# SketchyGAN: Towards Diverse and Realistic Sketch to Image Synthesis

**Authors:** Wengling Chen, James Hays  
**arXiv:** [1801.02753](https://arxiv.org/abs/1801.02753)  
**Date:** 2018-01-09  
**Summary page:** https://www.summarizepaper.com/en/paper/1801.02753v2/

## Key points

- Synthesizing realistic images from human-drawn sketches is a challenging task in computer graphics and vision.
- Existing approaches require exact edge maps or rely on retrieving existing photographs.
- "SketchyGAN: Towards Diverse and Realistic Sketch to Image Synthesis" proposes a novel approach using Generative Adversarial Networks (GANs).
- The proposed GAN model generates plausible images from 50 different categories, including motorcycles, horses, and couches.
- A fully automatic data augmentation technique for sketches significantly improves the performance of the GAN model.
- A new network building block enhances information flow by injecting the input image at multiple scales.
- Compared to state-of-the-art image translation methods, the authors' approach generates more realistic images and achieves significantly higher Inception Scores.
- The study presents an innovative solution leveraging GANs, data augmentation, and network building blocks for synthesizing realistic images from human-drawn sketches.
- Chen and Hays demonstrate significant improvements over existing methods in terms of image realism and quality.
- The research has important implications for various applications in computer graphics and vision fields.

## Summary

, , , , 

In the field of computer graphics and vision, synthesizing realistic images from human-drawn sketches has always been a challenging task. Existing approaches either require exact edge maps or rely on retrieving existing photographs. However, in this study titled "SketchyGAN: Towards Diverse and Realistic Sketch to Image Synthesis," authors Wengling Chen and James Hays propose a novel approach using Generative Adversarial Networks (GANs) to generate plausible images from 50 different categories, including motorcycles, horses, and couches. One key contribution of this work is the demonstration of a fully automatic data augmentation technique for sketches. The authors show that augmenting the data in this way significantly improves the performance of their proposed GAN model. Additionally, they introduce a new network building block that can be used in both the generator and discriminator components of the GAN. This building block enhances information flow by injecting the input image at multiple scales. Compared to state-of-the-art image translation methods, the authors' approach generates more realistic images and achieves significantly higher Inception Scores. The Inception Score is a widely used metric for evaluating the quality and diversity of generated images. Overall, this study presents an innovative solution to the problem of synthesizing realistic images from human-drawn sketches by leveraging GANs and introducing novel techniques such as data augmentation and network building blocks. Chen and Hays demonstrate significant improvements over existing methods in terms of image realism and quality. This research has important implications for various applications in computer graphics and vision fields.