clip-search

(WIP) A simple tool for searching images inside a local folder with text/image input using CLIP

git 10 results for “a blonde woman” in a folder of 30k randomly generated faces

Installation

Install with GPU support if you want to use GPU
Initially encoding images might run faster with GPU, but from limited testing it seems like searching with a saved dictionary might be faster on cpu.

git clone https://github.com/kanttouchthis/clip-search.git
cd clip-search
git clone https://github.com/openai/CLIP.git
pip install -r CLIP/requirements.txt

Usage

Basic Usage:

#Any amount/combination of text/image prompts is possible
#Copy 5 results for the text search "a car" to the results folder (and initiate dictionary if it doesn't exist)
python clip_search.py -f path/to/image/folder -r 5 -c -cr -t "a car"

#Return 10 results for image.png and "a car" each using dictionary dict.pt, without copying the images (only return names)
python clip_search.py -f path/to/image/folder -r 10 -d dict.pt -t "a car" -i image.png


usage: clip_search.py [-h] [-t TEXTS [TEXTS ...]] [-i IMAGES [IMAGES ...]] [-r RESULTS] [-se SAVE_EVERY] [-f FOLDER] [-d DICT] [-de DEVICE] [-fo FORMAT] [-cf COPY_FOLDER] [-in] [-c] [-cr]

(WIP) A simple tool for searching images inside a local folder with text/image input using CLIP

optional arguments:
  -h, --help            show this help message and exit
  -t TEXTS [TEXTS ...], --texts TEXTS [TEXTS ...]
                        Texts to search for: "a house" "an old man"
  -i IMAGES [IMAGES ...], --images IMAGES [IMAGES ...]
                        Images to search for: "/path/to/img1.jpg" "path/to/img2.jpg"
  -r RESULTS, --results RESULTS
                        Number of search results to return. Default: 5
  -se SAVE_EVERY, --save_every SAVE_EVERY
                        Dictionary save frequency. Default: 1000
  -f FOLDER, --folder FOLDER
                        Folder to scan. Default: images
  -d DICT, --dict DICT  Stored dictionary file
  -de DEVICE, --device DEVICE
                        Device to use ("cuda" or "cpu"). Default: cuda if available
  -fo FORMAT, --format FORMAT
                        Text search formatting. Default="a picture of {text prompt}"
  -cf COPY_FOLDER, --copy_folder COPY_FOLDER
                        Results folder. Default: results
  -in, --initiate       Initiate new dictionary (overwrite)
  -c, --copy            Copy images to results folder
  -cr, --copy_remove    Remove old results from folder

GitHub

View Github