Airborne Laser Scanning (ALS) enables efficient large-scale mapping of three-dimensional forest structure due to its extensive spatial coverage, but its low point density often limits the accurate representation of forest structural attributes required for applications such as wildfire management and forest health monitoring. This study proposes a deep generative upsampling framework to enhance low-density ALS tree point clouds and generate dense representations for improved structural analysis. To improve tree reconstruction, the framework incorporates a synthetic pretraining strategy that exposes the model to complete tree geometries before fine-tuning on real data. Experimental results demonstrate that the proposed framework generates high-density tree point clouds with strong spatial similarity to ground-truth point clouds. Synthetic pretraining further improves reconstruction performance, achieving lower Chamfer Distance and higher F1-scores across multiple upsampling rates. In addition, the reconstructed point clouds increase estimation accuracy of key structural attributes, including tree height, crown area, and tree volume. Qualitative experiments on real ALS data confirms the ability of the proposed framework to recover realistic tree structures from sparse point clouds. Overall, the proposed method provides a practical approach for bridging the wide spatial coverage of ALS with the detailed structural information typically obtained from high-density lidar data.