Tuna Meral’s  interest in generative artificial intelligence began with the first machine learning course he took while earning a bachelor’s degree in computer engineering at Boğaziçi University in Istanbul, Turkey.

He later gained practical experience by working full-time as a machine learning engineer while completing a master’s degree in computer engineering at the same university.

“This helped me develop an engineering mindset alongside research curiosity,” said Meral, who applied his studies to real-world systems and co-founded an autonomous driving team at his alma mater. 

“My interest in generative vision models grew and when I had the opportunity to join Professor Pinar Yanardag’s lab to research more deeply, I seized it,” he said. “Her research aligns closely with mine and I was also excited by Virginia Tech’s strong engineering reputation and the collaborative, interdisciplinary research culture at the Sanghani Center.”

Meral’s research is particularly focused on diffusion and transformer-based models for image and video generation. He works on controllability and mechanistic interpretability, aiming to make these models behave more reliably when users specify detailed instructions, and to understand why they fail when they do. 

To achieve this, he analyzes internal representations such as attention patterns and feature activations to see how concepts and relationships are formed inside the model. 

“Based on these insights,” he said, “I develop lightweight ‘steering’ methods that act like control knobs at inference time, often improving reliability without requiring expensive retraining. My broader goal is to help turn impressive generative demos into dependable systems that can be used in real creative and engineering workflows.”

For example, Meral identified a common failure pattern in text-to-image models: they consistently struggled when asked to generate multiple objects simultaneously. Observing this limitation across both open-source and closed-source model and inspecting the internal attention mechanisms of open-source models, he developed a solution to fix this issue without requiring any further training. He implemented this on open-source models and subsequently collaborated with Google to apply his method to its closed-source architecture. 

This project was a success and his paper, “CONFORM: Contrast is All You Need For High-Fidelity Text-to-Image Diffusion Models,” was published at the 2024 Conference on Computer Vision and Pattern Recognition (CVPR). 

Another paper has been accepted at CVPR 2026, coming up in June: "RoPE: Action-Controllable Infinite Video Generation Emerges From Autoregressive Self-Rollout."

Meral’s other published work includes:

·      “MotionFlow: Attention-Driven Motion Transfer in Video Diffusion Models,” presented in 2026 AAAI Conference on Artificial Intelligence main track.

·      “CLoRA: A Contrastive Approach to Compose Multiple LoRA Models,” presented in 2025 International Conference on Computer Vision (ICCV) highlights

·      “ConceptAttention: Diffusion Transformers Learn Highly Interpretable Features,” oral presentation at 2025 International Conference on Machine Learning (ICML)

·      “Conditional Information Gain Trellis,” published in Pattern Recognition Letters, 2024

“I value the visibility and support I’ve received from the Sanghani Center,” Meral said. “The center actively highlights student achievements, including publications, presentations, awards, and internships, which helps bring attention to our work beyond our immediate research circles.”

After his graduation, projected for May 2027, he plans to pursue a role that combines theoretical research with engineering pragmatism, likely as a research scientist in an industrial lab. 

“I enjoy research that leads to systems people can actually use, so I would like to continue developing methods that translate into tools that empower users. In the long run, I hope to contribute to generative AI systems that are both highly capable and genuinely controllable,” he said.