Reproduce “DSE-GAN: Dynamic Semantic Evolution Generative Adversarial Network for Text-to-Image Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Mengqi Huang, Zhendong Mao, Penghui Wang, Quan Wang, Yongdong Zhang
A new Text-to-Image framework, DSE-GAN, dynamically re-composes text features during multi-stage image generation to provide adaptive semantic guidance aligned with evolving visual content. This method yielded improved image quality and semantic consistency, with FID scores of 13.23 on CUB-200 and 15.30 on MSCOCO, outperforming prior models by up to 37.8%.
Key insights
- Dynamically adapting textual guidance to the current visual generation stage significantly improves image quality and semantic consistency in multi-stage T2I processes.
- Re-composing word features by routing both relevant words and their specific semantic subspaces provides more granular and adaptive control over the generative process.
- A simplified adversarial training setup with a single discriminator can effectively manage a multi-stage generator, enabling increased interaction stages and dynamic semantic evolution without proportional increases in training complexity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.