Intelligent GPU-accelerated frame extraction for presentation and lecture videos

Smart Extract is a high-performance Python tool that automatically finds and exports the best possible still frames from long presentation recordings such as talks, lectures, demos, and keynotes.

Instead of dumping thousands of near-identical frames, Smart Extract scans the entire video, scores visual quality, enforces temporal diversity, and optionally applies AI super-resolution to produce a curated, portfolio-ready image set.

Most frame-grab tools blindly sample frames by time.

Smart Extract understands quality.

It evaluates every candidate frame and keeps only the best moments.

• 🧠 Global Best-Frame Selection
  Scores frames by sharpness + brightness across the entire video

• ⏱ Temporal Diversity Enforcement
  Prevents near-duplicate frames from the same moment

• ⚡ GPU-Accelerated Scoring (Optional)
  Uses PyTorch and CUDA for fast analysis on RTX GPUs

• 🖼 AI Super-Resolution (4x)
  Optional RealESRGAN upscaling for ultra-clean slides and thumbnails

• ✂ Auto-Crop for Presentation Screens
  Removes dead space and zooms into the slide content

• 🎨 Optional Color Normalization
  Corrects projector color casts such as purple or yellow shifts

• 🚀 Fast Video Decoding
  Uses Decord when available with OpenCV fallback

1. Scan the entire video at a configurable low FPS
2. Score each sampled frame using:
   • Laplacian sharpness
   • Mean brightness
3. Rank all frames globally by visual quality
4. Select the top K frames while enforcing a minimum time gap
5. Post-process selected frames:
   • Optional crop
   • Optional color correction
   • Optional AI upscaling
6. Save only the final best frames

Result: A clean, diverse, high-quality image set instead of noise.

flowchart TD
    A[Input Video File] --> B[Frame Sampling]
    B --> C[Sharpness Scoring]
    B --> D[Brightness Scoring]
    C --> E[Global Frame Ranking]
    D --> E[Global Frame Ranking]
    E --> F[Temporal Diversity Filter]
    F --> G[Best Frame Selection]
    G --> H[Optional Auto Crop]
    H --> I[Optional Color Correction]
    I --> J[Optional AI Super Resolution]
    J --> K[Final High Quality Frames]

• 🎤 Conference talks and keynote recordings
• 🎓 University lectures and classes
• 🧑‍🏫 Internal presentations and demos
• 📊 Slide reconstruction from recorded sessions
• 🖼 Thumbnail generation
• 📁 Portfolio or documentation screenshots

• Python 3.9 or higher
• Windows, Linux, or macOS

• NVIDIA GPU with CUDA support
• RTX series GPU for best performance

Clone the repository:

git clone https://github.com/EyalPasha/smart-extract.git
cd smart-extract

Install dependencies:

pip install -r requirements.txt

This installs:

• torch (CUDA-enabled if available)
• opencv-python
• numpy
• decord for fast video decoding
• realesrgan and basicsr for AI upscaling

All behavior is controlled via config.py with no code changes required.

VIDEO_PATH = "IMG_8313.MOV"
OUTPUT_DIR = "photos"
OUTPUT_FORMAT = "PNG"

USE_GLOBAL_BEST_FRAMES = True
GLOBAL_TARGET_FRAMES = 100
GLOBAL_SAMPLE_FPS = 2.0
GLOBAL_MIN_TIME_BETWEEN = 8.0
GLOBAL_START_TIME_SECONDS = 27.0

This mode:

• Scans the entire video
• Picks the top K frames globally
• Ensures visual and temporal diversity

Output directory:

photos/best

USE_GPU_SCORING = True
ENABLE_SUPER_RESOLUTION = True

• GPU scoring accelerates sharpness and brightness evaluation
• RealESRGAN upscales selected frames by 4x for maximum clarity

Disabled by default for maximum fidelity:

ENABLE_AUTO_CROP = False
ENABLE_COLOR_CORRECTION = False

Enable these options if your recording contains:

• Excess dead space
• Projector color casts

python smart_extract.py

You will see:

• Device detection (CPU or CUDA)
• Global scan progress bar
• Final selection summary

Example output:

Done! Saved 100 global best frames to photos/best

Each saved frame includes metadata in the filename:

best_042_t318.4s_score912.png

• 042 is the rank
• 318.4s is the timestamp
• score912 is the final quality score

Perfect for sorting, filtering, and automation.

• If frames look too similar, increase GLOBAL_MIN_TIME_BETWEEN
• If key moments are missing, increase GLOBAL_SAMPLE_FPS
• If VRAM is limited, reduce the RealESRGAN tile size in smart_extract.py

Smart Extract is a production-ready visual curation pipeline.

It demonstrates:

• Computer vision fundamentals
• GPU acceleration with PyTorch
• Practical AI upscaling
• Performance-aware video processing
• Clean and configurable system design

Ideal for portfolio showcase, research tooling, or real-world media workflows.

MIT License

⭐ If this project helped you, consider starring the repository.