By Nan Zhao ([email protected])
Supervisor: Rebecca Fiebrink
MSc Advanced Project
Creative Computing Institute, University of the Arts London, London, United Kingdom
Abstract: The murals in Dunhuang Mogao Grottoes have been comprising the great treasure house of ancient Chinese art and religious heritages. However, the murals are under a severe situation of environmental damage and withering away. This paper proposes a set of generative approaches to automatically generate the Dunhuang mural styled images and translate human faces into Dunhuang artistic portraits. The generation is optimized by styleGAN, while the translation applied Unsupervised Generative Attentional Networks with Adaptive Layer Instance Normalization for Image-to-image Translation (U-GAT-IT). To better evaluate the proposed approaches, a large set of both Dunhuang murals and portraits has been collected and processed. Additionally, the paper discovered the data bias on the original dataset of U-GAT-IT. This paper further fed the network with an Asian face dataset. This paper argues that such a portrait translation could better represent the characteristic of Asian ancient art. The project aims to preserve the essence of Dunhuang mural arts and portrait styles, at the same time calling for public attention on the Dunhuang Arts. The experimental results prove that the system can generate satisfied Chinese Dunhuang mural styled images and portraits.
Keywords: Dunhuang Mogao Grottoes arts, Automatic image generation, Portrait generation, styleGAN2-ADA, unconditional image-to-image translation
- The networks and checkpoints: Drive Link
- The documentation blog for training the networks: Link
- The web blog recording the whole research process: Link
- The jypter notebooks implemented for data preprocessings: Link
- The interpolation videos can be seen here and some generated images can be seen here.
This project collects and processes two Dunhuang datasets. One is the dataset of Dunhuang mural paintings. The other is the dataset of Dunhuang portraits, with the faces cropped, aligned, and resized. For the sake of the data protection, the datasets are now only accessible under requests and permissions.
Last Update: 19 November 2021