Research — Math + Code + Art

RenderingOptimizationGenerative Models

arXiv preprint · 2025

Birth of a Painting

Differentiable Brushstroke Reconstruction

Ying Jiang, Jiayin Lu, Yunuo Chen, Yumeng He, Kui Wu, Yin Yang, Chenfanfu Jiang

Tackles the challenge of making a computer paint — not just generate an image, but build it up stroke by stroke, as a human would. The method optimizes Bézier strokes using a differentiable paint renderer, incorporates a texture stage driven by a StyleGAN prior, and applies a differentiable smudge renderer that blends neighboring strokes naturally. The result faithfully reproduces the look and feel of oil, watercolor, ink, and digital painting.

Project page → arXiv:2511.13191

GeometryMesh ReconstructionVoronoi3D Printing

arXiv preprint · 2026

VoroLight

Learning Voronoi Surface Meshes via Sphere Intersection

Jiayin Lu, Ying Jiang, Yumeng He, Yin Yang, Chenfanfu Jiang

Asks what it takes to learn a smooth Voronoi surface from data. Standard differentiable Voronoi methods tend to produce locally jagged geometry. VoroLight addresses this by associating a trainable sphere with each surface vertex and introducing a sphere-intersection loss that encourages smooth, higher-order configurations. The result reconstructs clean, watertight meshes from point clouds, images, or implicit fields, and extends naturally to volumetric meshes ready for 3D printing.

Project page → arXiv:2512.12984

// connection to tutorials

Voronoi geometry runs through multiple tutorials — from Voronoi Photo Mosaic to Style Transfer Voronoi. Stroke-based rendering connects to the Music Painting and Neural Music Visualizer tutorials at a conceptual level. Together, these works show the depth reachable from the same interplay of mathematics, code, and art.

Browse tutorials →