Influence estimation for attributing the unfairness of GNNs to training data

In recent years, huge progress has been made in image synthesis through the improvement of state-of-the-art generative models (e.g., VAEs, autoregressive models, and GANs). Now with the help of diffusion models, it reaches a brand-new level. However, applying diffusion models on conditional image synthesis still faces limitations and requires updates, especially for coarse spatial layout which generates realistic images with the corrected objects in the desired locations.
During my master thesis, I’m going to discover the possibility to achieve layout-to-image synthesis via diffusion models, to evaluate whether it will get better results than other GAN-based layout-to-image models. To process the layout information, I will extend the object instance-specific and layout-aware feature normalization (ISLA-Norm) used in LostGAN and adjust it for diffusion models. In addition, I’m going to test the proposed model by the dataset from Bosch to generate defective product surfaces while maintaining both diversity and fidelity for possible data augmentation and further industrial purposes. Due to the limited amount of the images in the bosch dataset, I’m going to try fine-tuning and discovering more possible methods to adapt the training of diffusion models.

• Author: Xuefeng Hou, B.Sc.
• Main Examiner: Prof. Dr. Mathias Niepert
• Supervisor: Ruyu Wang, M.Sc.