<p>Sensor generated point clouds are often noisy and sparse, which pose challenges in various downstream tasks. Recent approaches have tackled the problem through various upsampling approaches. Existing methods suffer from outlier and variable point cloud density leading to loss in object geometry. This work introduces Point-PAD, an architecture which utilizes kernel points and positional encodings in order to encode the local geometry, which is then processed by a multi-head attention block for adjusting predicted points with the entire architecture supported by a density aware loss function. The architecture, when tested on the PU-GAN dataset, achieves state-of-the-art results. The model outperformed the state-of-the-art by 3.22% and was able to upsample points even at sharp curves while preserving the geometry. We also report results on one of the first outdoor dataset comprising real life LiDAR (Light Detection and Ranging) scans of outdoor scenes for this task. Our model outperforms the existing work, proving its generalization, which is crucial for applications related to digital twins and autonomous systems. The code is available at: <a href="https://github.com/geoai4cities/PointPAD">https://github.com/geoai4cities/PointPAD</a></p>