Abstract: This study aimed to systematically evaluate the applicability of quality-limited and unstructured imagery acquired with a consumer-grade smartphone for estimating individual-tree parameters to support low-cost urban tree inventory and dynamic monitoring. Twenty isolated deciduous broadleaf trees with diameter at breast height (DBH) values ranging from 8.9 to 58.6 cm were selected on the campus of Northwest A&F University in Yangling, Shaanxi, China, during the winter leaf-off period. Between 71 and 263 images per tree were collected with a smartphone via 360° circumferential photography at distances of 2 to 5 m from the trunk, with a graduated rod placed beside each tree for metric scale recovery. The same image sets were processed using multi-view stereo (MVS), neural radiance fields (NeRF), and 3D Gaussian splatting (3DGS), implemented with COLMAP, Nerfacto, and Splatfacto, respectively. The resulting point clouds were manually segmented in CloudCompare to retain only the target tree, scaled using the known intervals on the graduated rod, and corrected for orientation by principal component analysis. Tree height was extracted directly from the vertical extent of each orientation-corrected single-tree point cloud. The corrected point clouds were processed with TreeQSM to construct quantitative structure models (QSMs) for DBH extraction. A total of 24 parameter combinations were tested, with each combination repeated 5 times to generate 120 QSMs per point cloud, and the optimal model was selected using the minimum mean point-to-model distance. Estimation accuracy was assessed against field measurements using root mean square error (RMSE) and relative root mean square error (RRMSE). Processing time was recorded for MVS dense reconstruction and for NeRF and 3DGS training, excluding the structure-from-motion stage common to all three methods. All three methods completed reconstruction for all 20 trees, with mean final single-tree point counts of approximately 1.055 million for MVS, 0.590 million for NeRF, and 0.024 million for 3DGS. MVS point clouds contained relatively complete trunks but showed discontinuities and noise in some smaller branches; NeRF point clouds retained more continuous branches and more complete canopies; 3DGS point clouds were relatively sparse and represented fewer fine branches. The average processing times of MVS, NeRF, and 3DGS were 149.3, 127.8, and 20.2 min per tree, respectively, and the time required by 3DGS was approximately one-sixth that of NeRF and one-seventh that of MVS. TreeQSM generated QSMs with relatively continuous main-stem cylinders from the point clouds of all three methods. Differences occurred mainly in canopy branches: MVS-based QSMs fitted major and some secondary branches but contained local discontinuities or abrupt directional changes; NeRF-based QSMs showed more continuous branch connections; and some 3DGS-based QSMs contained locally enlarged cylinder radii or abnormal branch trajectories. For DBH estimation, MVS yielded the lowest errors, with an RMSE of 0.71 cm and an RRMSE of 3.87%. For NeRF and 3DGS, the RMSE values were 2.19 and 4.09 cm, respectively, and the RRMSE values were 12.00% and 22.38%, respectively. Both NeRF and 3DGS underestimated DBH for all 20 trees. For tree height estimation, the RRMSE values of all three methods were below 5%. NeRF yielded the lowest errors (RMSE = 0.09 m and RRMSE = 1.47%), followed by 3DGS (0.12 m and 2.10%) and MVS (0.26 m and 4.53%). The differences among the methods in tree height estimation were smaller than those in DBH estimation. The findings can provide a technical reference for developing a low-cost and efficient forest observation system with public participation and for improving fine-scale forest resource management.