PixelBox

A desktop image search and indexing tool.

---

PixelBox is still pre-alpha. Database schema and feature prioritization are subject to change.

Features

• Cross-platform (Windows, Linux, MacOS) and FOSS
• Search across filenames and exif tags
• Drag and drop search for visually similar images
• Fast parallel indexing of images
• User-moddable image similarity engine (!)
• Portable and inspectable database format

Technologies

• Rust as the primary language (with egui and tract-onnx)
• SQLite as a storage medium for the image database
• Torch for training the image similarity model
• ONNX for running the similarity model

TODOs for Alpha Release

• Compress thumbnails in database [DONE - 2x Compression for No Loss in Speed]
• Remove from index on folder clear [DONE]
• Settings Page [DONE]
• Start removing those unwraps

TODOs for Roadmap

• Better similarity search
• OCR for images (search on text in images)
• Editable tags
• Face search
• Search on image contents in plaintext
• Watched directories via notify crate
• If a model is unavailable, don't perform image hash and just disable similarity search so people can use it for just tags
• Index inside of zip files

Project Structure

• .github - Links to demo pictures for readme and, eventually, CI/GitHub Action build scripts
• models - The final ONNX files to be used by the application for visual similarity
• resources - Non-shipped experiment logs and python training files
• src - The main application code
  • image_hashes - Wrappers for different image hashing methods
  • ui - Code for each of the major UI panels like search view, folder view, etc.

Using Your Own Image Hash (Advanced)

PixelBox's search uses the cosine distance between byte-quantified n-dimensional floats. For example, if you represent your image as [-1.0, 1.0, 0.0, 0.1] then this will be mapped to a 4-byte vector of [0x00, 0xFF, 0x80, 0x8C].

There are two ways to use your own image hash methods:

1. Replace the image_similarity.onnx file with your own trained model. The inputs should be channel-first 128x128 RGB images and the outputs should be a 1D vector of floats between -1 and 1. See image_hashes/efficientnet.rs for constraints.
2. Replace the 'hash' in the 'semantic_hash' table of your database. This should be an array of u8s as described above. You will not be able to drag-and-drop images for search if using this approach, but after finding a seed image you can right-click and do 'find similar'.