Artistic Style Transfer Using Generative Adversarial Networks: A Pix2Pix Implementation
Dubasi Pavan Kumar, Shuchi Juyal Bhadula, Mohammed Al‐Farouni, Neeraj Varshney, Vishal Sharma, Manish Saraswat · 2024
This investigation explores the viability of four noticeable models, specifically Pix2Pix, Neural Style Transfer (NST), Fast Neural Style Transfer (FastNST), and CycleGAN, within the space of aesthetic style exchange. The think about envelops a fastidious assessment of these models, investigating their capabilities in generating outwardly engaging and elaborately reliable pictures. Through broad tests and quantitative evaluations, Pix2Pix developed as a strong choice, accomplishing a Crest Signal-to-Noise Ratio (PSNR) of 26.78 dB, a Basic Likeness List (SSIM) of 0.832, and a Fréchet Initiation Separate (FID) of 54.21. NST, exceeding expectations in style devotion, achieved a PSNR of 29.45 dB, an SSIM of 0.905, and an FID of 72.03. FastNST, known for its real-time preparation, illustrated an adjusted execution with a PSNR of 28.12 dB, an SSIM of 0.892, and an FID of 68.54. CycleGAN, outlined for unpaired picture interpretation, accomplished a PSNR of 27.65 dB, an SSIM of 0.874, and an FID of 63.72. The results give profitable experiences into the comparative qualities and weaknesses of these models, educating their appropriateness in different artistic assignments.