Core Concepts
SketchGPT is a flexible autoregressive transformer model that can generate, complete, and recognize sketches by learning neural representations of sketches and their sequential drawing patterns.
Abstract
The paper presents SketchGPT, an autoregressive transformer model inspired by the GPT architecture, for versatile sketch-related tasks. The key contributions are:
SketchGPT employs a sequence-to-sequence autoregressive model to learn neural representations of sketches, capturing their dynamic drawing process. This allows the model to perform tasks like sketch generation, completion, and recognition.
The authors propose a stroke-to-primitive abstraction strategy to simplify the input data and enhance model generalization across diverse sketches. This discretization of sketches into a finite set of abstract primitives streamlines the learning process and reduces overfitting.
SketchGPT is a multi-task model capable of predicting the next stroke, generating, completing, and recognizing sketches, showcasing its overall versatility in sketch-related applications.
The paper provides a quantitative study for sketch generation, comparing SketchGPT with state-of-the-art models, and a comprehensive human evaluation study to assess the quality of generated sketches.
The experiments demonstrate SketchGPT's strong performance in sketch generation, completion, and recognition tasks, outperforming or matching existing approaches. The model's ability to adapt to various sketch-related applications highlights its potential as a versatile framework for understanding and generating sketches.
Stats
The model was evaluated on the QuickDraw dataset, which consists of over 50 million hand-drawn sketches across 345 different categories.
Quotes
"SketchGPT leverages the next token prediction objective strategy to understand sketch patterns, facilitating the creation and completion of drawings and also categorizing them accurately."
"Our findings exhibit SketchGPT's capability to generate a diverse variety of drawings by adding both qualitative and quantitative comparisons with existing state-of-the-art, along with a comprehensive human evaluation study."