International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 11 Issue: 11 | Nov 2024
p-ISSN: 2395-0072
www.irjet.net
State-of-the-Art Review on Image Synthesis with Generative Adversarial Networks 1Chetna L Shapur, 2Sirisha R, 3Dhanalakshmi S, 4Prof. Swati Srikanth Achanur 1,2,3 UG student, Dept. of CSE-AIML, AMC Engineering College, KARNATAKA, INDIA. 4Assistant professor, Dept. of CSE-AIML, AMC Engineering College, KARNATAKA, INDIA
-------------------------------------------------------------------------***------------------------------------------------------------------------
ABSTRACT - This paper explores advanced image-
Traditional models often face issues such as mode collapse and require significant computational power. This paper aims to overcome these challenges by integrating the Stable Diffusion Pipeline and optimizing the training process with advanced data augmentation techniques.
generation techniques with Generative Adversarial Networks (GANs). Assess our approach in comparison to current methods, and we highlight it by comparing it with current techniques in Visual quality and training stability. Our primary contributions include integrating the Stable Diffusion Pipeline and optimizing the training process with improved data augmentation methods. Experimental results showcase the efficiency of our methods in producing high-quality images across various domains.
4. ARCHITECTURE Our proposed architecture builds on the standard GAN architecture integrated with significant modifications. The generator and discriminator networks are designed to be enhanced with deeper layers and optimized activation functions. We incorporate the Stable Diffusion Pipeline to enhance the diffusion process during image generation. Moreover, data augmentation techniques are applied to the training dataset to enhance diversity and improve model robustness.
Key Words: GANs (Generative Adversarial Networks) for image synthesis, Neural network configurations, Data Enhancement, Computational learning.
1. INTRODUCTION Generative Adversarial Networks (GANs) have transformed the field of image generation by facilitating the generation of lifelike images from random noise. GANs feature two neural networks: the generator and the discriminator, which challenge one another to enhance the quality of the generated images. Despite significant advancements, including issues like mode collapse and training difficulties instability persists. This research proposes an enhanced GAN framework incorporating the Stable Diffusion Pipeline to address these issues and achieve superior image creation.
4.1.1 DESIGN The architecture of our Generative Adversarial Network (GAN) for text-to-image generation can be divided into several key components, each playing a critical role in the overall functionality and efficacy of the. The design is visually represented within the given image text.
4.1.2 Text Description The process begins with a written description input that is evaluated to derive meaningful attributes. These semantic attributes are crucial as they encapsulate the meaning and details of the text, which the generator will later use to produce related images.
2. EXISTING SYSTEM Conventional GANs, including DCGANs and Pro GANs, have established the groundwork for image generation. DCGANs use convolutional layers to create images while Pro GANs progressively grow the generator and discriminator networks to generate high-resolution images. Nonetheless, these models frequently encounter difficulties with stability and require extensive computational resources. Our approach leverages recent advancements within the framework GANs and incorporates the Stable Diffusion Pipeline to improve performance and stability.
4.1.3 Generator The generator network is responsible for converting the meaningful attributes gathered from the textual description into premium images. The transposed convolutions help in up-sampling the input attributes in order to generate an enlarged image, while batch normalization ensures stable and efficient training by normalizing the inputs to each layer.
3. PROBLEM STATEMENT The primary challenge in image generation using GANs is achieving high-quality images with stable training.
© 2024, IRJET
|
Impact Factor value: 8.315
|
ISO 9001:2008 Certified Journal
|
Page 691