TeethGenerator: Paired 3D Dental Data Synthesis
- TeethGenerator is a two-stage framework synthesizing paired pre- and post-orthodontic 3D dental models while enforcing anatomical constraints.
- Stage I employs a 3D U-Net VQ-VAE with diffusion modeling to generate realistic post-orthodontic morphologies, achieving high fidelity and uniqueness.
- Stage II uses Transformer-based style transfer to apply clinically plausible per-tooth transformations, enhancing downstream tooth arrangement performance.
TeethGenerator most specifically denotes a two-stage framework for synthesizing paired pre- and post-orthodontic $3$D teeth models in order to mitigate the scarcity of paired clinical dental data for downstream tooth arrangement networks (Lei et al., 7 Jul 2025). In a broader technical usage, the same label also appears in implementation-oriented descriptions of adjacent dental generative systems, including $2$D orthodontic visualization from photographs, spectral tooth-shape modeling, panoramic X-ray to $3$D reconstruction, and crown generation (Dou et al., 2023, Kubík et al., 30 Jun 2026, Ma et al., 2024, Bae et al., 26 Dec 2025). Within this literature, the central problem is not merely isolated shape synthesis, but generation under strong anatomical structure: a dental model typically contains $24$–$32$ segmented teeth whose morphology, relative layout, and inter-tooth constraints must remain clinically plausible (Lei et al., 7 Jul 2025).
1. Terminological scope and problem formulation
In its strictest sense, TeethGenerator refers to the framework introduced in “TeethGenerator: A two-stage framework for paired pre- and post-orthodontic 3D dental data generation” (Lei et al., 7 Jul 2025). That work is motivated by a stated bottleneck in digital orthodontics: collecting paired $3$D orthodontic teeth models is labor-intensive, yet such paired data are crucial for training tooth arrangement neural networks. The paper further states that numerous general $3$D shape generation methods are insufficient for this setting because they typically focus on single-object generation rather than anatomically structured teeth models containing many segmented teeth (Lei et al., 7 Jul 2025).
A recurrent theme across related work is that “teeth generation” is modality-dependent. In frontal orthodontic visualization, the goal is to synthesize an aligned-teeth facial image from a facial photograph or from a photograph plus a patient-specific $3$D teeth model (Dou et al., 2023, Shen et al., 2022). In restorative workflows, the target is a patient-customized crown or a missing-tooth completion conditioned on neighboring dentition (Bae et al., 26 Dec 2025, Tian et al., 4 Mar 2026). In reconstruction settings, the target is a $3$D point cloud or mesh inferred from sparse $2$D observations such as a single panoramic X-ray or five intra-oral photographs (Ma et al., 2024, Xu et al., 2024). This suggests that TeethGenerator is best understood as a family of conditional generative formulations specialized to dental morphology, arch structure, and orthodontic or prosthetic constraints.
A common misconception is that dental generation is equivalent to generic shape synthesis. The paired-data formulation in TeethGenerator directly contradicts that view: the system is designed to synthesize both post-orthodontic morphology and a corresponding pre-orthodontic configuration, with explicit per-tooth transformations in the second stage (Lei et al., 7 Jul 2025). Related systems likewise introduce dental-specific priors such as orthodontic movement maps, inter-tooth attention, spectral synchronization, or collision penalties rather than treating teeth as unconstrained objects (Shen et al., 2022, Tian et al., 4 Mar 2026, Kubík et al., 30 Jun 2026).
2. Two-stage paired $2$0D generation architecture
The canonical TeethGenerator pipeline has two modules: a teeth shape generation module and a teeth style generation module (Lei et al., 7 Jul 2025). Stage I begins from real post-orthodontic $2$1D teeth point clouds $2$2, represented as $2$3 teeth with $2$4 points per tooth. A $2$5D U-Net-based VQ-VAE encodes $2$6 into discrete latent codes $2$7, with latent dimension $2$8, and a diffusion model learns the latent distribution so that a Gaussian latent $2$9 can be denoised into $3$0, which is then decoded into a generated post-orthodontic point cloud $3$1 (Lei et al., 7 Jul 2025).
The forward diffusion process in Stage I is written as
$3$2
$3$3
The denoising network $3$4 is trained by
$3$5
Stage I therefore models the morphology distribution of post-orthodontic dentitions in a learned latent space rather than operating directly on raw point sets (Lei et al., 7 Jul 2025).
Stage II is conditional style generation. Its input is the generated post-orthodontic point cloud $3$6 and a “style” point cloud $3$7, which may be any real or synthetic pre-orthodontic sample (Lei et al., 7 Jul 2025). The style extractor $3$8 computes per-tooth style vectors $3$9 from voxelized tooth representations using PVCNN, mean/std pooling, and small MLPs; the shape extractor $24$0 computes a global code $24$1 from the whole model by global voxelization and max-pooling (Lei et al., 7 Jul 2025). A Transformer with $24$2 blocks and $24$3 heads conditions on $24$4 as tokens and on $24$5 via cross-attention, then predicts per-tooth transformation parameters
$24$6
where the $24$7-D rotation representation is mapped continuously to a $24$8 rotation matrix and applied, together with translation, to each tooth of $24$9 to obtain $32$0 (Lei et al., 7 Jul 2025).
The conditional distribution is written as
$32$1
Training minimizes a distance loss on reconstructed pre-orthodontic clouds together with a collision-avoidance term: $32$2
$32$3
$32$4
This design distinguishes between morphology generation and orthodontic-style transfer rather than entangling both in a single generator (Lei et al., 7 Jul 2025).
3. Representation, training protocol, and evaluation methodology
The dataset for the paired $32$5D TeethGenerator contains $32$6 pre/post pairs, filtered to $32$7 training samples, $32$8 validation samples, and $32$9 test samples; each sample contains approximately $3$0–$3$1 teeth, with $3$2 points per tooth obtained via FPS (Lei et al., 7 Jul 2025). The framework also produces a synthetic set of $3$3 generated post models paired with pre models (Lei et al., 7 Jul 2025).
Stage I uses PVCNN voxelization with $3$4, yielding a grid $3$5 for the per-jawed structure, a $3$6D U-Net VQ-VAE with a $3$7-dimensional latent codebook per voxel, and a diffusion U-Net using the same $3$8D U-Net backbone with sinusoidal timestep encodings and $3$9 diffusion steps (Lei et al., 7 Jul 2025). Stage II uses per-tooth voxelization of $3$0, style vectors $3$1, a global shape code $3$2, a Transformer with hidden dimension $3$3, and an MLP head mapping $3$4 outputs per tooth token (Lei et al., 7 Jul 2025).
Training is implemented in PyTorch on Linux using $3$5V100 GPUs. Stage I uses batch size $3$6, $3$7 epochs, and AdamW with $3$8, learning rate $3$9, and weight decay $3$0. Stage II uses batch size $3$1, $3$2 epochs, AdamW learning rate $3$3, and weight decay $3$4 (Lei et al., 7 Jul 2025). Point sampling normalizes teeth to a unit box and uses FPS with $3$5 points per tooth; mesh reconstruction is explicitly described as visualization only, via non-rigid registration of a high-resolution template mesh to generated points (Lei et al., 7 Jul 2025).
The evaluation protocol emphasizes both fidelity and distributional similarity. The synthetic distribution is compared with the real distribution using $3$6-NN accuracy under both Chamfer Distance and EMD, where ideal $3$7-NNA is approximately $3$8 for perfectly matched distributions (Lei et al., 7 Jul 2025). Uniqueness is measured by
$3$9
with $3$0 cm (Lei et al., 7 Jul 2025). This metric suite reflects the stated purpose of the framework: not only to reconstruct plausible individual samples, but to provide synthetic data that are varied enough to improve downstream learning.
4. Empirical behavior and downstream orthodontic utility
For post-orthodontic generation in Stage I, the reported comparison on $3$1 samples includes PointFlow, DPM, PVD, LION, DiT-3D, and TeethGenerator (Lei et al., 7 Jul 2025). TeethGenerator reports $3$2, $3$3, and $3$4, whereas the listed baselines report higher CD and EMD percentages but substantially lower uniqueness, such as PointFlow with $3$5, $3$6, and $3$7, and DiT-3D with $3$8, $3$9, and $2$0 (Lei et al., 7 Jul 2025). The paper states that the synthetic dataset aligns closely with the distribution of real orthodontic data and promotes tooth alignment performance when combined with real training data (Lei et al., 7 Jul 2025).
For the Stage II style-transfer module, the reported qualitative behavior includes faithful transfer of crowding, over-bite, and open-bite styles while preserving per-tooth morphology (Lei et al., 7 Jul 2025). The ablation findings are also specific: swapping $2$1 and $2$2 voxelization, or removing voxelization, produces collapsed styles or gaps (Lei et al., 7 Jul 2025). These observations indicate that the decomposition into per-tooth style and whole-arch shape is not merely architectural convenience but functionally necessary for stable conditional synthesis.
The most direct downstream result is on a TANet backbone for tooth alignment. Training on real data only ($2$3 samples) yields $2$4 mm, $2$5-$2$6 mm, and $2$7 mm; training on real plus $2$8 synthetic data ($2$9 additional samples) improves these to $2$00, $2$01-$2$02, and $2$03 (Lei et al., 7 Jul 2025). The summary further states that improvement saturates around $2$04–$2$05 synthetic data (Lei et al., 7 Jul 2025). A plausible implication is that the principal value of TeethGenerator lies less in stand-alone synthesis than in controlled data augmentation for scarce orthodontic supervision.
5. Relation to adjacent dental generative paradigms
The TeethGenerator literature intersects with several neighboring paradigms that solve different dental generation problems with different modalities and conditioning signals. In $2$06D orthodontic visualization, “3D Structure-guided Network for Tooth Alignment in 2D Photograph” uses a segmentation module, a contour-alignment diffusion model, and a generation diffusion model to synthesize an aligned-teeth mouth patch from a facial photograph while learning orthodontic movement from projected pre/post $2$07D intra-oral scans (Dou et al., 2023). OrthoGAN likewise synthesizes identity-preserving frontal facial images with aligned teeth from a frontal face image and a patient’s $2$08D scanned teeth model, conditioning on rendered silhouette and depth-mask maps derived from doctor-specified per-tooth rigid motions (Shen et al., 2022). These systems target patient communication and treatment visualization rather than paired $2$09D data augmentation.
A second cluster addresses compact or intrinsic $2$10D tooth-shape modeling. ToothForge uses synchronized spectral embeddings and a $2$11-VAE on $2$12-dimensional spectral coefficients to model dental crown geometries in a low-dimensional latent manifold, reporting on the molar class $2$13, Coverage $2$14, train time $2$15 min, and sample time $2$16 ms for $2$17 (Kubík et al., 30 Jun 2026). This differs from TeethGenerator in both target and representation: ToothForge models single-tooth crown geometry from intrinsic spectral coefficients, whereas TeethGenerator models paired whole-dentition pre/post data with explicit per-tooth transformations (Kubík et al., 30 Jun 2026, Lei et al., 7 Jul 2025).
A third cluster reconstructs or completes $2$18D dentition from sparse observations. PX2Tooth reconstructs $2$19D teeth from a single panoramic X-ray using PXSegNet and TGNet with a Prior Fusion Module, reporting $2$20, $2$21-CD $2$22, and $2$23-EMD $2$24 on a $2$25-pair CBCT–panoramic dataset (Ma et al., 2024). TeethDreamer reconstructs a $2$26D dental mesh from five intra-oral photographs via a multiview diffusion stage and Neus-based implicit surface reconstruction, reporting $2$27 mm, $2$28 mm, and $2$29 on $2$30 test cases (Xu et al., 2024). These methods are closer to inverse reconstruction than to paired data synthesis.
Restorative generation introduces still other priors. CrownGen uses a boundary prediction module and a diffusion model over tooth-level point clouds to generate patient-customized crowns, and its clinical study reports a TeethGenerator-assisted workflow time of $2$31 s versus $2$32 s for manual CAD, with statistically non-inferior reader-study scores (Bae et al., 26 Dec 2025). DM-CFO addresses compositional $2$33D tooth generation with graph diffusion for layouts and score distillation sampling with collision regularization, reporting Chamfer Distance $2$34 mm, F-Score $2$35, and Penetration Distance $2$36 mm (Tian et al., 4 Mar 2026). TADPM, by contrast, uses diffusion probabilistic modeling over per-tooth $2$37-DoF transformation matrices for automatic tooth arrangement and reports $2$38 mm, $2$39-$2$40 mm, $2$41, and $2$42 mm (Lei et al., 2023). Together these works show that “teeth generation” spans morphology synthesis, arrangement, restoration, reconstruction, and visualization, with TeethGenerator (Lei et al., 7 Jul 2025) occupying the specific niche of paired orthodontic $2$43D data generation.
6. Limitations, misconceptions, and research directions
One misconception exposed by the broader literature is that realistic tooth generation can be achieved without explicit structural constraints. Related work shows that this is often false. The $2$44D-only regime in the structure-guided tooth-alignment network cannot enforce realistic $2$45D collisions or occlusal contacts, and it may fail on extremely rotated or crowded teeth or very wide smiles (Dou et al., 2023). DM-CFO explicitly introduces collision-free optimization because existing compositional $2$46D generation may omit collision conflicts and allow intersections among neighboring teeth (Tian et al., 4 Mar 2026). These observations suggest that dental generation benefits from constraints that go beyond appearance or isolated shape fidelity.
Another misconception is that sparse-view or weakly conditioned generation is sufficient for all dental anatomy. PX2Tooth notes that occluded tooth roots in panoramic X-rays lead to under-constrained apex geometry, and that jawbones and neurovascular canals are not reconstructed (Ma et al., 2024). TeethDreamer reports that neural surface reconstruction takes approximately $2$47–$2$48 minutes per patient and that very high-frequency enamel details remain under-recovered (Xu et al., 2024). Such limitations are not unique to those systems, but they indicate why future TeethGenerator-like models may need richer geometry priors, faster reconstruction backbones, or hybrid implicit–explicit representations.
For the paired-data setting itself, the major open direction is to combine morphological realism with stronger anatomical and biomechanical plausibility. The broader literature already points toward several mechanisms: actual $2$49D alignment followed by reprojection for realism in orthodontic visualization (Dou et al., 2023), surface-aware collision metrics in compositional generation (Tian et al., 4 Mar 2026), and compact intrinsic representations for limited datasets (Kubík et al., 30 Jun 2026). Within that landscape, TeethGenerator (Lei et al., 7 Jul 2025) represents a specific and practically important shift: from single-sample dental synthesis toward synthetic paired datasets designed to improve downstream orthodontic learning under clinical data scarcity.