Photo-z’s are crucial for the main cosmology, galaxy evolution, and transient science drivers of Roman. The transformative nature of the Roman dataset presents both new challenges but also new opportunities for photo-z estimation. First, I will describe our photo-z forecasts that helped guide the HLIS survey design recommendations, especially the filter choices across the medium and wide tiers. These efforts assumed a fair spectroscopic redshift dataset for training and calibration, but that assumption is not currently achieved in deep imaging surveys. To address this need, our group is pursuing a multi-faceted approach to obtain additional spectroscopy, develop better re-weighting/interpolation methods, and implement an independent cross-check on the redshift distributions using clustering redshifts. Finally, I will discuss our efforts to understand whether deep learning models can leverage the information in high-resolution space-based imaging to achieve better photo-z’s than is possible with photometry alone.