.png){kind=icon}
.png){kind=icon}
Return
Exploring text-to-image generation models: Applications and cloud resource utilization
.png){kind=icon}
PRE
AI
DOI：10.1016/j.compeleceng.2025.110194.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Generating images from text presents a significant challenge in computer vision. Moreover, manually acquiring images from multiple perspectives for object or product generation is a resource-intensive and expensive endeavor. However, recent breakthroughs in deep learning and artificial intelligence have opened doors to creating new images from diverse data sources, and cloud resources play a pivotal role in alleviating the resource-intensive nature of this endeavor. As a result, substantial research efforts have been directed toward advancing image generation techniques, yielding impressive results. This paper aims to provide a comprehensive overview of existing image generation methods, offering insights into this evolving field of text-to-image generation. It traces the historical development of this technology. It examines the key models that have shaped its evolution, including Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), Conditional GANs (CGANs), StackGAN, Transformers, and diffusion models. The paper offers insights into the functioning of text-to-image generation within the GAN architecture, elucidating the mechanisms behind transforming textual descriptions into visual content. Additionally, the integration of text-to-image generation with cloud and edge-cloud computing highlights the synergistic potential of these technologies while addressing the challenges and considerations associated with cloud infrastructure. The paper concludes by surveying the diverse applications of text-to-image generation across various domains, such as art, ecommerce, entertainment, and education. It also discusses the evaluation metrics commonly used in assessing the quality of generated images and the challenges that exist both within the methods and in their application across different domains. This review offers a comprehensive overview of the capabilities and limitations of text-to-image generation. Also, we have introduced a new HiResGAN model using a single generator/discriminator pair of networks to produce high-resolution 256 x 256 images from textual descriptions. We illustrate the efficacy of our model in producing high-resolution images based on contextually rich text descriptions that are visually plausible and semantically consistent through a series of experiments and analyses.
Keywords:
VAE
GAN
StackGAN
Diffusion model
Cloud resouce utilization
Edge-cloud computing
Text-to-image generation
Journal
C
IF:
4.9
Papers:
6.7K
Citations:
1.3W