Sayak Paul

Sayak Paul is a senior research engineer at Hugging Face, where he works on the Diffusers library and conducts research on test-time scaling, image editing, and model evaluation. He is also interested in custom kernels and contributes to different aspects of the Kernels project at Hugging Face.

Beyond maintaining developer infrastructure, Paul has contributed to several advanced research initiatives exploring the future of generative AI systems. One of them is HeadHunter, a research project developed with KAIST researchers that investigates controllable diffusion models through attention head manipulation and perturbation guidance techniques. The project improves precision and controllability in AI-generated images while offering deeper insights into how diffusion architectures behave internally.

He has also been associated with ReflectionFlow, an experimental project exploring reasoning and reflection mechanisms in diffusion-based systems. The work reflects a broader industry shift toward combining generative AI with chain-of-thought style reasoning and agentic workflows.