ProVCam - AI Virtual Camera for Linux

AI virtual camera for Linux with real-time background replacement

Runs on AMD Radeon (ROCm) · drops straight into OBS / Zoom / Google Meet / Discord

ProVCam is an AI virtual camera for Linux that provides real-time background replacement, blur, green screen, and webcam output through v4l2loopback. It is designed for Wayland and X11 desktops, ships as AppImage builds, and works with OBS, Zoom, Google Meet, Discord, and other webcam-enabled apps.

---

Keywords

linux · virtual-camera · webcam · v4l2loopback · background-replacement · ai-background-removal · rocm · amd · appimage · obs · zoom · wayland · python · rust · video

---

Demo

provcam-demo-720p.mp4

---

Download

Two editions ship side-by-side. Pick one, make it executable, run it.

Edition	File	Size	Backend	When to pick
Rust (recommended)	ProVCam-Rust-x86_64.AppImage	~1.1 MB	Rust inference + Python fallback	Faster startup, lower RAM, single binary
Python	ProVCam-Python-x86_64.AppImage	~230 KB	Pure Python / PyTorch	Easier to modify and extend

# Rust edition (recommended)
chmod +x ProVCam-Rust-x86_64.AppImage
./ProVCam-Rust-x86_64.AppImage

First run copies the payload to ~/.local/share/provcam/ and opens the installer (creates the .venv, installs PyTorch ROCm, pre-caches the AI model). Later runs go straight to the GUI.

---

What is it

ProVCam takes your physical webcam (or a phone camera via Iriun), runs RobustVideoMatting to remove the background, composites the foreground onto a blurred / green / black / custom image, and pipes the result into a v4l2loopback device that any Linux app can treat as a regular webcam.

Built for AMD Radeon cards where ONNX Runtime's ROCm support is missing. Falls back to CPU when ROCm isn't available.

Features

• 🎭 AI background removal — RVM models, cached for offline use
• 🖼 Background modes: blur, custom image, green, black
• 🎚 Matte profiles: clean, balanced, soft
• ⚡ AMD GPU acceleration via PyTorch ROCm (FP16 when supported)
• 🛡 Supervisor with automatic fps/resolution downgrade on failures
• 🧩 Zero-copy IPC (mmap seqlock) between capture / inference / output processes
• 🖱 CustomTkinter GUI with live preview
• 📦 Self-contained AppImage — no manual install required

---

Requirements

• Linux with Wayland or X11

• AMD Radeon GPU with ROCm support (or any GPU for CPU fallback)

• v4l2loopback kernel module — the first-run installer sets it up for you

• Python 3.9+

• tk — Tkinter runtime for the GUI. Install once:

  # Arch / Manjaro
  yay -S tk            # or: sudo pacman -S tk
  
  # Debian / Ubuntu
  sudo apt install python3-tk
  
  # Fedora
  sudo dnf install python3-tkinter

  All other Python dependencies (PyTorch ROCm, OpenCV, CustomTkinter, numpy) are installed automatically into the app's own venv on first run.

---

Troubleshooting

The installer covers the common cases (venv, PyTorch ROCm, v4l2loopback, Iriun autostart). If something goes wrong:

• No ProVCam Output in Zoom/OBS — v4l2loopback isn't loaded. Run setup_boot_autostart.sh from ~/.local/share/provcam/app/ once.
• Permission denied on /dev/video2 — add your user to the video group: sudo usermod -aG video "$USER" (log out & back in).
• GPU not detected — verify rocminfo shows your GPU; reinstall the matching PyTorch ROCm wheel.

Logs live at ~/.cache/provcam/provcam_gpu.log.

---

Source code

This repository ships the prebuilt AppImages only. The source tree (Python backend, shell launchers, Rust inference crate, build scripts) lives at ~/.local/share/provcam/app/ on any machine that has run the AppImage once — it's self-extracting. From there:

cd ~/.local/share/provcam/app
./build_appimage.sh        # rebuild your own AppImage
pytest tests/              # run the test suite

---

Links

• YouTube tutorials and Linux desktop videos: @linuxlifex
• GitHub profile: ekremx25

---

Licence

MIT — do what you want, just don't claim warranty.