OBS AI Matting vs Robust Video Matting

OBS AI Matting is a free, open-source Linux filter for OBS Studio that removes, blurs, or replaces webcam backgrounds in real time. It runs Robust Video Matting through ONNX Runtime and CUDA to create soft edges and stable cutouts. Transparent mode lets OBS image, video, or color sources appear behind the camera, while a built-in mode blurs the background. Auto light match can adjust exposure and white balance to follow a selected OBS background or scene. Controls include blur strength, brightness, gamma, matte hardness, and quality settings of 384, 512, or 720. The interface supports English and Spanish and follows OBS language settings. Requirements include OBS Studio 28 or later, ONNX Runtime with CUDA, an NVIDIA GPU with CUDA and cuDNN, CMake, and a C++17 compiler. A roughly 107 MB RVM model must be downloaded separately. CPU fallback exists if CUDA fails, though a GPU is recommended for real-time use.

Robust Video Matting is a free project for separating people from video backgrounds. It processes frames with a recurrent neural network that carries temporal memory, and the project says it can perform matting in real time without additional inputs. A browser webcam demo runs the model live and can visualize its recurrent states. The conversion API accepts a video or image sequence and can produce a composited video or PNG sequence, optionally including alpha and foreground outputs.

Models are available for PyTorch, TorchHub, TorchScript, ONNX, TensorFlow, TensorFlow.js, and CoreML. MobileNetv3 is recommended for most uses; ResNet50 is larger and described as offering small performance improvements. The README reports GPU throughput figures, including 4K at 76 FPS and HD at 104 FPS on a GTX 1080 Ti, and 4K at 154 FPS and HD at 172 FPS on an RTX 3090 with FP16 under specified settings. The project cautions that file conversion runs much slower because its script does not use hardware video encoding or decoding or parallel tensor transfers. CoreML models require iOS 13 or later and do not support dynamic resolution. The code is GPL-3.0 licensed.