Computer Vision · Generative AI · Visual Computing
Bardiya
Kariminia
I am a Computer Engineering student and a researcher at Shahid Beheshti-University, graduating in January 2027
Currently Seeking PhD-direct & Masters opportunities
Research Interests
My research broadly centres on building intelligent visual systems that understand, generate, and reason about visual content. I am particularly drawn to generative AI exploring diffusion models and diffusion transformers for image and video editing, alongside 3D Gaussian Splatting and neural rendering for reconstructing and generating immersive 3D scenes and asseets. Beyond generation, I am fascinated by how vision-language models reason about the world, and I work on detecting and understanding their failure patterns.
- Generative AI & Diffusion Models
- Image and Video Editing
- Neural Rendering (NeRF, Gaussian Splatting)
- Vision-Language Models & Multimodal Reasoning
- Optimisation & Diversity-Aware Design
Education
Shahid Beheshti University
GPA: 18.78/20.0 (3.90/4.0 on U.S Scale)
- Machine Learning
- Artificial Intelligence
- Linear Algebra
- Discrete Math and Statistics
- Computer Architecture
- Operating Systems
- Advanced Programming
Experience
- Developed SIA, a framework for Image Analogy in image editing upon FLUX.2 Klein diffusion transformer, introducing Selective LoRA and a Constraint Support Vector module to suppress unintended edits.
- Proposed Multi-View Image Analogy to transfer camera information under uncalibrated camera settings.
- Developed GAIA (Geometry-Aware Image Analogy) by modifying RoPE positional encoding and offset computation with dense feature matching.
- Created a GAIA dataset with over 100k samples for the proposed multi-view problem.
- Worked on AI-aided engineering design using diffusion models.
- Developed a training technique injecting diversity through Determinantal Point Processes during the forward phase for 2D beams and wind-turbine airfoils.
- Optimized Pareto-front performance using Bayesian Optimization and NSGA-II.
- Techniques: DDIM, DDPM, Bayesian Optimization, multi-objective DPP, adversarial training.
- Investigated the consistency space of VLMs across image and text modalities, analyzing logical polarity for black-box hallucination patterns.
- Designed a framework to detect and learn model-specific hallucination patterns from logical polarity and features.
- Techniques: LVLMs, hallucination detection, pattern matching.
- Developed a deep generative framework for facial age transformation to improve kinship verification.
- Built a system for aging and de-aging faces while maintaining identity, ethnicity, and visual realism.
- Reported kinship-verification improvements ranging from 0.39% to 5.22%.
- Techniques: GANs, feature extraction, data balancing, cyclic domain learning.
Selected Projects
A smaller set of research projects and engineering implementations I had.
Diff-MOBO
A diversity-aware diffusion framework for engineering design, combining DPP-based diversity injection with Bayesian and evolutionary optimization.
Race-Bias-Free Face Aging
Official implementation of RA-GAN, a race-unbiased face aging model for reliable kinship verification. Developed during the SBU Vision Lab internship.
Image Denoising
A neural restoration framework for image denoising using deep learning techniques. Implements state-of-the-art architectures for noise reduction.
DDPM from scratch
A complete from-scratch implementation of Denoising Diffusion Probabilistic Models (DDPM) in PyTorch. Includes training and sampling pipelines.
Vision Transformer (ViT)
PyTorch reimplementation of the Vision Transformer paper "An Image is Worth 16x16 Words" with modifications for efficient image classification.