In his research focusing on enhancing the controllability and interpretability of large-scale generative models such as diffusion and rectified flow transformers, Ph.D. student Yusuf Dalva has moved away from treating these artificial intelligence systems as “black boxes.” 

“As the AI field has moved toward the massive models we see today, incredibly powerful tools are released almost weekly,” he said, “and while we are all impressed by what they can generate, we don't actually know how to make them do exactly what we want. They feel like high-performance cars without steering wheels.”

This is the gap that drives Dalva’s research.

“I’m not just interested in what these models can produce; I want to understand their internal logic so I can build the controls and turn these 'black boxes' into reliable tools that users actually need,” he said.

In many current AI systems, generating multiple specific identities, such as two distinct people, often leads to concept blending, where the model confuses or blends their unique identities that could be included in the generated image.

“By investigating the model’s internals to isolate and manipulate these specific identities, such as facial features of celebrities, I’ve developed systems that allow for the seamless composition of multiple distinct characters in a single scene while maintaining the realism and interactions between them,” said Dalva, who is advised by Pinar Yanardag. “This has profound implications for content creation and filmmaking because it transforms generative AI from a random image generator into a professional-grade tool.” 

For example, he said, a director could use these frameworks to consistently cast specific digital characters across various scenes and complex interactions, ensuring they look exactly the same from every angle and in every shot, a level of reliability that is essential for high-end storytelling.

Dalva was named an Amazon Fellow supported by the Amazon -Virginia Tech Initiative for Efficient and Robust Machine Learning for the 2025-26 academic year. Amazon Fellowships are awarded to Virginia Tech doctoral students pursuing educational and research experiences in AI-related fields who are recognized for their scholarly achievements and potential for future accomplishments.

The following are included in Dalva’s published work:

Dalva said that whether it's a casual brainstorm in the lab or a formal collaboration, the environment at the Sanghani Center that encourages students to learn from one another's diverse expertise has been incredibly helpful in his own growth as a researcher.

“From the start, I felt empowered to explore my specific interests in generative models and controllability, which was built on the foundation I had from my master’s studies. The culture of taking the initiative to design projects, while having the support of world-class faculty, allows for a level of intellectual freedom that is essential for pushing the boundaries of AI,” he said.

After graduation, projected for May 2027, Dalva would like to join an industrial research lab developing models of the future -- systems that are not just more powerful, but are designed from the ground up to be reliable, precise, and capable of following complex human intent. 

“Ultimately,” he said, “I want to contribute to a future where AI serves as a high-fidelity partner for professionals, allowing them to realize their creative visions with absolute control.”