In positron emission tomography, accurate quantification is essential for reliable diagnosis and treatment assessment. But the emitted photons (signal) can be attenuated in tissues before reaching the detectors. Without proper correction, attenuation can lead to quantitation errors in images, difficulty in distinguishing between benign and malignant conditions, and even misdiagnosis. Attenuation correction is typically performed using co-registered computed tomography (CT) scans, which offer anatomical data to estimate signal loss, but come at the expense of extra radiation exposure, potential misregistration and additional costly hardware. Cutting edge deep learning methods can generate a pseudo-CT instead. Here, Conditional Diffusion Probabilistic Models (DDPMs) are shown to outclass the previous state-of-the-art UNet model. Using all three orthogonal views of a non-attenuation-corrected PET image, DDPM is able to generate a higher quality pseudo-CT while reducing artifacts and other slice-to-slice inconsistencies. Preliminary results from a dataset of 159 head scans from the Siemens Biograph Vision PET/CT scanner show both qualitatively and quantitatively significantly improved pseudo-CTs.