# SketchyCOCO: Image Generation from Freehand Scene Sketches

**Authors:** Chengying Gao, Qi Liu, Qi Xu, Limin Wang, Jianzhuang Liu, Changqing Zou  
**arXiv:** [2003.02683](https://arxiv.org/abs/2003.02683)  
**Date:** 2020-03-05  
**Summary page:** https://www.summarizepaper.com/en/paper/2003.02683v5/

## Key points

- Paper introduces a method for automatic image generation from freehand sketches
- Authors propose a model called EdgeGAN for controllable image generation through sketches
- EdgeGAN is an attribute vector bridged GAN that generates high-quality object-level image content without relying on sketch training data
- Authors constructed a large-scale dataset called SketchyCOCO to support and evaluate their solution
- Extensive analysis, evaluation, and studies demonstrate the capability of the method to generate realistic scene-level images from freehand sketches
- Approach leverages freehand sketches as input to produce visually appealing and contextually accurate images
- EdgeGAN model shows promising results in generating high-quality images while providing control over synthesis process
- Potential applications in computer vision and graphic design, opening new avenues for sketch-based interfaces and generative models.

## Summary

The paper titled "SketchyCOCO: Image Generation from Freehand Scene Sketches" introduces a groundbreaking method for automatic image generation from scene-level freehand sketches. The authors propose a model that enables controllable image generation by specifying the desired synthesis goal through freehand sketches. The key contribution of this research is the development of an attribute vector bridged Generative Adversarial Network (GAN) called EdgeGAN, which facilitates high visual-quality object-level image content generation without relying on freehand sketches as training data. To support and evaluate their solution, the authors have constructed a large-scale composite dataset known as SketchyCOCO. This dataset plays a crucial role in validating their approach for both object-level and scene-level image generation tasks. Through extensive quantitative analysis, qualitative results, human evaluation, and ablation studies, the authors demonstrate the capability of their method to generate realistic complex scene-level images from various freehand sketches. Overall, this research presents an innovative approach to automatic image generation that leverages freehand sketches as input to produce visually appealing and contextually accurate images. The proposed EdgeGAN model shows promising results in generating high-quality images while providing control over the synthesis process. With its potential applications in various domains such as computer vision and graphic design, this work opens up new avenues for exploring the intersection of sketch-based interfaces and generative models.