Strong deep generative models

Deep generative models (e.g., GAN, diffusion model) show outstanding performance in various fields such as image generation, image editing, and super-resolution. However, generative models yield poor generative performance for complex and large structured datasets (e.g., ImageNet). Most of the generative models learn only image level and have limited representation power. Language level can clearly express ambiguity by image level (e.g., a man with long hair). Therefore, we believe that generative models can learn abundant representation power by leveraging language level. In this project, we develop a powerful generative model to learn abundant representation and build the baseline for comparing the performance of various state-of-the-art generative models.

This project is suitable for (but not limited to) students majoring in electrical engineering, computer science, industrial engineering, and mathematics. Ideally, students in their third year or higher are preferred. Basic English skills for reading and presenting papers and proficiency in Python programming are required. Otherwise, it will be very difficult to proceed with the project. Students who have experience with deep learning projects or have read papers on the topic are preferred. If they have experience working with generative models, it would be even better.

Supervisors

(If you are interested, please contact the supervisor below)


Feel like you want to explore other projects? Go and look for the list here: Student (Intern) projects