Semantic Layout Manipulation With High-Resolution Sparse Attention

We tackle the problem of semantic image layout manipulation, which aims to manipulate an input image by editing its semantic label map. A core problem of this task is how to transfer visual details from the input images to the new semantic layout while making the resulting image visually realistic. Recent work on learning cross-domain correspondence has shown promising results for global layout transfer with dense attention-based warping. However, this method tends to lose texture details due to the resolution limitation and the lack of smoothness constraint on correspondence. To adapt this paradigm for the layout manipulation task, we propose a high-resolution sparse attention module that effectively transfers visual details to new layouts at a resolution up to 512x512. To further improve visual quality, we introduce a novel generator architecture consisting of a semantic encoder and a two-stage decoder for coarse-to-fine synthesis. Experiments on the ADE20k and Places365 datasets demonstrate that our proposed approach achieves substantial improvements over the existing inpainting and layout manipulation methods.

1 Introduction

Semantic layout manipulation refers to the task of editing an image by modifying its semantic label map, i.e., changing the semantic layout or inserting/erasing objects as illustrated in Fig. 1. It has many practical image editing applications such as photo-editing [1], image retargeting [2], restoration [3], composition [4] and image melding [5], but is relatively under-explored due to the challenges of predicting complex, non-rigid spatial deformations and the domain gap between the input image and the target semantic layout. Essentially, developing an effective method to transfer visual patterns from the input image to the target semantic layout is the key to solving the problem.