Introduction
In modern orthodontics, monitoring patient progress typically necessitates expensive and time-consuming 3D intra-oral scans. While patients can easily capture 2D photographs of their teeth using smartphones, reconstructing a high-fidelity 3D model from these sparse, uncalibrated images remains a significant technical challenge. Standard five-view intra-oral photography (anterior, left/right buccal, and maxillary/mandibular occlusal) offers minimal overlap, causing traditional Structure-from-Motion (SfM) and Multi-View Stereo (MVS) algorithms to fail.
TeethDreamer addresses this by leveraging a 3D-aware diffusion model to "hallucinate" consistent multi-view images and normal maps, enabling high-quality 3D dental reconstruction from limited input.
Methodology
The TeethDreamer framework operates in two primary stages:
1. Multiview Cross-domain Diffusion
The system adapts a pretrained diffusion model (Zero123) to generate additional viewpoints. To overcome the limitations of RGB data alone—which can be uniform across teeth—TeethDreamer incorporates a "domain switcher" to generate both color images and surface normal maps.
To ensure cross-view consistency, the authors introduce a 3D-aware feature attention mechanism. This process involves backprojecting 2D features into a 3D voxel grid, processing them through a 3D CNN to capture spatial relationships, and then using these refined features to guide the denoising process in the diffusion model. This prevents common generative artifacts like the "Janus problem," where different views of the same object are physically inconsistent.
2. Geometry-aware Neural Surface Reconstruction
Following the generation of dense views and normals, the system employs NeuS (Neural Implicit Surface Reconstruction) to synthesize the final 3D mesh. To mitigate the impact of generative noise, a geometry-aware normal loss is introduced. This loss function utilizes a weighting scheme that automatically discounts generated normals that are physically implausible (e.g., pointing away from the camera), ensuring the final geometry is anchored by the most reliable data.
Dataset and Training
The model was trained on a curated dataset of 3,200 clinical cases, each consisting of five standard intra-oral photographs and a corresponding high-accuracy 3D intra-oral scan. To facilitate training, the researchers used Blenderproc to render synthetic 2D color images and normal maps from the 3D scans at eight specific viewpoints. The dataset was partitioned into 3,000 cases for training, 100 for validation, and 100 for testing.
Results and Evaluation
TeethDreamer demonstrates significant improvements over existing state-of-the-art methods such as Zero123 and SyncDreamer.
Quantitative Performance
TeethDreamer achieved a Chamfer Distance (CD) of 0.1670 mm and a Hausdorff Distance (HD) of 2.1126 mm, outperforming SyncDreamer (0.2045 mm CD / 2.5505 mm HD). Furthermore, the generated multi-view images reached a PSNR of 18.85, indicating superior fidelity to the actual dental structures.
Qualitative Comparison
Visually, TeethDreamer produces complete 3D models that capture personalized dental details, such as specific tooth misalignments, which are often missed by parametric template-based methods.
Conclusion
TeethDreamer represents a major step forward in remote dental monitoring by enabling accurate 3D reconstruction from standard smartphone photography. While the current reconstruction process takes approximately 10–20 minutes—limiting real-time feedback—future work aims to integrate faster techniques like Instant-NGP. Additionally, the team plans to explore the use of low-level features, such as edges, to further refine the boundaries between individual teeth and capture even finer surface details.
Summary updated