A Neural Green Screen Inside Your OBS Pipeline

What it does This plugin sits inside OBS Studio as a filter, using ONNX Runtime to segment portrait video and drop the background pixels so you can layer a new scene behind the subject. It also offers low-light enhancement for gloomy footage, keeping the entire pipeline inside the streaming software you already use.

The interesting bit Rather than betting on a single model, it bundles several pretrained segmentation networks—SINet, PP-HumanSeg, MediaPipe Meet, RobustVideoMatting, and a monocular depth estimator—so the plugin can adapt to different subjects and lighting. The authors explicitly target CPU inference as a first-class citizen, which is unusual for real-time video segmentation.

Key highlights

• Runs entirely within OBS as a source filter, composable with other plugins.
• Mac users get CoreML acceleration on Apple Silicon; Linux builds can target CUDA or MIGraphX if compiled from source.
• CPU thread count is exposed in the UI, and the authors recommend a 2-thread setting as the sweet spot.
• Includes secondary low-light enhancement models (TBEFN, URetinex-Net) for scenes that need more than background replacement.
• Pretrained weights are pulled from existing open-source segmentation projects, making this a curated integration rather than a from-scratch model.

Caveats

• Windows GPU acceleration is listed as planned (WinML), not currently shipped.
• Official source-build scripts only support Debian Forky; other platforms rely on release binaries or Flatpak.
• Mac binaries do not run under Rosetta 2, so Intel and Apple Silicon builds are not interchangeable.

Verdict Grab it if you want a virtual green screen without leaving OBS or buying extra gear. If you are not an OBS user, or you need a standalone segmentation pipeline to embed in your own app, this will not fit.