MEMFOF Optical Flow

MEMFOF Optical Flow

Estimates optical flow, the per-pixel motion between video frames, for Full HD video while using little GPU memory, with a Gradio demo and several trained models. It suits computer vision researchers and video engineers.

Made by egorchistov 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: MEMFOF is a method for estimating optical flow, the motion of every pixel from one frame to the next. It targets Full HD video and aims to keep accuracy high while using little VRAM. The Space is its demo, and the code comes from a paper shown at ICCV 2025.

Who it suits: It suits computer vision researchers and video engineers who need motion estimates for high-resolution footage without an oversized GPU.

Industry | Who uses it | How they use it | Example
Computer vision | Researchers | Estimate optical flow on video and compare models | A researcher runs two model versions on the same sequence.
Video engineering | Video processing teams | Get motion data for Full HD clips | An engineer tests a clip to see the flow results.
Robotics and driving | Perception engineers | Try models trained for everyday and driving scenes | An engineer tries the real-video model and the KITTI model.
Education | Students | Study a recent optical flow paper with working code | A student reads the code in the core folder and the demo notebook.

If you need ready-made video effects, skip it, because the project produces flow predictions and not edited video.

Where to use it:

• A web browser, through the Space's page on Hugging Face
• A local copy on Windows, macOS or Linux, using Python and a CUDA GPU
• The demo notebook, which the repository includes

Use case | Who | What they get back
Predict flow for a video | Researchers | Optical flow results for the sequence you give it
Compare model versions | Engineers | Results from versions tuned for real video, KITTI, Sintel or Spring
Train or evaluate | Researchers | The train.sh, eval.sh and submission.sh scripts in the repository

How to set it up:

• Python 3.10 or newer, PyTorch 2.5.0 or newer and CUDA 12.6 or newer, as the README states
• The packages listed in requirements.txt

1. Install the huggingface_hub package.
2. Download the Space into a local folder named optical-flow-MEMFOF.
3. Move into that folder.
4. Install gradio at version 5.34.2.
5. Install the packages listed in requirements.txt.
6. Start the demo.

pip install -U huggingface_hub
hf download egorchistov/optical-flow-MEMFOF --repo-type space --local-dir optical-flow-MEMFOF
cd optical-flow-MEMFOF
pip install gradio==5.34.2
pip install -r requirements.txt
python demo.py

How to use it: Open the Space's page, or start the demo locally with the command below. Give it a video sequence and it generates optical flow predictions. For ordinary camera footage, the README prefers the MEMFOF-Tartan-T-TSKH model.

python demo.py

Tool | What it does
demo.py | Hosts the Gradio demo page
demo.ipynb | A notebook that shows example usage
MEMFOF-Tartan-T-TSKH | The model the README recommends for real videos
train.sh, eval.sh and submission.sh | Scripts for training, evaluating and submitting results

Example: You open the demo and give it a short video sequence. It returns the predicted optical flow for the frames. You can switch to the model the README recommends for ordinary footage and compare the output.

For newcomers and experts: If you are new, optical flow shows where things move between two frames. Open the Space's page and try a short sequence with the default model.

If you know the field, the method trains at high resolution and uses several frames to keep memory low. The README lists six trained models, including versions tuned on KITTI, Sintel and Spring, and shows badges for those benchmarks.

Give it a video sequence
Choose a MEMFOF model
The model estimates motion between frames
Optical flow results appear

Good to know:

• The README prefers MEMFOF-Tartan-T-TSKH for footage from ordinary cameras
• Training needs several public datasets that you download yourself
• The README requires a CUDA GPU setup for local use
• License: bsd-3-clause
• Cost to run: none named
• Data it reaches: the video you give it, plus the model files it names

On Hugging Face, 13 November 2026

• Likes: 8
• Built with: gradio
• License: bsd-3-clause

Links

• Space: https://huggingface.co/spaces/egorchistov/optical-flow-MEMFOF
• Files: https://huggingface.co/spaces/egorchistov/optical-flow-MEMFOF/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.