Abstract: Traditional photosensitizer discovery relies on slow, trial-and-error methods. We present \textbf{A}I-\textbf{A}ccelerated \textbf{P}hoto\textbf{s}ensitizer \textbf{I}nnovation (AAPSI), a closed-loop workflow combining scaffold-based generative AI, graph transformers, and Bayesian optimization to design high-performance photosensitizers efficiently. Leveraging a database of 102,836 samples (23,763 unique molecules) and 23 natural product-derived scaffolds, AAPSI generates 6,148 candidates, screened via models predicting singlet oxygen quantum yield ($\phi_\Delta$) and absorption maxima ($\lambda_{max}$). Pareto optimization prioritizes top candidates, including HB4Ph ($\phi_\Delta$=0.85, $\lambda_{max}$=650nm), which exhibits superior photodynamic therapy performance. AAPSI ensures synthetic feasibility and novelty while drastically reducing discovery timelines.