Meme Search

Meme Search is an open-source, self-hosted web application that indexes and retrieves images using semantic content search. It uses machine learning to analyze image content and text, allowing users to find specific memes based on meaning rather than just filenames. The application is deployed as a server cluster, typically via Docker, and provides a web interface for managing and searching an image library.

The software handles the full pipeline from image-to-text extraction to vector embedding and retrieval. It can be configured to perform all processing locally or use an OpenAI-compatible vision API for description generation while maintaining local embeddings and search. Users can organize images across multiple subdirectories or upload files directly through the browser, with the system automatically scanning these directories for new content to index.

Key features

• Support for multiple image-to-text models including Florence-2, SmolVLM, and Moondream2
• Semantic vector search powered by Postgres and pgvector
• Automatic and bulk generation of image descriptions
• Manual editing of descriptions and custom tag assignment
• Drag-and-drop upload support for JPG, PNG, and WEBP formats
• Directory rescanning to automatically index new files
• Filtering by tags, directory paths, and description embeddings
• Dark mode interface toggle

The application is built using Ruby on Rails, Python, and PostgreSQL. It is designed for users with homelab setups or those who prefer local AI processing, though it supports external API integrations for vision tasks. The architecture utilizes a background job worker to handle bulk indexing and description generation without blocking the web interface. For memory-constrained hardware, the software offers quantized model versions, such as Moondream2-INT8, which reduces memory requirements from 5GB to approximately 1.5-2GB for CPU-only machines.

Deployment is streamlined through Docker Compose, which manages the app, database, and local auto-description generator. The system uses bind-mounted directories to ensure that database data, model downloads, and uploaded images persist across software upgrades. This makes it suitable for installation on NAS devices or dedicated home servers where data persistence is required.

This tool is positioned as a self-hosted utility for organizing and retrieving large collections of image-based memes.