ICCV2019-LearningToPaint

Why we included this project

This repo is the reference implementation of an ICCV 2019 paper that treats painting as a reinforcement learning problem. Instead of copying pixels directly, an agent works with a neural renderer to decide where each brushstroke goes, what shape it takes, and what color it should be, planning across many steps until a photo is rebuilt from a few hundred strokes. It is a good place to see reinforcement learning drive continuous, structured actions rather than discrete choices, and it is also relevant if you are exploring stroke-based or non-photorealistic image reconstruction. Pretrained renderer and agent weights plus a Colab notebook let you watch the model redraw images without training anything yourself. Treat it as a research demo and learning resource, not a production library; the pinned dependencies target older PyTorch versions, so expect some setup work.