A sophisticated AI-powered image gallery application that enables intelligent searching through your photo collection using natural language queries. SmartGallery combines computer vision, embeddings, and a user-friendly desktop interface to organize and discover images effortlessly.
- AI-Powered Search: Search images using natural language descriptions powered by CLIP embeddings
- Automatic Captions: Generate descriptive captions for images using Vision Encoder-Decoder models
- Smart Tagging: Automatically extract relevant tags from image captions using KeyBERT
- Album Organization: Organize images into albums based on folder structure
- Infinite Scroll: Smooth loading of image thumbnails with infinite scroll functionality
- Full-Screen Viewer: View images in a dedicated full-screen dialog
- Real-Time Updates: Monitor directories for new/deleted images and update embeddings on the fly
- GPU Support: Utilizes CUDA for faster processing when available
Demo.mp4
Shows album-based navigation with infinite scrolling thumbnails.
Natural language search example using query "cat" filtered inside albums.
Click any image to open it in a full-size dedicated viewer.
SmartGallery uses a multi-process architecture for optimal performance:
- Main Application (
app.py): PyQt5-based GUI for browsing and searching - Search Server (
search_server.py): Handles AI-powered image search queries - Encoder Server (
encoder_server.py): Generates embeddings for new images in real-time - Pipeline (
image_captioning_clip_pipeline.py): Preprocessing script to generate captions and embeddings
- Python 3.8+
- PyQt5: Desktop GUI framework
- PyTorch: Deep learning framework
- CLIP: Multi-modal embeddings
- FAISS: Efficient similarity search
- Transformers: Pre-trained models
- KeyBERT: Keyword extraction
- NumPy, Pandas, Pillow: Data processing
- GPU with CUDA support (optional but recommended)
- Minimum 8GB RAM for processing large image collections
- SSD storage for faster image loading
git clone https://github.com/Uni-Creator/SmartGallery.git
cd SmartGallerypip install -r requirements.txtOrganize your images in a folder structure:
PHOTOS/
├── Category1/
│ ├── Folder1/
│ │ ├── image1.jpg
│ │ └── image2.jpg
│ └── Folder2/
│ └── image3.jpg
└── Category2/
└── image4.jpg
Before running the application, generate embeddings for your images:
python image_captioning_clip_pipeline.pyThis will:
- Generate captions for each image
- Extract tags from captions
- Create CLIP embeddings for both images and captions
- Save outputs to the
data/andembeddings/directories
python app.pyUpdate the BASE_FOLDER variable in app.py to point to your photo collection:
BASE_FOLDER = r"D:\Your\Photo\Path" # Update this path- Browse Albums: Select albums from the left sidebar to view grouped images
- View Thumbnails: Scroll through image thumbnails in a 4-column grid
- Open Full Image: Click any thumbnail to view the full-resolution image
- Enter a natural language query in the search box (e.g., "sunset over water", "people at beach")
- Press ENTER to search
- Results will be filtered from the current album
- Clear the search box and press ENTER to show all images again
- JPG/JPEG
- PNG
- BMP
- WebP
SmartGallery/
├── app.py # Main PyQt5 GUI application
├── search_server.py # Search query processing server
├── encoder_server.py # Real-time embedding encoder
├── image_captioning_clip_pipeline.py # Preprocessing pipeline
├── text_search.py # Search engine logic
├── pre-processing.py # Data preprocessing utilities
├── requirements.txt # Python dependencies
├── data/
│ ├── final_cleaned_data.csv # Original image metadata
│ └── images_with_captions_and_tags.csv # Captions and tags
└── embeddings/
├── image_embeddings.npy # Raw image embeddings
├── image_embeddings_normalized.npy # Normalized image embeddings
├── caption_embeddings.npy # Caption embeddings
└── image_faiss_index.idx # FAISS index for fast search
Asynchronous background worker for generating and caching image thumbnails to prevent UI freezing.
Main PyQt5 window managing:
- Album selection and navigation
- Grid layout with infinite scroll
- Search interface
- Subprocess management (search & encoder servers)
Handles semantic search using CLIP embeddings with configurable similarity thresholds.
Real-time service that:
- Listens for new image paths
- Generates embeddings using CLIP
- Updates FAISS index
- Maintains metadata CSV files
In text_search.py, modify the alpha parameter in the search methods to adjust filtering sensitivity.
Edit in app.py ThumbnailWorker.process_queue():
image = image.scaled(200, 200, ...) # Adjust dimensionsModify in app.py load_batch():
if col >= 4: # Change 4 to desired columns per rowIn app.py GalleryWindow.__init__():
self.batch_size = 100 # Images loaded per scroll- Use GPU: Ensure CUDA is properly installed for ~5-10x faster processing
- Normalize Embeddings: Pre-normalized embeddings improve search speed
- FAISS Optimization: For large collections (>100k images), consider GPU-accelerated FAISS
- Caching: Thumbnails are cached in memory; increase batch size for slower systems
- Verify
search_server.pycan run independently:python search_server.py - Check that embeddings files exist in
embeddings/directory
- Reduce
BATCH_SIZEinimage_captioning_clip_pipeline.py - Process images in smaller chunks
- Reduce thumbnail batch size
- Install PyTorch with CUDA support:
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118 - Verify NVIDIA drivers are installed
- Run
image_captioning_clip_pipeline.pyto generate missing embeddings - Ensure CSV file paths are correct
- [❌] Multi-threaded image processing pipeline
- [❌] Web-based interface (Flask/React)
- [✅] Batch image uploading with automatic processing
- [❌] Advanced filtering and faceted search
- [❌] Image clustering and recommendations
- [❌] Database backend (PostgreSQL + pgvector)
- [❌] REST API for programmatic access
- [❌] Image deduplication detection
- [❌] Automatically remove deleted image's embeddings
This project is licensed under the MIT License. See the LICENSE file for details.
Contributions are welcome! Please feel free to submit pull requests or open issues for bugs and feature requests.
For issues, questions, or suggestions, please open an issue on the GitHub repository.
- OpenAI CLIP for multi-modal embeddings
- Facebook FAISS for efficient similarity search
- PyQt5 for the GUI framework
- Hugging Face Transformers for pre-trained models