Abstract:
The field of image-to-image translation faces several challenges, including issues
related to model convergence speed and generated image quality. In response to these
challenges, this project focuses on refining existing methodologies, specifically
targeting the Pix2Pix and CycleGAN models. By substituting instance normalization
with layer normalization in these architectures, we successfully overcome common
hurdles, resulting in significantly improved convergence speeds and enhanced image
quality. Through comprehensive experimentation, our adapted models demonstrate
superior performance in generating high-quality images. A thorough comparison with
InstructPix2Pix, a leading image translation model, further validates the effectiveness
of our approach, positioning our modified architectures as competitive solutions in the
realm of image-to-image translation. This project not only addresses key problems in
the field but also contributes valuable insights for advancing the state of the art in
image translation techniques.
