Extensive machine-learning assisted research has been dedicated to predicting band gaps for perovskites, driven by their immense potential in photovoltaics. Yet, the effectiveness is often hampered by the lack of high-quality band-gap datasets, particularly for perovskites involving d orbitals. In this work, we consistently calculate a large dataset of band gaps with a high level of accuracy, which is rigorously validated by experimental and state-of-the-art GW band gaps. Leveraging this achievement, our machine-learning derived descriptor exhibits exceptional universality and robustness, proving effectiveness not only for single and double, halide and oxide perovskites regardless of the underlying atomic structures, but also for hybrid organic-inorganic perovskites. With this approach, we comprehensively explore up to 15,659 materials, unveiling 14 unreported lead-free perovskites with suitable band gaps for photovoltaics. Notably, MASnBr₃, FA₂SnGeBr₆, MA₂AuAuBr₆, FA₂AuAuBr₆, FA₂InBiCl₆, FA₂InBiBr₆, and Ba₂InBiO₆ stand out with direct band gaps, small effective masses, low exciton binding energies, and high stabilities.