Colour-based Image Generation using CGAN

Generating photographic images conditioned on text descriptions, is a challenging task. Many works have been carried out to address problems related to such generative models. When results are analysed, each approach has its own achievements, drawbacks and areas need improvement. One observation is that most of the recent projects have used Generative Adversarial Network (GAN) as the basic architecture and have gained plausible results. In this paper, the results of experiments carried out using colour as the main attribute for generating low-resolution 64 × 64 images of birds are presented. Another specialty is that, a lightweight model is used for generating images that reduces the training time and the size of the model file, which can be considered as an advantage when compared with existing work having more complex models resulting longer training time and larger model files. Moreover, the presented model can also generate realistic new images, which are not available in the dataset by creating images of same bird in different colours and with different backgrounds. Therefore, attempted to create new designs or styles from the available dataset.