-wav2vec2-base --local-dir ./models/chinese-wav2vec2-base
```

- Run the single-GPU script; it needs more than 64 GB of graphics memory, or about 40 GB with the --cpu_offload switch.

```
bash inference_script_single_gpu.sh
```


### License
Apache 2.0 for the code and the model weights, so paid use is allowed. Keep the license and notices when you share the files, and say which files you changed.

### Good to know
- It needs very large hardware: more than 64 GB of graphics memory on one card, or about 40 GB with a memory-saving switch.
- Real-time speed needs eight H800 data-center cards or better. On a single card you get a finished video file, not a live stream.
- The same team released SoulX-FlashHead, a separate talking-head project that runs in real time on home graphics cards like the RTX 4090.
- There is no online demo yet, so you cannot try it in a browser before setting it up.
- The pages do not say how the videos may be used and add no rules on faces or voices. Animate only people who agreed to it.

Checked 2026-11-19 against: https://huggingface.co/Soul-AILab/SoulX-FlashTalk-14B, https://huggingface.co/Soul-AILab/SoulX-FlashTalk-14B/raw/main/README.md, https://huggingface.co/Soul-AILab/SoulX-FlashTalk-14B/raw/main/LICENSE.txt, https://huggingface.co/api/models/Soul-AILab/SoulX-FlashTalk-14B, https://github.com/Soul-AILab/SoulX-FlashTalk, https://github.com/Soul-AILab/SoulX-FlashTalk/blob/main/README.md, https://github.com/Soul-AILab/SoulX-FlashTalk/blob/main/LICENSE, https://github.com/Soul-AILab/SoulX-FlashTalk/blob/main/inference_script_single_gpu.sh, https://github.com/Soul-AILab/SoulX-FlashTalk/blob/main/generate_video.py, https://github.com/Soul-AILab/SoulX-FlashTalk/blob/main/requirements.txt, https://soul-ailab.github.io/soulx-flashtalk/, https://huggingface.co/api/models/TencentGameMate/chinese-wav2vec2-base

### Mistakes to avoid
- Don't: Trying to run it on a Mac or a 24 GB gaming card → Do: Use a Linux NVIDIA GPU with 40 GB or more and --cpu_offload, or SoulX-FlashHead for consumer cards (The model needs more than 64 GB of VRAM on one card, or 40 GB with offload, and the install pins CUDA 12.8 wheels.)
- Don't: Downloading only the SoulX-FlashTalk-14B weights → Do: Also download TencentGameMate/chinese-wav2vec2-base into models/ and pass it as --wav2vec_dir (The audio is turned into features by this separate wav2vec model before the video is made.)
- Don't: Running bash inference_script_single_gpu.sh and expecting your own photo and voice in the video → Do: Pass --cond_image and --audio_path (and --save_file) to generate_video.py, or edit the script (The script hard-codes the example files examples/man.png and examples/cantonese_16k.wav.)
- Don't: Expecting live, real-time output from a single GPU → Do: Plan on offline rendering to an mp4 on one GPU; real time needs 8xH800 or better (The README limits real-time speed to large multi-GPU setups.)
- Don't: Skipping FFmpeg because Python packages installed fine → Do: Install FFmpeg (apt, yum or conda-forge ffmpeg 7) before running (The README lists FFmpeg as a separate install step; the video with sound is written through it.)

### Questions
- Can I run it on my laptop or a gaming PC? No. One card needs more than 64 GB of graphics memory, or about 40 GB with --cpu_offload; the team points to SoulX-FlashHead for consumer cards.
- What do I give it? A portrait image, a speech audio file and a short text prompt; the example uses a PNG and a 16 kHz WAV.
- Where does the video go? To the file set with --save_file, or into a sample_results folder by default.
- Can I use it commercially? The model and code are under Apache 2.0, which allows commercial use; keep the license and notices with any copies.
- What is it built on? On the InfiniteTalk and Wan models, with code from Self Forcing.

---
AgentGid's catalog is curated: every entry was checked against the maker's official pages (registry, repository, license and pricing pages) on the date shown; a license value shows the quote it stands on wherever AgentGid has one. Usage numbers come from AgentGid's own scan of public repositories set up for AI coding agents. Claims by makers are reported, not tested by AgentGid.
