Research Highlights

Image Generative Model breaking away from Denoising in Diffusion-Like Environment

Update as of 3 August 2026



Generative AI with deep learning has many useful applications, such as image super-resolution, image inpainting, image/video compression, data augmentation, virtual reality. etc. Recently, diffusion models have been considered as the most effective way to make image generation, for which Noise is a means to bring and form the required statistical distributions to generate photo-realistic images. This research proposes a new domain transfer approach to break away from the requirement of using explicit noise operations for image super-resolution; hence no noise operation is required which can speed up the realization speed for over 100 times. It is also expected that this new Domain Transfer in Latent Space (DTLS) approach not only can perform high-quality image super-resolution, but can also be used to create an efficient image generative model, for various hi-tech applications.

 Quality faces generated with our novel idea in DTLS
 

Team Members:

  1. PI: Prof. SIU Wan Chi, Yam Pak Charitable Foundation School of Computing and Information Sciences, Saint Francis University
  2. Prof. CHAN Hing Hung Anthony, Yam Pak Charitable Foundation School of Computing & Information Sciences, Saint Francis University

Publication:

  1. Hon Man Hammond Lee, Wan-Chi Siu and Ivan Wang-Hei Ho, "DTLS-Inpaint: A Novel Impainting Framework with Latent Space Domain Transfer", Accepted with Minor Modification, IEEE Transactions on Image Processing.
  2. Shaohua Jia, Pengyu Liu, Kebin Jia, Anthony H Chan and Wan-Chi Siu, "Domain Transfer in Latent Space for Face Video Super-Resolution", IEEE Signal Processing Letter, July 2026.
  3. Hon Man Hammond Lee, Wan-Chi Siu, Felix Ming-Fei Duan, Yi-Hao Cheng and H. Anthony Chan, "Intelligent Picture Painting under Deep Learning with Text Enhancement", Proceedings, pp.1-5, 25th Conf on Digital Signal Proc, Messioia, Greece, 25-27 June 2025.
  4. Hon Man Hammond Lee and Wan-Chi Siu, "DTLS-Inpaint: Yet Another Efficient Image Inpainting with Domain Transfer", Proceedings, pp.1990-1995, IEEE International Conference on Image Processing (ICIP'2025), Anchorage, Alaska, USA, 14-17 September 2025.
  5. Chun-Chuen Hui, Wan-Chi Siu and H. Anthony Chan, "Domain Transfer Generative Model for New Face Generation", Proceedings, pp.1978-1983, IEEE International Conference on Image Processing (ICIP'2025), Anchorage, Alaska, USA, 14-17 September 2025.

Reference no.: UGC/FDS11/E06/24