Differentiable Vector Graphics Rasterization for Editing and Learning

摘要

We introduce a differentiable rasterizer that bridges the vector graphics andraster image domains, enabling powerful raster-based loss functions, optimizationprocedures, and machine learning techniques to edit and generatevector content.We observe that vector graphics rasterization is differentiableafter pixel prefiltering. Our differentiable rasterizer offers two prefiltering options:an analytical prefiltering technique and a multisampling anti-aliasingtechnique. The analytical variant is faster but can suffer from artifacts suchas conflation. The multisampling variant is still efficient, and can renderhigh-quality images while computing unbiased gradients for each pixel withrespect to curve parameters.We demonstrate that our rasterizer enables new applications, includinga vector graphics editor guided by image metrics, a painterly renderingalgorithm that fits vector primitives to an image by minimizing a deepperceptual loss function, new vector graphics editing algorithms that exploitwell-known image processing methods such as seam carving, and deepgenerative models that generate vector content from raster-only supervisionunder a VAE or GAN training objective.