TAlignDiff: Diffusion-Based Tooth Alignment

• TAlignDiff is a learning-based method for automatic orthodontic tooth alignment that predicts per-tooth 4x4 rigid transformation matrices from pre-treatment 3D dental point clouds.
• It fuses a primary regression network using global and local PointNet encoders with a diffusion-based denoising module to regularize anatomically plausible tooth movements.
• Empirical results on a clinical dataset show significant improvements in Target Registration Error and Absolute Arch Error compared to established baselines.

Searching arXiv for the exact term and closely related papers to ground the article. Search query: "TAlignDiff" TAlignDiff is a learning-based method for automatic orthodontic tooth alignment that predicts per-tooth rigid motions from pre-treatment 3D dental point clouds to a target post-treatment arrangement by combining point cloud regression with diffusion-based transformation modeling (Liu et al., 6 Aug 2025). In the cited literature, the name specifically refers to “TAlignDiff: Automatic Tooth Alignment assisted by Diffusion-based Transformation Learning” (Liu et al., 6 Aug 2025). The broader diffusion-alignment literature also uses superficially similar terminology for inference-time alignment of generative diffusion models, but that constitutes a different research line; in that separate context, TAlignDiff is best understood as an unrelated naming collision rather than a shared method family (Kim et al., 25 Nov 2025).

1. Definition and scope

TAlignDiff addresses tooth-wise alignment in orthodontics: given a pre-treatment 3D dental model represented as tooth point clouds, it predicts how each tooth should be moved into its post-treatment arrangement (Liu et al., 6 Aug 2025). The method assumes that orthodontic treatment can be represented as per-tooth rigid transformations from the initial to final arrangement, rather than tooth deformation. Its output is a set of $4\times 4$ transformation matrices, one for each tooth, comprising a rotation matrix and a translation vector (Liu et al., 6 Aug 2025).

The input point cloud domain is written as

$$P=\{p_i \mid i=1,\dots,N\}\subseteq \mathbb{R}^{N\times 3},$$

where $N$ is the total number of sampled 3D points. Each case is organized into $M=32$ tooth sub-point-clouds corresponding to the 32 permanent teeth. In the reported dataset, each case contains 4096 points total, arranged as 32 sub-point clouds of 128 points each (Liu et al., 6 Aug 2025). The paper denotes the pre-orthodontic point cloud as $P_{in}$ and the post-treatment target as $P_{gt}$ (Liu et al., 6 Aug 2025).

For each tooth $i$, TAlignDiff predicts

$$T_i= \begin{bmatrix} R_i & D_i\ 0 & 1 \end{bmatrix},$$

where $R_i\in\mathbb{R}^{3\times 3}$ is a rotation matrix and $D_i\in\mathbb{R}^{3\times 1}$ is a displacement vector (Liu et al., 6 Aug 2025). The target aligned point cloud is expressed as

$$P=\{p_i \mid i=1,\dots,N\}\subseteq \mathbb{R}^{N\times 3},$$0

This formulation places TAlignDiff in the class of transformation-regression methods, but with an additional distribution-learning prior over the space of anatomically plausible transformations (Liu et al., 6 Aug 2025).

2. Motivation and conceptual basis

The paper argues that prior automatic tooth alignment methods typically regress transformation parameters using point-to-point geometric constraints, such as minimizing discrepancies between predicted and target aligned point clouds (Liu et al., 6 Aug 2025). That approach is treated as insufficient because clinically valid transformation matrices are not arbitrary: they are associated with the anatomical structure of the human oral cavity and have particular distribution characteristics that deterministic geometric supervision alone does not capture (Liu et al., 6 Aug 2025).

This motivates the central hypothesis of TAlignDiff: tooth transformation matrices possess latent anatomical and distributional structure. In the paper’s framing, valid orthodontic movement patterns form a structured distribution in transformation space, reflecting plausible rotation and translation ranges, correlations between neighboring teeth, and whole-arch organization (Liu et al., 6 Aug 2025). A plausible implication is that a model constrained only by reconstruction loss may fit local geometry while still producing transformations that are less plausible at the level of the full dentition.

The method therefore combines two complementary signals. The first is explicit geometric supervision over aligned point clouds. The second is diffusion-based modeling of the latent distribution of valid transformation matrices (Liu et al., 6 Aug 2025). The paper describes this coupling as integrating point cloud-based transformation regression and diffusion-based transformation modeling into a unified framework with bidirectional feedback between geometric constraints and diffusion refinement (Liu et al., 6 Aug 2025).

3. Architecture and mathematical formulation

TAlignDiff comprises two components: a primary point cloud-based regression network (PRN) and a diffusion-based transformation matrix denoising module (DTMD) (Liu et al., 6 Aug 2025). PRN is the main inference-time predictor. DTMD is an auxiliary training-time module that learns the latent distribution of transformation matrices and regularizes PRN outputs (Liu et al., 6 Aug 2025).

The PRN regression is written as

$$P=\{p_i \mid i=1,\dots,N\}\subseteq \mathbb{R}^{N\times 3},$$1

where $$P=\{p_i \mid i=1,\dots,N\}\subseteq \mathbb{R}^{N\times 3},$$2 is a global PointNet encoder, $$P=\{p_i \mid i=1,\dots,N\}\subseteq \mathbb{R}^{N\times 3},$$3 is a local PointNet encoder, $$P=\{p_i \mid i=1,\dots,N\}\subseteq \mathbb{R}^{N\times 3},$$4 denotes feature concatenation, and $$P=\{p_i \mid i=1,\dots,N\}\subseteq \mathbb{R}^{N\times 3},$$5 is an MLP decoder (Liu et al., 6 Aug 2025). The feature extractor uses the encoder part of PointNet with three 1D convolutional layers of channels $$P=\{p_i \mid i=1,\dots,N\}\subseteq \mathbb{R}^{N\times 3},$$6. The decoder is an MLP with three fully connected layers of channels $$P=\{p_i \mid i=1,\dots,N\}\subseteq \mathbb{R}^{N\times 3},$$7 (Liu et al., 6 Aug 2025). The global encoder captures whole-dentition context, whereas the local encoder captures tooth-level geometry (Liu et al., 6 Aug 2025).

The predicted aligned point cloud is

$$P=\{p_i \mid i=1,\dots,N\}\subseteq \mathbb{R}^{N\times 3},$$8

PRN is trained using two geometric losses. The reconstruction loss is

$$P=\{p_i \mid i=1,\dots,N\}\subseteq \mathbb{R}^{N\times 3},$$9

which penalizes pointwise distance between predicted and target aligned point clouds (Liu et al., 6 Aug 2025). The tooth centroid offset loss is

$N$0

where $N$1 and $N$2 denote predicted and target tooth centroids (Liu et al., 6 Aug 2025). The centroid term is intended to enforce tooth-level displacement consistency beyond raw pointwise reconstruction.

DTMD operates in transformation space rather than on raw geometry. The target transformation matrices $N$3 are reshaped into $N$4, and the forward diffusion process is

$N$5

where $N$6 is the noisy transformation representation at timestep $N$7 and $N$8 is a timestep-dependent coefficient (Liu et al., 6 Aug 2025). The denoiser $N$9 is trained by standard noise prediction: $M=32$0 The paper does not specify the internal DTMD architecture, the exact matrix vectorization, the number of diffusion timesteps, or the reverse sampling equations (Liu et al., 6 Aug 2025). Those omissions are notable because they limit full reimplementation from the paper alone.

The bridge between PRN and DTMD is a contrastive denoising loss. Let $M=32$1 and $M=32$2 denote noisy ground-truth and noisy predicted transformations at timestep $M=32$3. Then

$M=32$4

The interpretation given in the paper is that if predicted transformations lie near the true transformation distribution, their denoising responses should resemble those of noisy ground-truth transformations (Liu et al., 6 Aug 2025). The full training objective is

$M=32$5

with best reported weights

$M=32$6

(Liu et al., 6 Aug 2025).

4. Training procedure and inference behavior

The training strategy is staged. For the first 200 epochs, PRN and DTMD are trained jointly using $M=32$7, $M=32$8, $M=32$9, and $P_{in}$0 (Liu et al., 6 Aug 2025). For the next 200 epochs, DTMD is frozen and PRN continues training while still receiving supervision from the pretrained DTMD (Liu et al., 6 Aug 2025). The reported training setup uses PyTorch, one NVIDIA 3090 GPU, Adam, PRN learning rate $P_{in}$1, DTMD learning rate $P_{in}$2, batch size $P_{in}$3, and total training time of 400 epochs (Liu et al., 6 Aug 2025).

Inference is deliberately simpler than training. Only PRN is used at test time: given $P_{in}$4, PRN predicts $P_{in}$5, and $P_{in}$6 is applied to the input teeth to obtain the aligned arrangement (Liu et al., 6 Aug 2025). DTMD is not used at inference, which the authors present as a practical advantage because diffusion-based supervision improves training without introducing deployment-time overhead (Liu et al., 6 Aug 2025).

This training–inference asymmetry distinguishes TAlignDiff from diffusion methods that perform iterative sampling at test time. A plausible implication is that the method is closer to diffusion-regularized discriminative regression than to generative diffusion inference. The paper itself supports this reading by emphasizing that DTMD acts as an auxiliary training module and does not explicitly refine transforms during deployment (Liu et al., 6 Aug 2025).

5. Data, evaluation, and empirical performance

The experiments use the Automatic Tooth Arrangement Challenge dataset from ISICDM 2024 (Liu et al., 6 Aug 2025). The dataset contains 124 patients collected in a clinical environment, split into 74 training, 20 validation, and 30 test cases (Liu et al., 6 Aug 2025). Each sample contains 32 permanent teeth and 4096 points total, arranged into 32 tooth sub-point clouds of 128 points each (Liu et al., 6 Aug 2025). The supervision consists of orthodontists’ treatment plans encoded as $P_{in}$7 transformation matrices (Liu et al., 6 Aug 2025).

To increase diversity given the limited sample size, the paper uses two augmentations. The first is multi-tooth rotation: randomly select $P_{in}$8 teeth and apply independent Euler-angle rotations under biomechanical constraints. The second is single-tooth translation. For consistency, inverse transformation matrices are computed for manipulated teeth to maintain correct labels (Liu et al., 6 Aug 2025).

The evaluation uses two metrics: Target Registration Error (TRE), defined as the distance between predicted and target tooth point clouds, and Absolute Arch Error (AAE), a newly introduced metric measuring difference between predicted and target post-arranged dental arches (Liu et al., 6 Aug 2025). Exact formulas are not provided in the paper excerpt.

The main comparison includes PointNet++, PointMLP, PSTN, TADPM, and the proposed TAlignDiff (Liu et al., 6 Aug 2025). Reported results are as follows.

Method	Validation	Test
PointNet++	TRE $P_{in}$9, AAE $P_{gt}$0	TRE $P_{gt}$1, AAE $P_{gt}$2
PointMLP	TRE $P_{gt}$3, AAE $P_{gt}$4	TRE $P_{gt}$5, AAE $P_{gt}$6
TADPM	TRE $P_{gt}$7, AAE $P_{gt}$8	TRE $P_{gt}$9, AAE $i$0
PSTN	TRE $i$1, AAE $i$2	TRE $i$3, AAE $i$4
TAlignDiff	TRE $i$5, AAE $i$6	TRE $i$7, AAE $i$8

The paper states that the improvements are statistically significant with $i$9 (Liu et al., 6 Aug 2025). Relative to PSTN, which is the strongest non-proposed baseline in the table, TAlignDiff improves both geometric alignment and whole-arch arrangement quality (Liu et al., 6 Aug 2025). The paper further argues that TADPM performs worse in this small-data setting because it conditions diffusion on high-dimensional geometric features, making it less data-efficient (Liu et al., 6 Aug 2025).

Ablations support the contribution of both centroid supervision and diffusion regularization. With only reconstruction loss, the test result is TRE $$T_i= \begin{bmatrix} R_i & D_i\ 0 & 1 \end{bmatrix},$$0, AAE $$T_i= \begin{bmatrix} R_i & D_i\ 0 & 1 \end{bmatrix},$$1. Adding centroid loss improves this to TRE $$T_i= \begin{bmatrix} R_i & D_i\ 0 & 1 \end{bmatrix},$$2, AAE $$T_i= \begin{bmatrix} R_i & D_i\ 0 & 1 \end{bmatrix},$$3. The best full setting $$T_i= \begin{bmatrix} R_i & D_i\ 0 & 1 \end{bmatrix},$$4 reaches TRE $$T_i= \begin{bmatrix} R_i & D_i\ 0 & 1 \end{bmatrix},$$5, AAE $$T_i= \begin{bmatrix} R_i & D_i\ 0 & 1 \end{bmatrix},$$6 (Liu et al., 6 Aug 2025). The main trend reported is that moderate centroid and denoising weights help, whereas larger $$T_i= \begin{bmatrix} R_i & D_i\ 0 & 1 \end{bmatrix},$$7 or $$T_i= \begin{bmatrix} R_i & D_i\ 0 & 1 \end{bmatrix},$$8 can hurt performance (Liu et al., 6 Aug 2025).

Qualitative analysis includes performance histograms, 3D scatter plots of predicted transformation parameters, and mesh visualizations for deep overbite cases (Liu et al., 6 Aug 2025). The paper interprets the more concentrated transformation distributions produced by TAlignDiff as evidence that transformation-distribution modeling stabilizes predictions (Liu et al., 6 Aug 2025). This suggests the diffusion prior acts less as a generative end in itself than as a regularizer on clinically plausible transformation space.

6. Terminological context and relation to other diffusion-alignment literature

The name TAlignDiff is specific to the orthodontic method above (Liu et al., 6 Aug 2025). However, the diffusion literature contains a separate line of work on inference-time alignment of text-to-image models, where papers such as “Test-Time Alignment of Text-to-Image Diffusion Models via Null-Text Embedding Optimisation” (Kim et al., 25 Nov 2025), “DeRaDiff: Denoising Time Realignment of Diffusion Models” (Manujith et al., 28 Jan 2026), and “HyperAlign: Hypernetwork for Efficient Test-Time Alignment of Diffusion Models” (Xie et al., 22 Jan 2026) address reward-conditioned alignment during sampling rather than clinical tooth arrangement.

That adjacent literature matters because one cited paper explicitly refers to “TAlignDiff-style methods” as a broad family of inference-time diffusion alignment methods that manipulate latents, injected noise, denoising trajectories, or particles to optimize a target reward (Kim et al., 25 Nov 2025). In that usage, the term does not denote the orthodontic method and instead functions as an informal label for test-time diffusion alignment approaches (Kim et al., 25 Nov 2025). Null-TTA distinguishes itself from such methods by optimizing the unconditional text embedding in classifier-free guidance rather than latent or noise variables (Kim et al., 25 Nov 2025). DeRaDiff instead performs denoising-time realignment by geometrically mixing aligned and reference reverse posteriors to emulate different KL regularization strengths during sampling (Manujith et al., 28 Jan 2026). HyperAlign trains a hypernetwork to generate input- and timestep-dependent low-rank weight updates for efficient test-time diffusion alignment (Xie et al., 22 Jan 2026).

This terminological split is substantial. TAlignDiff in orthodontics is a point-cloud-based regression framework with diffusion-assisted transformation learning used only during training (Liu et al., 6 Aug 2025). The inference-time alignment papers are generative test-time control methods for text-to-image diffusion models (Kim et al., 25 Nov 2025, Manujith et al., 28 Jan 2026, Xie et al., 22 Jan 2026). The shared lexical elements—alignment and diffusion—do not imply methodological continuity.

A second neighboring literature concerns extraction or calibration of text-image grounding rather than reward alignment. For example, ELBO-T2IAlign studies pixel-level image/class-level text alignment in pretrained diffusion models and calibrates cross-attention maps using ELBO-derived classwise scores (Zhou et al., 11 Jun 2025). That work is relevant to alignment in diffusion models but again unrelated to automatic orthodontic tooth alignment (Zhou et al., 11 Jun 2025).

The most precise encyclopedic interpretation is therefore twofold. First, TAlignDiff properly names the orthodontic framework introduced in 2025 (Liu et al., 6 Aug 2025). Second, similar strings appear in diffusion-model alignment discourse, but those refer to distinct research problems and should not be conflated (Kim et al., 25 Nov 2025).

7. Limitations and open questions

The paper’s explicit limitations discussion is brief, but several constraints are identifiable from the method description (Liu et al., 6 Aug 2025). The dataset is relatively small, comprising 124 patients, and the authors note future work on larger datasets and incorporation of more clinical context (Liu et al., 6 Aug 2025). The method assumes rigid $$T_i= \begin{bmatrix} R_i & D_i\ 0 & 1 \end{bmatrix},$$9 transforms for each tooth and does not model tooth deformation (Liu et al., 6 Aug 2025). It also does not specify explicit orthogonality enforcement for $R_i\in\mathbb{R}^{3\times 3}$0, nor any regularization guaranteeing valid rotations (Liu et al., 6 Aug 2025).

Architectural detail for DTMD is incomplete. The paper does not report the exact transformation-matrix vectorization, whether all 16 matrix entries are used in diffusion, the diffusion timestep count, the noise schedule, or the denoiser architecture (Liu et al., 6 Aug 2025). This leaves ambiguity about the exact inductive bias of the transformation-space model. Likewise, preprocessing details such as segmentation, normalization, and coordinate standardization are not described in the provided summary (Liu et al., 6 Aug 2025).

A broader methodological caveat is that DTMD is not used at inference. This is a deliberate efficiency choice, but it means the learned transformation prior affects predictions only through training-time regularization (Liu et al., 6 Aug 2025). A plausible implication is that deployment-time performance depends entirely on how well PRN internalizes the prior, rather than on explicit posterior refinement. The paper presents this as an advantage in runtime terms, but it also limits the model’s capacity for instance-specific correction during inference.

Despite these caveats, the reported evidence supports the central claim that modeling the distribution of anatomically plausible tooth motions improves automatic orthodontic alignment (Liu et al., 6 Aug 2025). In that sense, TAlignDiff occupies a specific methodological niche: it is neither a pure deterministic regressor nor a full generative diffusion inference system, but a hybrid framework in which diffusion serves as a transformation-space regularizer for clinically constrained geometric prediction.