An effective architecture for Image colorization using Adaptive GAN

More attention has recently been paid to image colorization. In order to produce realistic results, traditional reference-based methods frequently rely on external colour pictures. An extensive picture database or internet search engine will be required to obtain such instances. At a reasonable cost, recent deep-learning-based algorithms may colorize photos automatically. On the other hand, inconsistent hues and poor artifacts are always present. Utilizing the extensive and diverse colour priors stored within a pre-trained Adaptive Generative Adversarial Network (GAN), our goal in this study is to restore vibrant colors. In particular, we incorporate matched features comparable to exemplars into the recoloring procedure by using feature modulations and a GAN encoder to “extract” them. Our system could generate vivid colors with just one forward pass thanks to the dynamic generative colour prior and sensitive designs. Additionally, changing the GAN hidden codes makes it extremely simple to obtain a variety of outcomes. By traversing GAN latent space, our method enabled controlled and smooth transitions while maintaining the advantages of Adaptive GAN interpretable controls. Our approach outperforms previous efforts in terms of execution, as evidenced by extensive testing and client research.