# PLACE: Adaptive Layout-Semantic Fusion for Semantic Image Synthesis

**Authors:** Zhengyao Lv, Yuxiang Wei, Wangmeng Zuo, Kwan-Yee K. Wong  
**arXiv:** [2403.01852](https://arxiv.org/abs/2403.01852)  
**Date:** 2024-03-04  
**Summary page:** https://www.summarizepaper.com/en/paper/2403.01852v1/

## Key points

- Authors Zhengyao Lv, Yuxiang Wei, Wangmeng Zuo, and Kwan-Yee K. Wong address challenges in synthesizing high-quality images with consistent semantics and layout
- Introduce the adaPtive LAyout-semantiC fusion modulE (PLACE) to improve semantic image synthesis using pre-trained models
- Utilize a layout control map to accurately represent layouts in the feature space
- Combine layout and semantic features in a timestep-adaptive manner for realistic image generation
- Propose Semantic Alignment (SA) loss function to enhance alignment of layouts during fine-tuning
- Introduce Layout-Free Prior Preservation (LFP) loss to maintain priors of pre-trained models using unlabeled data
- Extensive experiments show that PLACE model outperforms existing methods in visual quality, semantic consistency, and layout alignment
- Availability of source code and models on GitHub for accessibility in further research and implementation

## Summary

In their paper titled "PLACE: Adaptive Layout-Semantic Fusion for Semantic Image Synthesis," authors Zhengyao Lv, Yuxiang Wei, Wangmeng Zuo, and Kwan-Yee K. Wong address the challenges in synthesizing high-quality images with consistent semantics and layout. They introduce the adaPtive LAyout-semantiC fusion modulE (PLACE) as a solution that leverages pre-trained models to improve semantic image synthesis. The key innovation of PLACE lies in its utilization of a layout control map to accurately represent layouts in the feature space. By combining layout and semantic features in a timestep-adaptive manner, the model is able to generate images with realistic details. To further enhance the alignment of layouts during fine-tuning, Lv et al. propose the Semantic Alignment (SA) loss function. Additionally, they introduce the Layout-Free Prior Preservation (LFP) loss, which utilizes unlabeled data to maintain the priors of pre-trained models. This approach ultimately enhances visual quality and semantic consistency in synthesized images. Extensive experiments conducted by the authors demonstrate that their PLACE model outperforms existing methods in terms of visual quality, semantic consistency, and layout alignment. The availability of source code and models on GitHub provides accessibility for further research and implementation in this field. Overall, Lv et al. 's work represents a significant advancement in large-scale text-to-image models and offers a promising solution for improving semantic image synthesis through adaptive layout-semantic fusion techniques.