Ganic Vision: GAN-based Image Reconstruction and Attribute Manipulation

Deep learning methods have been used more and more in the past few years to solve problems with image creation, modification, and reconstruction but current models frequently fail to recover and maintain visual structure, particularly when large chunks of data are absent. This research proposes a revolutionary two-stage image reconstruction and manipulation method. The initial step, which draws inspiration from sketch drawings, entails creating edge maps in order to forecast the missing image structure. The in-painting phase that completes the image is followed by these forecasted edge maps. This study also discusses about the challenges of image manipulation using descriptions in natural language. Semantically altering an object's visual characteristics in an image in accordance with the given textual descriptions without altering the structure itself in the second step. Existing techniques frequently fail to retain the original image's text-inconsequential information. This research study presents a text-driven discriminator that uses input text to generate high-level local discriminators in order to address this issue. This makes it possible to classify fine-grained attributes separately, enabling the generator to create visuals in which only the areas that correspond to the supplied text are altered. The outcomes of the experiment show that the two-stage model bridges the technological gap by reconstructing the image and modifying it in accordance with user preferences.