iptc-with-llava

Short Python scipt to create IPTC keywords for any images with Ollama and LlaVA

This Python script sequentially reads images from an input folder, sends them to a Llava 1.6 model within an Ollama instance using a special prompt and saves the resulting keywords in the IPTC block of the new image in the target folder.

Quick Install

You can clone this repository with

git clone https://github.com/rawar/iptc-with-llava.git

Then you need to create a virtual Python environment with python3.11 -m venv .venv and source .venv/bin/activate

After all, you need to install the dependencies with

pip install -r requirements.txt

1st Run

You can use the existing quellbilder and zielbilder folder for your source images and the images with the generated IPTC meta data. You can use a local Ollama service or a Ollama instance somewhere in your local network. If you would like to use a local network Ollama service, please use the base_url parameter inside Lanchains Ollama constructor. Then you can put your images into quellbilder and run python image-to-iptc.py.

Example

The following prompt generates keywords from the following image:

Please find very precise keywords and separate them with commas.

LlaVA 1.6 generates the following keywords: