Video Stabbot

Desktop video stabilization for quick fixes, manual FFmpeg tuning, OpenCV feature tracking, and RAFT dense optical flow.

Video Stabbot is an Electron app for stabilizing shaky footage from a single interface. It combines FFmpeg vidstab, OpenCV sparse optical-flow tracking, and RAFT dense optical flow so users can choose between speed, control, and maximum motion-analysis quality.

File selection Mode selection Custom settings Processing

• Quick FFmpeg Mode - Fast two-pass stabilization using FFmpeg vidstab with bicubic transforms, recommended sharpening, optional audio mapping, and practical defaults.

• Custom Mode - Full manual control over vidstab parameters:

  • Border mode, smoothing strength, shakiness detection
  • Accuracy, auto/manual zoom, zoom speed, interpolation method
  • Tripod mode, relative transforms, max shift, max rotation
  • Step size, minimum contrast, encoding quality

• OpenCV Feature Tracking (requires Python, OpenCV, NumPy, and SciPy) - Sparse optical-flow stabilization with robust trajectory smoothing.

  • Auto Track mode uses Shi-Tomasi corners and pyramidal Lucas-Kanade tracking.
  • Features are distributed across a grid so motion estimates do not overfit one textured area.
  • SIFT, ORB, and AKAZE can seed tracking points when preferred.
  • Forward/backward track validation rejects unstable points.
  • ECC refinement improves the tracked transform when image alignment converges.
  • Farneback dense optical flow is used as a fallback when sparse tracks are weak.
  • RANSAC estimates camera motion while rejecting moving-object outliers.
  • Output keeps the selected resolution and muxes source audio.

• RAFT Deep Learning (requires PyTorch and torchvision) - Dense mesh-flow stabilization for hard footage.

  • Uses torchvision RAFT large weights for Sintel-style or Things-style motion.
  • Uses a 7 x 9 local mesh by default instead of only one global camera path.
  • Smooths local mesh trajectories to reduce parallax and rolling local jitter.
  • Automatically caps analysis resolution for practical memory use.
  • Converts dense flow to robust global camera motion with RANSAC.
  • Falls back to sparse tracking when dense flow is unreliable.
  • Output keeps the selected resolution and muxes source audio.

• Faster startup checks - Python packages are probed in one lightweight pass instead of importing each dependency.
• GPU acceleration - Auto-detects NVIDIA NVENC, Intel QSV, AMD AMF, or Apple VideoToolbox; falls back to CPU libx264.
• Smart dependency detection - Advanced modes appear disabled with tooltip explanations when prerequisites are missing.
• Drag-and-drop or file picker - Easy video input.
• Real-time progress tracking - Phase labels, progress bar, and elapsed time.
• Tooltips - Hover over any setting for detailed explanations.
• Bidirectional slider/input controls - Type values manually or use sliders.

• Node.js v18 or later - https://nodejs.org
• FFmpeg with libvidstab support - must be on your system PATH
  • Download from https://ffmpeg.org/download.html
  • Verify: ffmpeg -filters | findstr vidstab on Windows or ffmpeg -filters | grep vidstab on macOS/Linux

• Python 3.8+ on your system PATH

• Required packages:

  pip install opencv-python numpy scipy

• Python 3.8+ on your system PATH

• PyTorch and torchvision:

  # GPU version for NVIDIA CUDA
  pip install torch torchvision --index-url https://download.pytorch.org/whl/cu128
  # CPU-only version
  pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu

• Additional packages:

  pip install opencv-python numpy scipy

First RAFT run may download model weights. GPU is strongly recommended.

git clone <repo-url>
cd Video-Stabbot
npm install
pip install -r requirements.txt

npm start

From the app:

1. Select a video - Drag onto the drop zone or click to browse.
2. Choose a stabilization mode - Quick FFmpeg, Custom, OpenCV Features, or RAFT Deep Learning.
3. Configure settings - Each advanced mode has its own settings view with tooltips.
4. Pick a save location - Choose where to save the stabilized video.
5. Wait for processing - Progress shows the current phase and elapsed time.
6. View result - Open the output folder or process another video.

Video-Stabbot/
|-- package.json
|-- requirements.txt
|-- README.md
|-- screenshot-file-selection.jpg
|-- screenshot-mode-selection.jpg
|-- screenshot-custom-settings.jpg
|-- screenshot-processing.jpg
|-- src/
| |-- main/
| | |-- main.js
| | |-- detect-worker.js
| | `-- preload.js
| `-- renderer/
| |-- index.html
| |-- renderer.js
| `-- styles.css
`-- scripts/
 |-- smoothing_lib.py
 |-- opencv_feature_tracking.py
 `-- raft_dense_motion.py

• OpenCV: Install pip install opencv-python numpy scipy.
• RAFT: Install pip install torch torchvision opencv-python numpy scipy.
• Hover over the disabled mode tile for the specific missing dependency.

• Install PyTorch with CUDA support for GPU acceleration.
• Reduce refinement iterations only when speed matters more than quality.
• Use OpenCV mode on CPU-only systems when speed matters.

• Download an FFmpeg build with libvidstab support.
• Verify with ffmpeg -filters | grep vidstab.
• Windows users can try builds from https://www.gyan.dev/ffmpeg/builds/.

To package the app as a standalone Windows executable:

npm run dist:win

Output files are created in dist/:

• video-stabbot Setup <version>.exe
• video-stabbot <version>.exe if the portable target is generated

Optional build commands:

npm run pack
npm run dist

Python dependencies must still be installed separately by end users for advanced modes.

Commit source and project metadata only:

• src/
• scripts/
• package.json
• package-lock.json
• requirements.txt
• README.md
• screenshot jpgs or other intentional assets
• .gitignore

Do not commit generated build artifacts such as dist/, installers, portable executables, or unpacked app folders.

After running npm run dist:win, upload release artifacts from dist/:

• video-stabbot Setup <version>.exe
• video-stabbot <version>.exe if generated

Recommended release steps on Windows:

1. Update version in package.json.
2. Commit and push changes.
3. Build release artifacts with npm run dist:win.
4. Create and push a tag such as v2.0.1.
5. Draft a GitHub release for that tag and upload the .exe assets.

MIT License - see LICENSE file for details.

• FFmpeg vidstab filters for the quick and custom stabilization paths.
• OpenCV for feature detection, Lucas-Kanade optical flow, and robust affine estimation.
• RAFT for dense optical flow.
• SciPy for trajectory smoothing.