Parham Rezaei

Parham Rezaei

I am an undergraduate in Computer Engineering at Sharif University of Technology.

Last summer, I had the previlege to work under the supervision of Prof. Farnia at CUHK. We explored Multi-Armed bandit approaches to online learning of the optimal mixture of models.

Before that, I collaborated with Prof. Soleymani and Prof. Rohban at Sharif, working on compositional generation and methods to enhance spatial relationship alignment in text-to-image diffusion models.

Two summers ago, I interned at the L3S Research Center, where we explored ways to improve the CLIP score for uncovering the reasoning behind image classification.

Currently, I am working under the supervision of Prof. Mondelli and Prof. Locatello at ISTA on weak-to-strong generalization in generative models.

Email  /  GitHub  /  Google Scholar  /  LinkedIn

profile photo

Research

I am interested in generative models. Mostly, theoretical perspective on how the generative models behave in various tasks. What I describe as a cool research is using a analytical view to figure out ways to improve the generative models in practical scenarios.

project image

Be More Diverse than the Most Diverse: Online Selection of Diverse Mixtures of Generative Models


Parham Rezaei, Farzan Farnia, Cheuk Ting Li
ICLR, 2025
arxiv / code / website /

We proposed an online algorithm for finding the optimal mixture of generative models.

project image

Diffusion Beats Autoregressive: An Evaluation of Compositional Generation in Text-to-Image Models


Arash Mari Oriyad, Parham Rezaei, M.S. Baghshah, M.H. Rohban
NeurIPS 2024 Workshop on Compositional Learning, 2024
arxiv /

We investigated the performance of diffusion models in text-to-image generation tasks. We found that diffusion models outperform autoregressive models in terms of compositional generation.




Research Experiences

I have been fortunate to work with some amazing researchers and professors.

ISTA
Theoretical foundations of weak-to-strong generalization in generative models • Feb. 2025 to Now
Advisor: Prof. Marco Mondelli and Prof. Francesco Locatello
CUHK
Optimizing mixture of generative models using mixture-ucb algorithms • Jul. 2024 to Oct. 2024
Advisor: Prof. Farzan Farnia
Sharif ML Lab
Compositional generation and spatial accuracy in T2I models • Dec. 2023 to Jul. 2024
Advisor: Prof. Mahdieh Soleymani, Prof. Mohammad Hossein Rohban
L3S Research Center
Improving rationales behind CLIP • Jul. 2023 to Oct. 2023
Advisor: Prof. Wolfgang Nejdl

Teaching

These include both teaching and non-paid TAing experiences.

TA

Deep Learning [Graduate]: Spring 2025, Sharif, with instructor: Mahdieh Soleymani

System 2 [Graduate]: Spring 2025, Sharif, with instructor: Mahdieh Soleymani

Advanced Programming: Spring 2024, Spring 2023, Sharif, with instructor: Mohammad Amin Fazli

Artificial Intelligence: Spring 2024, Fall 2023, Sharif, with instructor: Mohammad Hossein Rohban

Computer Simulation: Fall 2024 (Head), Spring 2024, Fall 2023, Sharif, with instructor: Bardia Safaei

Design of Algorithms: Spring 2024, Sharif, with instructor: Mohammad Ali Abam

Database Design: Spring 2024, Sharif, with instructor: Maryam Ramezani

Discrete Structures: Spring 2024, Sharif, with instructor: Hamid Zarrabi-Zadeh

Fundamentals of Programming: Fall 2023, Fall 2022, Sharif, with instructor: Mohammad Amin Fazli

Linear Algebra: Spring 2023, Fall 2023, Sharif, with instructor: Maryam Ramezani

Mathematics Olympiad Instructor

Allameh Tabatabei High School: Sep. 2022 to Jul. 2023, Teaching Algebra

Allameh Helli High School: Sep. 2021 to Jul. 2022, Teaching Combinatorics and Geometry

Other Projects

These include coursework, side projects and unpublished research work. (To be updated)

project image

A 3D Tennis Complete Analysis


Sharif 3D Computer Vision
2023-10-05
paper / code /

Used both classical computer vision techniques and deep learning models to automate the process of analyzing tennis matches.

Miscellaneous

I enjoy reading and playing tennis in my free time. I also love traveling and generally new adventures.


Design and source code from Jon Barron's website