min-dalle

Why we included this project

You give it a plain-English prompt and it returns a grid of images, using the DALL·E Mini and Mega weights rebuilt in PyTorch and trimmed down to the inference path. The dependency list is short (numpy, requests, pillow, torch), so a pip install and a free Colab GPU are enough to get running without dragging in a heavyweight framework. You load the model once and reuse it for many prompts, with sampling controls like temperature and top-k plus a supercondition factor that balances how closely the output tracks your text against how varied the images come out. Progressive decoding output suits interactive notebooks, and a command-line path handles batch generation. For developers who want a compact, readable text-to-image model they can run and adapt locally, this is a straightforward place to start.