Sequential Sketch Stroke Generation

In this project, we designed and implemented a sequence prediction model for generating handwritten strokes, focusing on a stroke-by-stroke generation process that mimics how users naturally draw. The goal was to create a conditional diffusion model capable of generating sequential strokes to form complete sketches. The project utilized the Quick, Draw! dataset, which contains user-captured sketch strokes with timestamps across 345 different object categories.

To achieve this, we explored various perspectives on data representation, including both image-based and sequential approaches. The model was trained on three specific categories: cat, garden, and helicopter, with a separate diffusion model developed for each category.

Datasets

Poster