Application of Diffusion Models in Limited Supervision Based Learning

This thesis explores ways in which Diffusion Models (a special class of Generative AI models) can be applied in real-world situations where human-labelled training data is scarce or imbalanced. It introduces three approaches: (a) to make these models more data-efficient, (b) better at generating rare categories and (c) better at personalizing to individual subjects. Finally, it presents a benchmark for advanced, context-aware deepfake detection. These advances can improve applications such as medical imaging and content authenticity research by generating underrepresented examples, reducing bias and creating more controllable image outputs. Overall, the work strengthens the reliability and versatility of Generative AI in practical settings.