# 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/1801.02753v1/

## Key points

- Synthesizing realistic images from human-drawn sketches is a complex task in computer graphics and vision.
- Existing methods rely on precise edge maps or access to a database of images.
- The authors propose using Generative Adversarial Networks (GANs) to generate lifelike images across 50 different categories.
- A key contribution is the development of a fully automatic data augmentation technique for sketches, which significantly aids in the synthesis process.
- They introduce a new building block that enhances information flow and leverages input images at multiple scales for both the generator and discriminator networks.
- Their approach produces more realistic images and achieves notably higher Inception Scores compared to state-of-the-art image translation methods.
- The GAN-based method effectively generates diverse and authentic images from sketches.

## Summary

, , , , 

In the field of computer graphics and vision, synthesizing realistic images from human-drawn sketches is a complex task. Existing methods either rely on precise edge maps or require access to a database of images. However, in this study, the authors propose an innovative approach using Generative Adversarial Networks (GANs) to generate lifelike images across 50 different categories such as motorcycles, horses, and couches. One key contribution of this work is the development of a fully automatic data augmentation technique for sketches. The authors demonstrate that this augmented data significantly aids in the synthesis process. Additionally, they introduce a new building block that enhances information flow and leverages input images at multiple scales for both the generator and discriminator networks. Compared to state-of-the-art image translation methods, their approach not only produces more realistic images but also achieves notably higher Inception Scores—a metric used to evaluate image quality. These findings highlight the effectiveness of their GAN-based method in generating diverse and authentic images from sketches. The paper titled "SketchyGAN: Towards Diverse and Realistic Sketch to Image Synthesis" by Wengling Chen and James Hays provides valuable insights into addressing the challenges associated with synthesizing realistic images from hand-drawn sketches. Their novel GAN approach, combined with data augmentation techniques and improved network architecture, demonstrates significant advancements in this domain.