hzwer/ICCV2019-LearningToPaint

Uses a differentiable neural renderer as the painting environment paired with a policy network agent trained via model-based DRL, enabling the system to learn stroke sequencing and parameter selection (position, color, shape) without paired stroke annotations. Supports multiple rendering backends (round brushes, triangles, Bezier curves) and integrates PyTorch with TensorBoard for training visualization, pre-trained on CelebA with inference-ready models available.

2,297 stars. No commits in the last 6 months.