Lj Miranda writes about machine learning, Python tooling, and Kubernetes, with explainers on VQGAN and generative art projects.
HN commenters link his VQGAN explainer for a technical account of how VQGAN and CLIP generate images, and as a companion to related discussions.
Due to how VQGAN + CLIP works, it can't memorize its inputs in the way language models like GPT-3 do.
VQGAN does the generation work, CLIP just says if it's good not, improve the latents, repeat. Here's a good technical writeup: https://ljvmiranda921.github.io/notebook/2021/08/08/clip-vqgan/
And of course, the most-common VQGAN was trained on ImageNet, which likely doesn't have every movie poster as training data. (it could be in CLIP though)
If you have a notebook / sketchbook / journal and it needs to be held open with an object look for another notebook [1]. It seems trivial but that is a poorly designed notebook.
1. https://ljvmiranda921.github.io/assets/png/pen-and-paper/blank.jpg
I also wrote an explainer for VQGAN here: https://ljvmiranda921.github.io/notebook/2021/08/08/clip-vqgan/ to coincide with the growing interest in these models