Stacked deep fusion GAN for enhanced text-to-image generation

Chen, Wenli, Sun, Yaqi, Rosin, Paul L. ORCID: https://orcid.org/0000-0002-4965-3884 and Lai, YuKun ORCID: https://orcid.org/0000-0002-2094-5680 2025. Stacked deep fusion GAN for enhanced text-to-image generation. Visual Computer 10.1007/s00371-025-03908-7 Item availability restricted.

PDF - Accepted Post-Print Version Restricted to Repository staff only until 5 May 2026 due to copyright restrictions. Download (10MB)

Abstract

Generating high-quality, semantically consistent images from text descriptions remains a challenging task in computer vision. Current methods often struggle with effectively integrating textual information into the image generation process, resulting in images that lack realism or contain significant artifacts. To address these issues, we propose SDeep, a novel framework utilizing a generative adversarial network (GAN) architecture with a channel attention mechanism. SDeep deepens the text-to-image fusion process through stacked deepening blocks (SD blocks) and enhances image detail through multilayer channel attention (MLCA). Extensive experiments on the CUB and COCO datasets demonstrate that SDeep outperforms state-of-the-art methods in terms of image quality and semantic alignment with text descriptions. Our approach not only generates more realistic images but also better preserves the semantic consistency between text and generated images, marking a significant advancement in text-to-image synthesis. Code can be found at https://github.com/zxcnmmmmm/SDeep.

Item Type:	Article
Date Type:	Published Online
Status:	In Press
Schools:	Schools > Computer Science & Informatics
Publisher:	Springer
ISSN:	0178-2789
Date of First Compliant Deposit:	4 June 2025
Date of Acceptance:	28 March 2025
Last Modified:	05 Jun 2025 09:15
URI:	https://orca.cardiff.ac.uk/id/eprint/178780

Actions (repository staff only)

Edit Item