Kandinsky S2V 1.0

Kandinsky S2V 1.0

Makes an avatar video from one reference frame and a speech track, generated in 5 second chunks. A Gradio demo of the Kandinsky S2V 1.0 pipeline for video and avatar developers.

Made by kandinskylab on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: Kandinsky S2V 1.0 is a Gradio demo of a speech-driven avatar pipeline. You give it one reference frame and a speech track, and it produces an avatar video. The video is generated in chunks of five seconds each.

The Space runs the Lite distilled version of the pipeline. It names a Qwen2.5-Omni model, a CLIP model and HunyuanVideo among its parts.

Who it suits: Developers and researchers who work on talking avatars and want to try the pipeline without installing it.

Industry | Who uses it | How they use it | Example
Media | Video creators | Animate a portrait from a voice track | Use an example image with its matching audio
Software | Avatar developers | Judge the pipeline before integrating it | Run a short clip and review the result
Research | Speech and video researchers | Study speech-driven video | Compare the 5 second and 10 second options

You can skip it if you need long videos on a free account, since a free account can run only the 5 second option.

Where to use it:

• A web browser, on the Space's page at Hugging Face, with nothing to install
• Windows, macOS or Linux with Python, from a copy you download, on a GPU with plenty of memory

Use case | Who | What they get back
Animate a portrait | Creators | An avatar video that follows the speech
Try an example | Newcomers | A result from a bundled image and audio pair
Test the pipeline | Developers | A generated clip of 5 or 10 seconds

How to set it up:

• Python 3.12 on your computer
• The Python packages listed in requirements.txt
• A GPU with a lot of memory, since the reference peak is 57 GiB on the Lite model
• A Hugging Face read token, only while the model repositories are private
• The huggingface_hub package, which gives you the hf command

1. Install the huggingface_hub package.
2. Download the Space with the hf command.
3. Move into the downloaded folder.
4. Install the pinned Gradio version.
5. Install the packages from requirements.txt.
6. Start the app with Python.

pip install -U huggingface_hub
hf download kandinskylab/Kandinsky-S2V-1.0-Demo --repo-type space --local-dir Kandinsky-S2V-1.0-Demo
cd Kandinsky-S2V-1.0-Demo
pip install gradio==6.27.0
pip install -r requirements.txt
python app.py

How to use it: Open the Space's page in your browser, or start your own copy with the command below. Add a reference frame and a speech track, choose 5 or 10 seconds of the track, and run it.

python app.py

Example: You pick an example image and its audio, and ask for a 5 second clip. The demo returns an avatar video of that image speaking along with the track.

For newcomers and experts: If you are new to avatar video, start with an example pair and the 5 second option, since it needs the least GPU time. Then try your own image and audio.

If you know the field, the Space runs on ZeroGPU with a 96 GB slot because the video decoder pushes memory use past the smaller 48 GB slot. Attention goes through PyTorch SDPA because FA3 is not available there, so generation is slower than on an H100. Two optional variables pick the config and the slot.

Title: GPU time requested per clip
5 second clip: 100 seconds
10 second clip: 180 seconds

You add a reference frame and a speech track
The viewer picks 5 or 10 seconds of audio
The pipeline generates the video in 5 second chunks
The avatar video plays on the page

Good to know:

• License: the Space lists mit
• Cost to run: GPU time at Hugging Face, where a request on the large slot counts double, so a free account with 300 seconds a day fits only the 5 second clip
• Data it reaches: the images and speech you add, plus the model weights, and a read token while the repositories are private
• Not for: long clips on a free account, or running on a 48 GB GPU, which the README says does not fit

On Hugging Face, 15 November 2026

• Likes: 5
• Built with: gradio
• License: mit

Links

• Space: https://huggingface.co/spaces/kandinskylab/Kandinsky-S2V-1.0-Demo
• Files: https://huggingface.co/spaces/kandinskylab/Kandinsky-S2V-1.0-Demo/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.