GitHub - Hkshoonya/nvidia-broadcast-linux: NV Broadcast — Unofficial NVIDIA Broadcast for Linux and other OS. AI-powered virtual camera with background removal, blur, replacement, video enhancement, and noise cancellation. GPU accelerated. Open source.

by DocZeus | AI Powered

NV Broadcast — Unofficial NV Broadcast and other OS. Open source. GPU accelerated. Built with passion.

Why I Built This

I left Windows. Millions of Linux users left Windows. But we all missed one thing — NVIDIA Broadcast.

That one app that made every video call look professional. Background blur that actually worked. Virtual backgrounds that didn't look like a PowerPoint slide. Noise cancellation that silenced your mechanical keyboard.

On Linux? Nothing. You had to cobble together 5 different tools, fight with v4l2loopback configs, and still get janky edges. That's over now.

I built this because I believe Linux users deserve the same broadcast-quality experience. Not a half-baked wrapper. Not a "good enough" hack. A real, proper implementation that uses your NVIDIA GPU to do what NVIDIA Broadcast does on Windows — but open source, and in some ways, better.

"Not saying this is perfect. But I believe it will be."

This is fast. This is optimized. And the quality already rivals Windows Broadcast. With the community behind it, we'll surpass it.

— DocZeus

What's New

v1.1.13 — Auto Frame Tracking and Framing Patch

• Center Face Tracking Smoothed — Auto Frame now follows side-to-side face movement continuously instead of waiting for a large lateral jump before the crop catches up
• Stable Background Mode Added — The new Auto Frame Framing selector lets users choose Center Face tracking or Stable Background framing when they do not want the room to drift
• Minimum Zoom Framing Fixed — Center Face now keeps a small internal crop margin at the minimum zoom setting, so face centering still works when users reduce zoom as far as the UI allows
• Mode Switching Fixed — Switching from Stable Background back to Center Face snaps to the next detected face immediately, then resumes normal smoothing
• Regression Tests Added — Release checks now cover lateral tracking smoothness, framing mode switching, config persistence, and minimum-zoom auto-frame behavior

If you are still on v1.1.12 or older, update to v1.1.13. This is the recommended stable patch for smoother Auto Frame tracking and clearer framing behavior.

v1.1.12 — OBS Startup, Phone Webcam, and Audio Filtering Patch

• OBS Startup Race Fixed — The main app now stops the optional headless virtual-camera service before opening the camera, preventing blank preview and busy-camera conflicts after login
• Phone Webcam Fallback Improved — Android/phone-as-webcam sources now get safer mode retries, and busy /dev/video10 is reported clearly before fallback attempts
• Audio Filtering Dependency Fixed — RNNoise mic/speaker cleanup now installs its PyAV dependency with the core app, so audio routing does not silently run unfiltered
• Complete Runtime Bundle Preserved — Meeting transcription and summaries keep the bundled faster-whisper, ctranslate2, soundfile, and support-package path; openai-whisper remains optional and guarded for compatible Python versions
• Virtual Camera Setup Recovery Improved — Existing installs update the loopback configuration safely and avoid live reloads while OBS, browsers, or meeting apps are using the camera
• Regression Tests Expanded — Release checks now cover GUI/headless ownership, phone webcam fallback, audio-filter dependency imports, Snap/package metadata, and hardware-independent vcam tests

If you are still on v1.1.11 or older, update to v1.1.12. This is the recommended stable patch for OBS startup reliability, phone webcam compatibility, audio filtering, and package runtime completeness.

v1.1.11 — OBS Camera Compatibility and Packaging Reliability Patch

• OBS White Preview Fixed — The camera pipeline now handles cameras that expose raw video modes instead of MJPEG for the selected resolution
• Safer Camera Auto-Detection — Startup now avoids stale, metadata-only, and virtual-loopback camera nodes after reboot, reducing blank preview and “no effects” cases
• Headless Virtual Camera Fixed Too — nvbroadcast-vcam uses the same camera compatibility path as the main app, so OBS-only workflows get the same fix
• CUDA Runtime Packaging Fixed — Source, Debian, RPM, and amd64 Snap installs now use the correct package paths for the CUDA mode runtime
• Regression Tests Added — Release checks now cover raw-camera fallback, camera-node filtering, headless virtual camera behavior, and package metadata consistency

If you are still on v1.1.10 or older, update to v1.1.11. This is the recommended stable patch for OBS camera compatibility and package install reliability.

v1.1.10 — Live Edge Quality and Compute Control Update

• Cleaner Live Background Edges — Background replace is steadier around hair, shoulders, raised hands, and finger gaps during motion
• Less Edge Cleanup Cost — The DocZeus/fused CUDA path now uses the same replace-mode foreground cleanup while the CPU fringe cleanup does less full-frame work
• New Compute Selector — Users can choose Auto, GPU Focused, or CPU Focused depending on whether they want automatic tuning, CUDA preference, or lower GPU load
• CUDA Install Path Fixed — Source, Debian, RPM, and amd64 Snap installs now install the CUDA mode runtime on NVIDIA systems instead of installing only CuPy and leaving ONNX inference on CPU
• Clearer GPU Runtime Checks — CUDA modes now require both CuPy compositing and the ONNX Runtime CUDAExecutionProvider, so the app no longer treats a partial GPU install as ready
• Meeting Dependencies Stay Safe — Packaged installs continue using the lighter faster-whisper runtime, and openai-whisper remains guarded for Python versions that support it
• Sponsors Are Visible — Public GitHub Sponsors now show in the README, dedicated sponsor wall, and About window so users can see who is backing the project

If you are still on v1.1.9 or older, update to v1.1.10. This is the recommended stable update for live background edge quality, compute-mode clarity, and release packaging safety.

v1.1.9 — Meeting Runtime Dependency Hotfix

• Meeting Transcription Install Fixed — The app now installs faster-whisper safely without skipping the support packages required for local transcription
• Package Installers Fixed Too — Debian, RPM, and macOS package install paths now use the same corrected meeting runtime recipe as the in-app installer
• Clearer Recovery Steps — Missing-transcription messages now show the complete command instead of only pip install faster-whisper
• No OpenAI Whisper Regression — openai-whisper remains optional and guarded on newer Python versions; the default packaged path stays on the lighter faster-whisper runtime

If you are still on v1.1.8 or older, update to v1.1.9. This is the recommended hotfix for meeting transcription dependency reliability.

v1.1.8 — Audio Helper and Installer Reliability Patch

• Fixed Echo-Like Mic Doubling — Old background audio helpers now exit with the app and are cleaned up before a new helper starts, preventing stale helpers from feeding duplicate nvbroadcast mic audio into calls
• CuPy Installer Flow Fixed — The source installer no longer aborts the whole install when the optional CuPy verification step fails after package installation
• Clearer Installer Errors — Install failures now report the real exit code and show useful CuPy verification output for future troubleshooting
• Sponsor Links Easier to Find — The About window and README now point users to GitHub Sponsors more clearly

If you are still on v1.1.7 or older, update to v1.1.8. This is the recommended stable patch for microphone reliability and source-install GPU setup.

v1.1.7 — Live Edge and Mic Reliability Patch

• Cleaner Hair and Hand Edges — Background replace now holds hair edges, finger gaps, and hands near the body more cleanly during motion
• Better Live Quality Mode Behavior — Quality profiles now stay closer to the fresher inline matte path, which reduces the “edges follow motion” look that showed up after the earlier lag fixes
• Less Face-Effect Spill Into Hair — Beautify and relighting now use tighter face-tone masks, so head hair and side hair look less bright and washed out
• Stronger Live Processing Path — The live GPU/background path, shared face-landmark worker flow, and fused face-overlay handling were tuned further for steadier real-time behavior
• Mic Always Ready — The exported nvbroadcast microphone now stays available even when voice effects and noise removal are turned off
• Broader Release Verification — This release was rechecked across video, audio, meeting transcription, summaries, packaging metadata, and release smoke before shipping

If you are still on v1.1.6 or older, update to v1.1.7. This is the recommended stable patch for current live edge quality and microphone reliability.

v1.1.6 — Live Background Performance and Stability Patch

• Fixed Background Reset Loops — The live alpha path now uses one dedicated worker instead of bouncing inference across short-lived threads, which stops the CUDA invalid-resource-handle failures that could make replace mode collapse into repeated RVM resets
• Lower Replace-Mode Live Cost — Relighting now reuses the same-frame final matte instead of rebuilding it, and replace-mode matte work is cached more aggressively on the live path
• Better Motion Handling Around Face Effects — Beautify keeps GPU work local to the face ROI and preserves raw denoise history more carefully, reducing motion smear around the face and glasses
• Safer Heavy Live Stack Behavior — The app is better at keeping the background path responsive on heavier Meeting-style stacks instead of compounding lag with duplicate work

If you are still on v1.1.5 or older, update to v1.1.6. This is the recommended stable patch for the recent live background lag regression.

v1.1.5 — Stability and Live Quality Patch

• Safer Effect + Mode Switching — Video pipeline rebuilds now wait for teardown properly, which reduces camera freezes, device-busy failures, and mode-switch crashes
• TensorRT Detection Fixed — Zeus and Killer now recognize current tensorrt-cu12 installs correctly and handle the TensorRT handoff more safely
• Duplicate Audio Devices Cleaned Up — Stale nvbroadcast mic and speaker duplicates are deduped, and startup restore no longer churns the virtual-audio path as aggressively
• Better Motion On Face + Glasses — Beautify denoise is now limited to the face ROI and keeps raw history, which reduces motion smear and disappearing glasses during movement
• Replace Mode Overlap Improved — Raised hands near shoulders and underarms are less likely to blow false holes through the background during motion

If you are still on v1.1.4 or older, update to v1.1.5. This is the recommended stable patch for switch stability, TensorRT detection, and current live-quality fixes.

v1.1.4 — Audio and Packaging Reliability Patch

• Browser-Safe Processed Mic — The Linux nvbroadcast microphone path is now stable for Chrome, Discord, Meet, and similar apps instead of hanging or opening with silence
• Meeting Runtime Fixed — The optional Whisper runtime installer now validates real imports before claiming success and includes the missing httpx dependency
• Packaged Meeting Runtime — Release installers now bundle the local meeting transcription runtime more consistently so packaged installs do not depend on the in-app runtime path as often
• No External ffmpeg Requirement For Saved Meetings — The app now reads its own saved WAV meeting capture directly for the final transcript pass
• Release Checks Tightened — Packaging CI is opted into Node 24 early, and release smoke now covers dependency-installer, transcriber, summarizer, meeting-store, and packaging metadata checks

If you are still on v1.1.3 or older, update to v1.1.4. This is the recommended stable patch for meeting audio reliability and packaged runtime consistency.

v1.1.3 — Meeting-First Live Quality Update

• Lower Visible Delay — Live video now prefers the newest frames instead of letting stale buffers build up, which reduces the odd “lips move late” effect in meetings
• Auto - Adaptive Mode — Hardware-aware tuning now persists across restart, warns on very weak devices, and can recommend lighter capture modes when real-time FPS collapses
• Better Live Face Effects — Face relighting is now fill-light biased instead of darkening the face, and eye contact is more conservative so it distorts less
• Meeting-Selectable Mic — The processed mic is exported as nvbroadcast, so Zoom, Meet, Teams, OBS, and similar apps can select it directly
• Correct Speaker Routing — Speaker denoise now honors the selected output device instead of whichever sink the system guessed
• Release + Update Flow — The app now surfaces platform-aware upgrade targets for GitHub releases, macOS .pkg downloads, and Snap installs

If you are still on v1.1.2 or older, update to v1.1.3. This is the recommended stable build for live meetings, adaptive tuning, and corrected audio routing.

v1.1.2 — Priority Stability Update

• Meeting Transcript Quality — Better chunk cleanup and a stronger final full-audio pass improve saved transcripts and notes
• No End-Meeting Freeze — Meeting transcript, notes, and summary finalization now run off the UI thread
• Persistent Speaker + Profile State — Speaker selection and the active profile now restore correctly after restart
• Reset to Defaults — One-click recovery back to a known-good baseline
• Mic Test Fixed — Recording and playback are more reliable, with 30s, 45s, and 60s capture options plus early stop

If you are still on v1.1.0 or v1.1.1, update to v1.1.2. It is the recommended stable patch for meeting quality, persistence, and audio test reliability.

v1.1.1 — Stability Patch

• Virtual Camera Stability — Safer Linux v4l2loopback sink startup and retry handling
• Lower Live Lag — Shared face landmarks and face-ROI relighting reduce delay in heavier effect stacks
• Better Replace Edges — Tighter shoulders, ear-side hair, and under-arm gaps during background replace
• Meeting Transcription Reliability — Faster startup, shorter chunking, and cleaner saved meeting audio
• Resolution Change Safety — Resolution changes are saved safely and applied after restart instead of hanging the stream

If you are still on v1.1.0, update to v1.1.1. It improved virtual-camera behavior, lower lag, and cleaner live compositing.

v1.1.0 — Meeting Assistant Update

• Meeting Assistant Sidebar — Collapsible live transcript and rolling summary inside the app
• Meeting History — Local session history stays on-device for 7 days with automatic cleanup
• Two-Way Meeting Audio — Meeting capture records both sides for better local notes and transcripts
• Background Runtime Installs — Optional CUDA, TensorRT, and meeting runtimes install in the background with progress
• Improved Setup Guidance — First-run flow explains modes, downloads, and skip/install choices more clearly

v1.0.0 — AI Release

• AI Meeting Transcription — Local Whisper speech-to-text (tiny/base/small/medium models, GPU-accelerated)
• AI Meeting Summarizer — Extracts action items, questions, key points from transcripts (fully local)
• Voice Effects — Bass boost, treble, warmth, compression, noise gate, gain (GPU + CPU)
• 6 Voice Presets — Natural, Radio, Podcast, Deep Voice, Bright, Studio
• Microphone Selection — Full PipeWire/PulseAudio device enumeration
• Speaker Detection — All output devices via PipeWire
• Audio Level Monitor — Real-time VU meter with peak hold
• Mic Test — Record 30s / 45s / 60s and play back to test your setup
• Meeting Mode — Combined video+audio recording with live transcription and AI summary
• Recording Fix — MP4 now includes audio track (NVENC video + AAC audio)
• Voice FX GPU Acceleration — CuPy CUDA for warmth/gate/gain, scipy for filters (2.8ms/chunk)

v0.3.0

• Eye Contact Correction — MediaPipe iris tracking redirects your gaze to look at camera
• Face Relighting — Fill light guided by the scene
• Recording Mode — NVENC hardware encode to MP4 (x264 fallback on non-NVIDIA)
• Performance Overlay — Real-time FPS, GPU usage, VRAM, temperature monitoring
• User Profiles — 5 built-in (Meeting, Streaming, Presentation, Gaming, Clean) + custom save/load
• Multi-Camera Support — Hot-switch between cameras without restarting
• Apple-Inspired UI — Glassmorphism cards, collapsible sections, smooth transitions
• Shared FaceLandmarker — Single MediaPipe instance shared across all face effects (3x faster)
• macOS Support — CPU modes with CoreML, AVFoundation camera, Homebrew installer
• CI Pipeline — GitHub Actions builds .deb, .rpm, .pkg + Swift Camera Extension on macOS

v0.2.0

Premium GPU Modes

• Killer Mode — Fused CUDA kernel + 360p inference = 48fps at 1080p (20ms/frame)
• Zeus Mode — 480p optimized inference = 33fps at 1080p (30ms/frame)
• DocZeus Mode — Fused CUDA kernel compositing = CUDA Max quality at 150x faster blend (0.1ms vs 15ms)

Edge Refinement Neural Network

• Toggle-activated second-pass inference at 720p for Zeus/Killer modes
• Uses RVM ResNet50 at full resolution with morphological edge band blending
• 89.9% quality recovery — brings fast modes close to max quality edges

Video Enhancement

• 5 independent effects: Skin Smooth, Denoise, Enhance, Sharpen, Edge Darken
• 4 presets: Natural, Broadcast, Glamour, Custom
• Per-effect toggle + intensity slider
• MediaPipe FaceLandmarker at half-res, every 5th frame
• GPU batch processing (CuPy) for enhance + sharpen + vignette

Resolution & FPS Selector

• Auto-detects camera capabilities via v4l2
• Shows only supported resolutions (360p to 4K)
• FPS dropdown adapts per resolution (e.g., 4K shows 30fps, 1080p shows 30+60fps)
• Validated before pipeline start — no more cap negotiation hangs

UI Improvements

• Resizable preview — drag the divider between preview and controls
• Pause View — freeze the preview display (camera keeps running)
• Hide Preview — collapse preview entirely for more control space
• Mirror toggle — horizontal flip for webcam view
• Scrollable controls — all settings accessible regardless of window size
• Grouped cards — Input, Processing, Background, Auto Frame, Beauty

Performance Optimizations

• Pre-downsampling: Frames above 720p are downsampled before inference (124ms -> 29ms at 1080p)
• Async effects processing: Capture thread never blocks — zero preview latency
• Python-side frame throttling: No pipeline restart for mode/profile changes
• Fused CUDA kernel: Single GPU pass for alpha blend + enhance + vignette (0.1ms)

What It Does

Processing Modes

9 modes from maximum speed to CPU fallback:

Edge Refine toggle available for Killer and Zeus modes — adds ~27ms but recovers 89.9% of max quality edges.

Switch modes anytime from the Mode dropdown. No restart needed.

CUDA modes require the CUDA mode runtime: CuPy for compositing plus ONNX Runtime with CUDAExecutionProvider for model inference. Source, .deb, .rpm, and amd64 Snap installs handle this automatically on NVIDIA systems. The arm64 Snap build stays CPU-safe because ONNX Runtime GPU wheels are not published for Linux arm64 yet.

Architecture

                         NV Broadcast v1.0.0
                         ─────────────────────────────────

  ┌───────────┐      ┌──────────────────────────────────────────┐      ┌──────────────┐
  │  Webcam   │─────▶│           GStreamer Pipeline              │─────▶│ Virtual Cam  │
  │(360p-4K)  │      │                                          │      │ /dev/video10 │
  └───────────┘      │  JPEG Decode ─▶ Color Convert ─▶ appsink │      └──────┬───────┘
                     └──────────────────────┬───────────────────┘             │
                                            │                        ┌───────▼───────┐
                            ┌───────────────▼──────────────┐         │ Chrome / Zoom │
                            │    Async Effects Thread       │         │ Firefox / OBS │
                            │   (never blocks capture)      │         │ Discord/Meet  │
                            │                               │         └───────────────┘
                            │  ┌─────────────────────────┐  │
                            │  │   AI Segmentation        │  │
                            │  │                          │  │
                            │  │  Pre-downsample to 720p  │  │
                            │  │  (or 480/360 for Zeus/   │  │
                            │  │   Killer modes)          │  │
                            │  │                          │  │
                            │  │  ┌────┐ ┌─────┐ ┌────┐  │  │
                            │  │  │RVM │ │ISNet│ │BiR │  │  │
                            │  │  └──┬─┘ └──┬──┘ └─┬──┘  │  │
                            │  │     └───┬──┘      │     │  │
                            │  │         ▼         │     │  │
                            │  │   Alpha Refine    │     │  │
                            │  │  (sigmoid+dilate) │     │  │
                            │  └────────┬──────────┘     │  │
                            │           │                 │  │
                            │  ┌────────▼───────────────┐ │  │
                            │  │  Edge Refiner (opt.)   │ │  │
                            │  │  720p 2nd pass RVM     │ │  │
                            │  │  (Zeus/Killer only)    │ │  │
                            │  └────────┬───────────────┘ │  │
                            │           │                 │  │
                            │  ┌────────▼───────────────┐ │  │
                            │  │     Compositing        │ │  │
                            │  │                        │ │  │
                            │  │  ┌────────┐ ┌───────┐  │ │  │
                            │  │  │ Fused  │ │ CuPy  │  │ │  │
                            │  │  │ CUDA   │ │ CUDA  │  │ │  │
                            │  │  │ 0.1ms  │ │ 15ms  │  │ │  │
                            │  │  └────────┘ └───────┘  │ │  │
                            │  └────────────────────────┘ │  │
                            │           │                 │  │
                            │  ┌────────▼───────────────┐ │  │
                            │  │   Video Enhancement  │ │  │
                            │  │  5 effects + presets   │ │  │
                            │  │  GPU batch (CuPy)      │ │  │
                            │  └────────────────────────┘ │  │
                            │           │                 │  │
                            │  Mirror flip (optional)     │  │
                            └───────────┬─────────────────┘
                                        │
                            ┌───────────▼──────────────┐
                            │ Preview (GTK4 Texture)    │
                            │ Pause / Hide / Resize     │
                            └──────────────────────────┘

  ┌───────────┐      ┌─────────────────────────────────┐      ┌──────────────┐
  │    Mic    │─────▶│     RNNoise AI Denoise          │─────▶│ Virtual Mic  │
  │           │      │     (48kHz, 10ms frames)        │      │  (PipeWire)  │
  └───────────┘      └─────────────────────────────────┘      └──────────────┘

Fused CUDA Kernel (DocZeus/Killer)

A custom CUDA kernel that performs alpha blend + enhance + sharpen + vignette in one GPU pass:

// Single kernel: fg*alpha + bg*(1-alpha) + enhance + vignette
// 0.1ms at 1080p — 150x faster than CuPy's multi-kernel approach
extern "C" __global__ void fused_composite(
    fg, bg, alpha, face_mask, vignette, output,
    total_pixels, enhance_i, vignette_i, brightness, contrast, warmth
);

Edge Refinement Network

When Edge Refine is toggled ON (Zeus/Killer modes):

1. Fast pass: RVM at 360p/480p → coarse alpha (18-21ms)
2. Refine pass: RVM ResNet50 at 720p → quality alpha (30ms, every 2nd frame)
3. Blend: On refine frames use quality alpha; on skip frames 80% quality + 20% coarse for tracking
4. Result: 89.9% quality recovery with minimal cost

AI Models

Quality Presets (RVM only)

Requirements

Hardware

Software

• Linux with NVIDIA driver 525+ (Pop!_OS, Ubuntu, Fedora, Arch, openSUSE, etc.)
• Python 3.11+
• PipeWire (virtual microphone)
• PulseAudio utilities (pactl) for speaker-monitor routing and device resolution
• GStreamer 1.20+ with plugins-base, plugins-good, plugins-bad
• GTK4 and Libadwaita
• v4l2loopback kernel module
• DKMS and kernel headers (to build v4l2loopback)

Installation

Linux — One Command Install

git clone https://github.com/Hkshoonya/nvidia-broadcast-linux.git
cd nvidia-broadcast-linux
./install.sh

macOS — One Command Install

git clone https://github.com/Hkshoonya/nvidia-broadcast-linux.git
cd nvidia-broadcast-linux
./install_macos.sh

Requires macOS 12+, Homebrew, Python 3.11+. Installs GStreamer, GTK4 via Homebrew. CPU modes with CoreML acceleration on Apple Silicon. GPU modes (Killer/Zeus/DocZeus/CUDA) are Linux-only and require an NVIDIA GPU.

Linux — Snap Package

sudo snap install nvbroadcast

Snap users typically receive background refreshes from snapd. When the app sees a newer stable release, the in-app update button opens the Snap Store listing so the user can move directly into the store-managed upgrade path.

The amd64 Snap build includes the CUDA mode runtime for NVIDIA systems. The arm64 Snap build stays CPU-safe because the required ONNX Runtime GPU wheels are not available for Linux arm64 yet. If CUDA modes are still unavailable on amd64 Snap, use the source installer, .deb, or .rpm release package as the fallback.

Packaged releases are intended to include the local meeting transcription runtime. Source installs from this repo can still use the in-app runtime installer flow for optional components.

Linux Installer Details

The installer:

1. Detects your distro and package manager
2. Checks all requirements (Python, PipeWire, GPU, DKMS, kernel headers)
3. Installs missing packages with the correct names for your distro
4. Installs NVIDIA CUDA mode runtime packages when an NVIDIA GPU is detected
5. Asks about compositing — CPU, GStreamer GL, or CuPy CUDA
6. Sets up virtual camera, launcher scripts, desktop entry, systemd service
7. Verifies GPU acceleration and writes initial config
8. Lets optional runtimes install later inside the app without blocking the rest of the UI

Update Behavior

• Git checkout / manual Linux packages — the app checks GitHub Releases and opens the matching release download page when a newer stable build is available
• macOS package installs — the app prefers the latest .pkg release asset when one is published
• Snap installs — the app opens the Snap Store listing; stable refreshes are normally handled by snapd

Optional: TensorRT (for Zeus/Killer modes)

.venv/bin/pip install tensorrt-cu12 onnx

TensorRT Python wheels are currently published for Python 3.8 through 3.13 on Linux x86_64. If you are on Python 3.14+, use DocZeus or the CUDA modes instead.

Supported Distros

# 1. System dependencies
sudo apt install -y \
    python3-gi python3-gi-cairo \
    gir1.2-gtk-4.0 gir1.2-adw-1 \
    gir1.2-gstreamer-1.0 gir1.2-gst-plugins-base-1.0 \
    gstreamer1.0-plugins-base gstreamer1.0-plugins-good \
    gstreamer1.0-plugins-bad \
    v4l-utils v4l2loopback-dkms \
    pipewire-bin pulseaudio-utils

# 2. Python venv
python3 -m venv .venv --system-site-packages
source .venv/bin/activate

# 3. Install
pip install -e .

# For NVIDIA GPU acceleration, install the CUDA extra instead:
pip install -e ".[cuda]"

# 4. Optional: CuPy-only retry for GPU compositing
pip install cupy-cuda12x nvidia-cuda-nvrtc-cu12

# 5. Virtual camera
sudo modprobe v4l2loopback devices=1 video_nr=10 \
    card_label="NVbroadcast" exclusive_caps=1 max_buffers=4

# 6. Run
python -m nvbroadcast

Usage

Setup Once, Forget Forever

nvbroadcast          # Launch GUI (first time: setup wizard)

1. Setup wizard detects your system and configures the best mode
2. App starts and auto-begins streaming
3. Configure effects, select resolution/FPS/mode
4. Close the window — app minimizes to background, virtual camera stays active
5. Open Chrome / Zoom / Discord / OBS — select "NVbroadcast" as your camera
6. Next login — app starts automatically with all your settings remembered

Controls

Headless Mode

nvbroadcast-vcam                    # No GUI, just the virtual camera
nvbroadcast-vcam --format i420      # Firefox-compatible format

As a System Service

Use this only for no-GUI/headless passthrough workflows. Do not run the headless service at the same time as the GUI app, because both need exclusive access to the physical camera and NVbroadcast virtual camera.

systemctl --user enable --now nvbroadcast-vcam

# If you use the GUI app instead:
systemctl --user disable --now nvbroadcast-vcam

The headless command is a passthrough producer for OBS/browser workflows. For full background effects, start the main NVbroadcast app first, then select the NVbroadcast camera in OBS or your meeting app.

Troubleshooting

sudo modprobe -r v4l2loopback
sudo modprobe v4l2loopback devices=1 video_nr=10 card_label="NVbroadcast" exclusive_caps=1 max_buffers=4

.venv/bin/pip install --upgrade ".[cuda]"

.venv/bin/python -c "import onnxruntime as ort; print(ort.get_available_providers())"

v4l2-ctl -d /dev/video0 --list-formats-ext   # Check supported resolutions

Project Structure

nvidia-broadcast-linux/
├── src/nvbroadcast/
│   ├── __init__.py              # Package version (1.1.13)
│   ├── app.py                   # GTK4 app: modes, effects, pipeline management
│   ├── vcam_service.py          # Headless virtual camera service
│   ├── core/
│   │   ├── config.py            # TOML config, performance profiles, compositing backends
│   │   ├── constants.py         # App ID, paths, GPU config
│   │   └── gpu.py               # GPU detection, CUDA device mapping
│   ├── video/
│   │   ├── effects.py           # Multi-model engine, fused CUDA kernel, edge refiner
│   │   ├── pipeline.py          # GStreamer pipeline, async effects, frame throttling
│   │   ├── beautify.py          # Video enhancement (5 effects + GPU batch)
│   │   ├── autoframe.py         # MediaPipe face tracking with smooth zoom/pan
│   │   └── virtual_camera.py    # v4l2loopback + camera capability query
│   ├── audio/
│   │   ├── effects.py           # RNNoise denoiser
│   │   ├── pipeline.py          # GStreamer audio pipeline
│   │   ├── monitor.py           # Speaker output denoise
│   │   └── virtual_mic.py       # PipeWire virtual microphone
│   └── ui/
│       ├── window.py            # Main window: resizable paned layout, 9 modes
│       ├── setup_wizard.py      # First-run wizard
│       ├── controls.py          # Effect toggles, sliders, file picker
│       ├── device_selector.py   # Dropdown selector (single-connect fix)
│       ├── video_preview.py     # Live video preview
│       └── style.css            # App styling with Adwaita/system theme integration
├── models/                      # AI models (auto-downloaded)
│   ├── rvm_mobilenetv3_fp32.onnx
│   ├── rvm_resnet50_fp32.onnx
│   ├── rvm_mobilenetv3_fp16.onnx   # Lightweight refiner model
│   ├── rvm_resnet50_fp32_trt.onnx  # TensorRT shape-inferred
│   └── rvm_mobilenetv3_fp32_trt.onnx
├── install.sh                   # Multi-distro installer
├── uninstall.sh                 # Clean removal
├── pyproject.toml               # Package config (v1.1.13)
└── README.md

Contributing

Contributions, feedback, and ideas are warmly welcome.

How to Contribute

1. Fork this repository
2. Create a branch (git checkout -b feature/amazing-thing)
3. Commit with clear messages
4. Open a Pull Request

Report Issues

Found a bug? Open an issue.

Ideas for Contribution

• Eye contact correction (v0.3.0)
• Virtual lighting / face relighting (v0.3.0)
• System tray indicator (v0.2.0)
• Multi-camera support (v0.3.0)
• Recording mode (v0.3.0)
• Performance overlay (FPS, GPU usage) (v0.3.0)
• GStreamer NVDEC/NVENC hardware codec pipeline (v0.3.0)
• NVIDIA Maxine SDK integration
• Flatpak packaging
• Snap packaging

Future Upgrades

• Meeting lip-sync compensation — explicit audio/video delay calibration so heavy live video stacks still land naturally in calls
• Per-device auto benchmark — benchmark each camera mode and effect stack once, then pin the best stable settings for that machine
• Speaker diarization — separate “me” vs “remote speaker” in live meeting transcripts and saved notes
• Local live captions — optional on-screen captions and confidence-aware subtitle output for streams and calls
• Multi-person framing — presenter mode for interviews, podcasts, and side-by-side calls
• AI meeting memory — on-device semantic search across prior meetings, summaries, action items, and decisions
• Scene-aware relighting — stronger face light that reacts to background direction, exposure, and skin tone without flattening the face
• Quality advisor — explain exactly which effect, resolution, or backend is costing FPS on the current hardware

Sponsor This Project

If NV Broadcast saves you from going back to Windows, consider sponsoring.

This project takes ongoing work across GPU runtimes, packaging, camera/audio compatibility, meeting features, and cross-distro bug fixes. Sponsorship helps keep that work moving without turning the app into adware or locking core features behind a paywall.

• Individual sponsors help fund day-to-day fixes, release maintenance, and new features
• Creator and business sponsors help cover hardware testing, distro support, packaging, and priority reliability work
• All sponsors help keep the Linux version improving instead of stagnating behind Windows-only tools

💎 Featured Sponsors

💚 Backers & Supporters

GitHub Sponsors · Report bugs · Share ideas

License

• Python app & Linux code: GPL-3.0 — see LICENSE
• macOS Camera Extension (macos/): Proprietary — see macos/LICENSE

Any redistribution or derivative work must retain the original author attribution.

Created with passion by DocZeus
Because Linux users deserve broadcast-quality video too.

_{Copyright (c) 2026 DocZeus. All rights reserved under GPL-3.0.}