Using GANs for Indian Art Generation

Report : Generation of Indian Arts using Generative Adversarial Networks (GANs)

Cultural and geographic differences have led to the creation of a plethora of different art forms, and it was clear to us that the AI models should be trained to create unique and culturally specific art. This project is an effort in this direction as we investigate the capabilities of several GANs for generating Indian Art.

• We evaluated and compared the performance of DCGAN, LSGAN, GanGogh and CycleGAN on a curated dataset of Indian Art comprising 8 classes.

• Outputs generated by the LSGAN model are shown above. The paintings are beginning to take shape. Tanjore, Gond, and Kalighat especially are better defined because of the prevalence of borders in their style.

• The Structural Similarity Index Measure (SSIM) for all classes of Arts for the chosen GANs is shown in the results (Table II) of the report

• Out of the 8 classes considered, all the GANs were best at generating Tanjore art. This could be attributed to the distinct structure of the art itself, allowing it to be captured easily by the models. This was proved quantitatively as the SSIM metric was the largest for Tanjore art. Given the constraints, CycleGAN performs the best as it does not learn the features of the training image set from scratch.