canvon / CLI-P

CLI-P is commandline driven semantic image search using OpenAI's CLIP

CLI-P: Semantic commandline image search using CLIP

CLI-P is an easy, commandline interface driven way for the lazy photographer to look for something in that big old image archive. I quickly hacked it together, but it seems to be working alright.

It is a set of two python scripts to use CLIP to search through a library of image files from a simple commandline. It supports both search by text embedding or image similarity, similar to same.energy.

CLI-P uses LMDB as for storing filenames and 512D feature vectors and faiss to build a fast similarity search index, so it should scale reasonably well to even high numbers of images.

Face detection is being done with Retinaface, while face embeddings are calculated with a pretrained ArcFace model.

Make sure you have virtualenv and swig installed.

To set things up, follow (or run) setup.sh. You may want to look at the first few lines of build-index.py and adjust the variables there to your liking. If you run it, you should get a message All done. at the end. If you get an error about running out of pages for the LMDB, rename cli-p.conf.example to cli-p.conf and customize it.

Indexing images in a folder

Make sure you have activated the virtual environment in your current shell using:

. env/bin/activate

Run python build-index.py with a list of folders as arguments. It will not recurse into subfolder. Only files ending with .jpg, .jpeg or .png will be indexed, but you could add other extensions in the code.

Searching for images

Make sure you have activated the virtual environment in your current shell using:

. env/bin/activate

Run python query-index.py and you will get a prompt where you can enter search terms and a few commands. Enter h for a short help.

For a nicer image viewing experience, try:

a
r 1600x900

Known issues

• Running query-index.py while a new index is being generated will give bad results.
• Running multiple instances of build-index.py in parallel may mess things up, but running a single instance of it afterwards should fix them back up.

Testing

GUI:

There's a (currently very incomplete, but) emerging test suite which you can call via: python -m unittest discover tests

That is, invoke the Python unittest module's discover sub-command on the tests directory.

For more details, example commands and example output, see the test-suite's README.

Credits

By ps-auxw. OO/GUI work by canvon.