Robust Video Matting (RVM)
Open-source recurrent neural network model for real-time, high-resolution human video matting without trimaps or background images.
At a Glance
Robust Video Matting code and pretrained models released under the GNU General Public License v3.0; free to use, modify and distribute under GPL terms.
Engagement
Available On
Alternatives
Listed Nov 2026
About Robust Video Matting (RVM)
Robust Video Matting (RVM) is the official implementation of the paper "Robust High-Resolution Video Matting with Temporal Guidance", developed at ByteDance Inc. with the University of Washington. It removes the background from people in video, producing alpha mattes and foreground layers without a trimap or pre-captured background. The model and code are published on GitHub under GPL-3.0.
What It Is
RVM is a human video matting model. Unlike approaches that treat each frame as an independent image, it uses a recurrent architecture that carries temporal memory across frames, which the authors say improves temporal coherence and matting quality. A training strategy that combines matting and segmentation objectives is described as improving robustness. The authors report 4K at 76 FPS and HD at 104 FPS on an Nvidia GTX 1080 Ti.
Workflow and Outputs
Users load a MobileNetV3 or ResNet50 variant (the README recommends MobileNetV3 for most cases) and run it through a provided convert_video API or custom inference code. The input can be a video file or an image sequence directory. Outputs can be a composited video or PNG sequence, plus optional raw alpha and foreground predictions. A downsample_ratio hyperparameter can be tuned or left to auto, and the recurrent states are cycled between frames. Training and evaluation documentation is included for building custom models.
Model Formats and Demos
Pretrained weights are offered for PyTorch, TorchHub, TorchScript, ONNX, TensorFlow, TensorFlow.js and CoreML. The ONNX models were tested on ONNX Runtime with CPU and CUDA backends, and the CoreML models require iOS 13+ and do not support dynamic resolution. A browser webcam demo visualizes the recurrent states, and a Colab notebook lets users try their own videos. Third-party ports listed include NCNN for Android, MNN, TNN and a Unity demo.
Community Discussions
Be the first to start a conversation about Robust Video Matting (RVM)
Share your experience with Robust Video Matting (RVM), ask questions, or help others learn from your insights.
Pricing
Open Source (GPL-3.0)
Robust Video Matting code and pretrained models released under the GNU General Public License v3.0; free to use, modify and distribute under GPL terms.
- Licensed under GPL-3.0 (OSI-approved)
- Source code and pretrained models published
- Available in PyTorch, TorchScript, ONNX, TensorFlow, and TensorFlow.js
- Free Colab demo and browser webcam demo
- Self-hosted; your own GPU/compute costs apply
Capabilities
Key Features
- Real-time human video matting without trimap or background image
- Recurrent architecture with temporal memory
- 4K 76FPS and HD 104FPS on GTX 1080 Ti
- MobileNetV3 and ResNet50 model variants
- Outputs composite video, PNG sequence, alpha and foreground
- Exports for PyTorch, TorchScript, ONNX, TensorFlow, TensorFlow.js, CoreML
- TorchHub loading
- Browser webcam demo and Colab demo
- Training and evaluation documentation
Integrations
Demo Video

