Abstract:
The thesis introduces SiaGANClip, an innovative method for creating photorealistic images through unsupervised learning. The model combines a Generative Adversarial Network (GAN) with an additional discriminator based on the Siamese Network to resolve mode collapse problem, where the generator produces a limited variety of outputs despite being trained to generate diverse samples. By adding an extra discriminator and utilizing the Siamese Network’s capacity to assess the similarity between data pairs, SiaGANClip guarantees the production of varied and high-quality images. The study highlights the model’s success in generating realistic images from text prompts, significantly enhancing diversity and image quality. The proposed approach is evaluated on the Oxford 102 Flower dataset, showing promising results and potential for various applications in AI-generated imagery.
