Thing/Not Thing

This is a zero-shot binary image classifier. Type in the name of an object and AI predicts whether your uploaded photo matches it.

How it Works

This project uses OpenCLIP, an open-source implementation of OpenAI's CLIP.

This project is available as a web demo here. But it will be slower than when the project is run locally on a GPU.

You can expand "Additional Inputs" to allow adjusting the cosine similarity threshold below which your photo is deemed Not <object>.

Tested on Debian.

NVIDIA Container Toolkit
Docker (and Docker Compose)
An NVIDIA GPU with sufficient VRAM for your chosen ViT model (model size can be changed in app.py)

Create a .env file which points to the path where you downloaded the OpenCLIP model. Then run:

docker compose build
docker compose run torch

Finally, go to http://localhost:7860 in your browser.