metadata

title: Video Model Studio
emoji: 🎥
colorFrom: gray
colorTo: gray
sdk: gradio
sdk_version: 5.20.1
app_file: app.py
pinned: true
license: apache-2.0
short_description: All-in-one tool for AI video training

🎥 Video Model Studio (VMS)

Presentation

What is this project?

VMS is a Gradio app that wraps around Finetrainers, to provide a simple UI to train AI video models on HF中国镜像站.

You can deploy it to a private space, and start long-running training jobs in the background.

Funding

VideoModelStudio is 100% open-source project, I develop and maintain it during both my pro and personal time. If you like it, you can tip! If not, have a good day 🫶

News

🔥 2025-03-12: VMS now officially supports Wan!
🔥 2025-03-11: I have added a tab to preview a model!
🔥 2025-03-10: Various small fixes and improvements
🔥 2025-03-09: I have added a basic CPU/RAM monitor (no GPU yet)
🔥 2025-03-02: Made some fixes to improve Finetrainer reliability when working with big datasets
🔥 2025-02-18: I am working to add better recovery in case of a failed run (this is still in beta)
🔥 2025-02-18: I have added persistence of UI settings. So if you reload Gradio, you won't lose your settings!

TODO

Add Aya-Vision-8B for frame analysis (currently we use Qwen2-VL-7B)

Features

Run Finetrainers in the background

The main feature of VMS is the ability to run a Finetrainers training session in the background.

You can start your job, close the web browser tab, and come back the next morning to see the result.

Automatic scene splitting

VMS uses PySceneDetect to split scenes.

Automatic clip captioning

VMS uses LLaVA-Video-7B-Qwen2 for captioning. You can customize the system prompt if you want to.

Download your dataset

Not interested in using VMS for training? That's perfectly fine!

You can use VMS for video splitting and captioning, and export the data for training on another platform eg. on Replicate or Fal.

Supported models

VMS uses Finetrainers under the hood. In theory any model supported by Finetrainers should work in VMS.

In practice, a PR (pull request) will be necessary to adapt the UI a bit to accomodate for each model specificities.

Wan

I am currently testing Wan LoRA training!

LTX-Video

I have tested training a LTX-Video LoRA model using videos (not images), on a single A100 instance. It requires about 18/19 Gb of VRAM, depending on your settings.

HunyuanVideo

I have tested training a HunyuanVideo LoRA model using videos (not images),, on a single A100 instance.

It requires about 47~49 Gb of VRAM, depending on your settings.

CogVideoX

Do you want support for this one? Let me know in the comments!

Limitations

No AV1 on A100

If your dataset contains videos encoded using the AV1 codec, you might not be able to decode them (eg. during scene splitting) if your machine doesn't support hardware decoding.

Nvidia A100 don't support hardware AV1 decoding for instance.

It might be possible to convert them on server-side or use software decoding directly from Python, but I haven't looked into that yet (you can submit a PR if you have an idea).

My recommendation is to make sure your data comes in h264.

You can use FFmpeg to do this, eg:

ffmpeg -i input_video_in_av1.mp4 -vcodec libx264 -acodec aac output_video_in_h264.mp4

One-user-per-space design

Currently CMS can only support one training job at a time, anybody with access to your Gradio app will be able to upload or delete everything etc.

This means you have to run VMS in a PRIVATE HF Space, or locally if you require full privacy.

Deployment

VMS is built on top of Finetrainers and Gradio, and designed to run as a HF中国镜像站 Space (but you can deploy it anywhere that has a NVIDIA GPU and supports Docker).

Full installation at HF中国镜像站

Easy peasy: create a Space (make sure to use the Gradio type/template), and push the repo. No Docker needed!

That said, please see the "RUN" section for info about environement variables.

Dev mode on HF中国镜像站

Enable dev mode in the space, then open VSCode in local or remote and run:

pip install -r requirements.txt

As this is not automatic, then click on "Restart" in the space dev mode UI widget.

Full installation somewhere else

I haven't tested it, but you can try to provided Dockerfile

Prerequisites

About Python:

I haven't tested Python 3.11 or 3.12, but I noticed some incompatibilities with Python 3.13 dependencies failing to install.

So I recommend you to install pyenv to switch between versions of Python.

If you are on macOS, you might already have some versions of Python installed, you can see them by typing:

% python3.10 --version
Python 3.10.16
% python3.11 --version
Python 3.11.11
% python3.12 --version
Python 3.12.9
% python3.13 --version
Python 3.13.2

Once pyenv is installed you can type:

pyenv install 3.10.16

Full installation in local

the full installation requires:

Linux
CUDA 12
Python 3.10

This is because of flash attention, which is defined in the requirements.txt using an URL to download a prebuilt wheel expecting this exact configuration (python bindings for a native library)

./setup.sh

Degraded installation in local

If you cannot meet the requirements, you can:

solution 1: fix requirements.txt to use another prebuilt wheel
solution 2: manually build/install flash attention
solution 3: don't use clip captioning

Here is how to do solution 3:

./setup_no_captions.sh

Run

Running the Gradio app

Note: please make sure you properly define the environment variables for STORAGE_PATH (eg. /data/) and HF_HOME (eg. /data/huggingface/)

python3.10 app.py

Running locally

See above remarks about the environment variable.

By default run.sh will store stuff in .data/ (located inside the current working directory):

./run.sh

Spaces:

jbilcke-hf
/

VideoModelStudio

Running