acceptodds
Under review as a conference paper at ICLR 2027

CrownGen: Clinically Applicable Dental Crown Generation via Physical-Reward Generative Modeling

Abstract

We present CrownGen, a fully automatic framework for clinically applicable single-crown generation from raw intraoral scans. Dental crown synthesis requires both high-fidelity anatomical morphology and precise functional relations to neighboring and opposing teeth, which are difficult to capture with reconstruction-driven generative objectives alone. CrownGen combines margin-aware latent diffusion, ambiguity-preserving morphology refinement, and reward-guided clinical applicability adaptation. The morphology generation stage produces anatomically coherent crowns with high-frequency dental structures, while the adaptation stage uses non-differentiable physical rewards from proximal contact and occlusal clearance to improve functional validity. Experiments and professional user studies demonstrate superior geometric fidelity, anatomical realism, functional validity, and clinical acceptability over existing learning-based baselines.

Primary area:
generative models
Keywords:
Generative 3D Modeling

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_7d6jOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CrownGen, a dental crown generation framework that combines margin-aware generative modeling, morphology refinement, and physical-reward adaptation. The method aims to generate anatomically plausible crowns while improving their compatibility with adjacent and opposing teeth. The experiments demonstrate promising geometric accuracy and functional adaptation. Overall, the article has clear ideas, reasonable structure, and sufficient experiments. Although, there are some concerns on the role and evaluation of the predicted margin line.

Critical strengths

The paper addresses a clinically meaningful problem by considering both anatomical morphology and functional constraints in dental crown generation. The proposed framework is technically reasonable, and the integration of morphology generation with physical reward adaptation is well motivated. The experiments are relatively comprehensive, including comparisons with existing methods, ablation studies, geometric evaluations, and expert assessments. The improvements in functional validity metrics, DNT and DTA, are particularly encouraging.

Critical weaknesses

Major: (1) The margin line is jointly predicted during the first DiT training, but is not explicitly used in subsequent mesh reconstruction or refinement. And MLD is computed between the expert annotated margin line and the generated crown surface rather than the directly predicted margin line. Please clarify how the margin branch contributes to the final crown geometry, whether the generated margin shown in Fig. 3(b) comes from the prediction branch or is extracted from the reconstructed mesh, and consider reporting direct margin line prediction accuracy. (2) Although the input and output of the high frequency morphology refinement DiT can be inferred from the overall pipeline, the description in Section 3.1.2 is too brief and not very easy to follow. Please add a few equations, which may be similar to those in the first stage, to make the process clearer, e.g., M^{(2)}=\mathrm{MC}({\mathrm{p}|D_{\psi }(\mathrm{p},\ \mathrm{DiT}(a_{i}))=0}). This is how I understand the process, please correct me if I misunderstood. At least, please provide a detailed explanation in the supplementary material. (3) Some recent dental crown generation methods, such as DentalGEN[1], are not discussed in the paper. Please include these studies in the Related Work to provide a more comprehensive overview of recent developments in this field. [1] Wu J, Liu Y, Chen Y, et al. DentalGEN: a large-scale controllable generative AI framework for automated dental crown restoration[J]. npj Digital Medicine, 2026. Minors: (1) The dataset covers multiple tooth positions, but the current visualizations are relatively limited. Additional examples across different tooth positions would help demonstrate the generalization of the method. (2) Although it is not explicitly mentioned in the paper, the visualizations seem to suggest that the input includes the prepared tooth geometry and labels. If so, how well does the tissue surface of the generated crown fit the preparation surface? Clarifying whether preparation surface clearance is considered would improve the clinical interpretation of the results. (3) Figure 3 contains separate descriptions for its subfigures but lacks a unified figure caption.

Questions for authors

N

1 comment
AuthorsOpenReview ↗

Response to Reviewer 7d6j

Response to Reviewer 7d6j

R3-1 · Major 1 · Margin line prediction

At inference, the first-stage DiT jointly generates a crown shape latent and a 256-point preparation margin from the scan-derived context. It learns the relationship between crown geometry and the preparation margin through joint prediction and a dedicated margin loss. Mesh reconstruction uses the crown shape latent, and refinement uses spatial anchors derived from the reconstructed mesh.

The MLD reported in the paper measures the distance from the reference margin to the generated crown surface. We also evaluated the predicted curve directly on the validation set. Its mean symmetric distance to the reference margin was 0.21 ± 0.28 mm (median 0.15 mm). We will add this result to the revised manuscript.

R3-2 · Major 2 · Morphology refinement

We added an equation in Section 3.1.2 to clarify the morphology refinement stage.

R3-3 · Major 3 · DentalGEN

Thank you for the suggestion. We have added DentalGEN (Wu et al., 2026) to the first paragraph of Section 2, discussing its controllable multi-view diffusion framework for joint crown geometry and texture synthesis and its incorporation of clinical constraints.

R3-4 · Minor 1 · Visualizations across tooth positions

We thank the reviewer for pointing this out, and agree that the original gallery did not sufficiently cover different tooth positions. We have rebuilt the gallery with crowns from all seven tooth positions (1–7, central incisor to second molar), so that results across positions can be viewed side by side. The new gallery is in Additional Experiments (Fig. 11 in the appendix).

R3-5 · Minor 2 · Internal Fit and preparation clearance

CrownGen uses the intraoral scan and a target tooth ID at inference, with tooth labels obtained by automatic segmentation. The functional rewards address proximal contact and occlusal clearance. To assess preparation fit, we evaluated the intaglio surfaces of the technician-designed reference crowns (GT), CrownGen, and its ablated variants. We evaluated the unprocessed crowns on 100 cases using area-uniform surface sampling, excluding regions within 0.5 mm of the finish line. We report unsigned intaglio–preparation distances, and local distance-map errors relative to GT. GT incorporates undercut relief, which may contribute to spatial distance variation.

MethodMean unsigned distance (μm)Within-case distance SD (μm)Distance-map MAE vs. GT (μm)
GT82.7 ± 17.524.6 ± 12.20.0 ± 0.0
Stage 1160.1 ± 34.063.3 ± 13.185.0 ± 19.8
Ours w/oRL100.7 ± 38.594.2 ± 78.266.3 ± 9.3
Ours67.3 ± 15.332.7 ± 10.227.4 ± 5.3

Values are case means ± between-case SD. The unsigned distances quantify geometric proximity. Distinguishing clearance from overlap requires a signed assessment.

R3-6 · Minor 3 · Figure 3 caption

Thank you for the suggestion. We revised Figure 3 to use a single caption with descriptions of both panels.

References

  • Wu, J., Liu, Y., Chen, Y., Zhang, Y., Gu, J., Liu, Y., Li, M., Wang, W., Fu, Y., Cui, Z., Tsoi, J.K.H., Shen, D. DentalGEN: a large-scale controllable generative AI framework for automated dental crown restoration. npj Digital Medicine (2026). https://doi.org/10.1038/s41746-026-03236-7
Reviewer_GDNyOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper describes an approach to generate the 3D geometry of dental crowns from intra-oral 3D scans. The method operates in two stages. First, a latent diffusion model, a finetuned version of Hunyuan3D, is used to generate a first instance of a crown geometry. A voxel-based detail refinement stage is then applied to further add finer structure of typical anatomical tooth detail. An RL based second stage with specifically designed rewards is then applied to make sure that the created dental geometries are functionally adequate and respect geometric / contact surface constraints in relation to neighboring teeth and teeth on the opposite side of the contact surface. The authors argue that this reward-based second stage is needed due to the challenge in formulation the respective constraints in a differentiable way. The paper reports quantitative results on a large scan database and shows that it is able to outperform related approaches from the literature. An ablation study is also provided and a user study shows the favorable results of the method compared to baseline approaches.

Critical strengths

-The paper proposes a new solution to a challenging constrained geometric design problem with some new ideas on how to ensure functional (crown) constraints within the framework of a learned / finetuned generative model

-Result and evaluations show an improvement over related approaches, both in terms of quantitative criteria, but also in terms of usability / user study evaluations with professional from the dental prosthetics field

Critical weaknesses

-Novelty: The paper does not introduce groundbreaking new concepts and relies on an orchestration of existing concepts ideas. The method is based on Hunyuna3D, and also the voxel-based geometric refinement stage bears similarities to other works, e.g. the TRELLIS model.

-Related work: The paper misses references from the geometric modeling and computer graphics literature on parametric dental modeling, such as Buchaillard et al., 3D Statistical models for tooth surface reconstruction, or Velinov et al., Appearance Capture and Modeling of Human Teeth. Some of these earlier works also aim to reconstruct detailed tooth appearance which is also an important angle for tooth replacement not addressed in this work. Would be good to dig through the related work of these earlier papers again.

Further, the paper misses one very important reference, Zhang et al., An Implicit Parametric Morphable Dental Model, ACM TOG 2022, which proposes a joint / coupled implicit parametric model of the gum and whole tooth row with a model for each individual tooth. Based on such a coupled model, also local edits / fill-ins in the tooth row are possible while respecting geometric constraints of the entire row of teeth. Explicit crown reconstruction with fitting to opposing teeth surfaces was not investigated in the paper, but this is still a very relevant baseline, see also later comments.

-Design / claims: The paper proposes a computationally comparably heavy method for essentially a parametric shape completion task in a very specialized and constrained domain. The results are good but it is not obvious that similar results could not have been achieved also with, for instance, an SDF based approach (like Zhang et al. 2022) for initialization, followed by an optimization of geometric constraints based on the SDF representation.

Connected to that, it is not clear to me if the claim that RL is inevitable in the presented form due to the non-differentiability of the functional constraints holds. Based on an implicit (e.g. SDF) representation one should be able to describe distance-based geometric constraints in a differentiable way, and one could use this for fitting also of parametric morphable models designed in the more traditional (yet also more compact / efficient) way.

I am not saying the results of the presented approach are bad, but a discussion on the above design considerations / or even evaluation of such a simpler alternative, would be nice.

-Evaluation: maybe I overlooked it but would be good to add penetration results for all methods tested, and maybe also add visual comparisons to the results with the tested alternative methods to the video.

Questions for authors

some questions appear above

3 comments
AuthorsOpenReview ↗

Response to Reviewer GDNy (1/2)

Response to Reviewer GDNy (1/2)

R2-1 · Novelty

Thank you for this assessment. CrownGen builds on Hunyuan3D, and its use of voxel-based spatial structure is related to prior methods such as TRELLIS. Our contribution is the modeling and training strategy for functional crown generation.

The central question we address is how to adapt a pretrained geometric prior to patient-specific boundary and contact requirements. Our joint crown–margin formulation incorporates boundary supervision into shape generation, while functional post-training directly targets proximal and occlusal relationships. These choices have measurable effects with the same backbone and preprocessing: removing joint margin modeling increases the reported MLD from 0.012 to 0.062 (Table 2). Moreover, morphology refinement improves contact-region overlap without improving contact-centroid localization. Adding reward-based post-training increases contact IoU from 0.290 to 0.565 and reduces centroid error from 1.085 to 0.296 mm (Table 3).

These results support the complementary roles of anatomical modeling and functional learning, and demonstrate the value of the proposed adaptation beyond the pretrained backbone alone. We will revise the contribution statements to distinguish the inherited components from our specific design choices and their experimentally supported benefits.

R2-2 · Related work

Thank you for highlighting these relevant works. We have expanded the first paragraph of Section 2 to cite and discuss Buchaillard et al. (2007), Velinov et al. (2018), and Zhang et al. (2022). In particular, we now describe Zhang et al.’s coupled implicit representation of teeth and gums and its support for reconstruction and individual tooth replacement, acknowledging its relevance to dental modeling and editing.

AuthorsOpenReview ↗

Response to Reviewer GDNy (2/2)

R2-3 · Design and claims

We thank the reviewer for suggesting SDF-based optimization. A differentiable SDF objective is a viable approach for distance-based constraints. Our motivation for reward-guided post-training is to use the functional scores produced by our evaluator directly and to learn functional adaptation in the generator across cases.

In an earlier experiment, we fine-tuned the generator directly with a constraint-specific loss based on ray-cast distances from selected query points to the crown surface. The fine-tuned crowns showed local dents or bumps in the highlighted regions of Fig. 9 in the appendix. Our reward-guided approach evaluates complete sampled crown meshes and uses their functional scores to weight the post-training objective. Figure 9 illustrates the difference for the two strategies we tested. We will clarify this motivation and discuss SDF-based geometric optimization in the revised manuscript.

R2-4 · Penetration results and visual comparisons

Inspired by the intersection evaluation in CrownFusion (Ye et al., 2026), we implemented a crown–antagonist penetration evaluation and applied it uniformly to all compared methodsusing the same 100 cases randomly sampled from our test set. The table reports results from our own implementation on the shared cases. CrownFusion’s published results were obtained on a different test set under its own evaluation protocol, and a runnable CrownFusion model was unavailable to us for evaluation on these cases. We report the percentage of cases with penetration and the mean penetration area across all 100 cases, assigning zero to nonpenetrating cases. Measurements use signed 3D surface distances at the original seating pose, with negative values indicating penetration.

MethodCases with penetration (%)Mean penetration area (mm²)
DCrownFormer74.01.6404
DMC63.03.3846
VBCD22.00.1561
Stage 118.00.3029
Ours w/oRL16.00.1825
Ours14.00.0839
GT (reference)15.00.6257

These metrics quantify crown–antagonist penetration and complement the functional distance errors in Table 1 and the expert assessments in Section 5.1. For proximal contact, Table 1 reports deviations from the 0 mm target.

We will add side-by-side visual comparisons with the tested baselines to the supplementary video before the discussion period closes.

References

  • Buchaillard, S.I., Ong, S.H., Payan, Y., Foong, K. 3D statistical models for tooth surface reconstruction. Computers in Biology and Medicine, 37(10), 1461–1471 (2007). https://doi.org/10.1016/j.compbiomed.2007.01.003
  • Velinov, Z., Papas, M., Bradley, D., Gotardo, P., Mirdehghan, P., Marschner, S., Novák, J., Beeler, T. Appearance capture and modeling of human teeth. ACM Transactions on Graphics, 37(6) (2018). https://doi.org/10.1145/3272127.3275098
  • Ye, J.Z., Yan, X., Hauberg, S., Chang, A.X., Zhang, H., Søndergaard, P.L. CrownFusion: 3D Dental Crown Generation Using Geometry Images and Latent Diffusion. In: MICCAI 2026, LNCS 16889, pp. 464–474. Springer, Cham (2027). https://doi.org/10.1007/978-3-032-38186-6_44
  • Zhang, C., Elgharib, M., Fox, G., Gu, M., Theobalt, C., Wang, W. An Implicit Parametric Morphable Dental Model. ACM Transactions on Graphics, 41(6) (2022). https://doi.org/10.1145/3550454.3555469
AuthorsOpenReview ↗

R2-4 Update · Supplementary Video

Following up on R2-4 visual comparisons, we have uploaded the updated supplementary video. The final segment shows CrownGen, DCrownFormer, DMC, VBCD, and the human-designed reference crowns (GT) side by side.

Reviewer_2AYZOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper provides a framework for generating single crowns from raw intraoral scans.

Critical strengths

The paper generally appears sound, and the results look good. In particular, the RL part appears quite novel.

Critical weaknesses
  1. The caption of Figure 1 states: "We present CrownGen, the first fully automatic framework for single-crown generation from oral scans that produces functionally plausible crown geometries without manual intervention." As far as I can tell, this claim is never backed up. I would want concrete evidence to support why previous methods fail this criterion, likely from dental technicians at a variety of schools.
  2. The overall idea pipeline mirrors Crownfusion: https://link.springer.com/chapter/10.1007/978-3-032-38186-6_44 The current paper should discuss differences clearly.
  3. While the presented metrics seem strong, they would have been more convincing if the authors had tried to incorporate recent approaches: https://papers.miccai.org/miccai-2026/0402-Paper1259.html
  4. The paper provides no information about its data sources. This is a critical omission.
  5. I severely doubt the user study, because in the field everyone knows that you cannot get clinicians to agree on what is the best crown. Getting a score of 4.85 out of 5 is unheard of on average, and suggests either cherry-picking or strong biases in the selection of users.
Questions for authors

Please respond to weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 2AYZ (1/2)

Response to Reviewer 2AYZ (1/2)

R1-1 · Scope of the automation claim

Thank you for pointing out this overstatement. We acknowledge that prior work, including CrownFusion, has already demonstrated automated crown generation. Our intended emphasis is an integrated scan-to-design workflow that automatically performs tooth segmentation and pose alignment and generates crown geometry without requiring a pre-annotated or separately supplied margin line. Beyond automation, our geometric and functional evaluations, together with expert assessments, support the suitability of the generated models as usable digital crown designs for subsequent clinician-led evaluation and refinement. This is the design-stage quality and practical utility we intended to convey, with final clinical suitability assessed through downstream professional review and any necessary adjustments. We will remove “first” from the caption and revise the wording to clarify both the scope of automation and the intended use of the generated designs.

R1-2 · Discussion of CrownFusion

We thank the reviewer for pointing out this work and will discuss CrownFusion in the revised Related Work. We did not have a runnable CrownFusion implementation for evaluation on our cases. Both methods condition a latent generator on the surrounding dentition, but they differ in two key aspects. First, CrownFusion flattens the crown surface into a 2D geometry image to reuse image generators, whereas we generate the crown natively in 3D, in the same frame as the oral scan. This avoids compressing the axial walls where proximal contacts lie, allows gaps and penetrations with neighboring and opposing teeth to be measured directly on every generated sample, and lets a single model cover all tooth positions.

Second, CrownFusion uses occlusal fit only for evaluation, and its generator learns to imitate technician-designed crowns, which alone does not ensure correct contacts. Our reward-guided RL stage scores generated crowns by their proximal and occlusal relations and adapts the generator accordingly, making functional fit a training objective.

R1-3 · CrownBench evaluation

We thank the reviewer for suggesting CrownBench. We used its official evaluation code on cases from our dataset with both adjacent teeth present and used the technician-designed crowns as the reference. CrownGen results are reported as mean ± SD across five sampling seeds. The first four rows reproduce values reported by Wu et al. in Table 1 on their CrownBench test set. The last two rows are results on our dataset. Comparisons between studies are for context because the test sets differ.

Metrics. OFR (Occlusal Feasibility Rate) is the fraction of cases whose minimum signed distance to the antagonist lies within [−0.2, 0.5] mm. PFR (Proximal Feasibility Rate) is the fraction of cases where both mesial and distal sides contact the adjacent teeth with at most 0.3 mm penetration. MSI (Marginal Seal Integrity) is the maximum distance from the margin line to the crown surface. COC (Contextual Occlusal Consistency) measures how closely the crown's occlusal contact ratio matches that of the adjacent teeth, as exp(−|Δ|). ADF (Anatomical Detail Fidelity) is the mean cosine similarity between crown surface normals and those of the technician design. OCD (Occlusal Contact Distribution) counts distinct occlusal contact regions and can be interpreted relative to the corresponding technician designs.

MethodOFR ↑PFR ↑MSI (mm) ↓COC ↑OCDADF ↑
DMC0.520.160.830.801.960.87
3DShape2VecSet0.780.620.620.812.590.95
VBCD0.820.760.630.823.960.95
Human design (CrownBench reference)0.920.900.000.833.121.00
Technician design (our dataset reference)0.940.920.000.873.361.00
CrownGen0.86 ± 0.0150.78 ± 0.0130.46 ± 0.0070.88 ± 0.0053.10 ± 0.1390.97 ± 0.001

We will add this evaluation to the comparison section (Section 4.2) of the revised paper.

AuthorsOpenReview ↗

Response to Reviewer 2AYZ (2/2)

Response to Reviewer 2AYZ (2/2)

R1-4 · Dataset source

The dataset was obtained from a single industry-operated dental CAD/CAM cloud platform under a research collaboration agreement permitting model training and academic publication. It comprises historical digital crown design records spanning approximately five years and associated with 148 client organizations.

The crowns were designed by professional dental design teams from two companies using exocad software and underwent quality inspection before delivery. Direct patient-identifying information was removed before the dataset was provided to the research team.

R1-5 · User study

Experts rated the anatomical plausibility of each candidate crown individually. In the CAD environment, they could open and rotate the anonymized crown meshes and inspect them alongside the corresponding upper and lower jaw scans. The original instructions asked experts to consider crown shape, contour, convexity, fit along the preparation margin, and marginal ridges when assessing how well each design matched the patient’s anatomy. Each design was scored on a 1–5 scale, with 5 being the highest score.

We used a script to randomly select four cases per tooth position from the 1,000-case test set, yielding 64 cases across 16 posterior tooth positions. Case selection was fully automated. With help from colleagues in medicine, we recruited 32 evaluators, including 26 practicing dentists and six senior dental technicians. None participated in model development or training data preparation. Two experts independently rated each case, and we included all selected cases and recorded ratings in the analysis.

The mean of the 128 anatomical plausibility ratings for CrownGen was 4.85 (SD = 0.49). The counts for scores of 1 through 5 were 0, 1, 4, 8, and 115, respectively. Both experts gave a score of 5 in 51 of the 64 cases. Their scores differed by one point in eight cases and by at least two points in five cases. Appendix A.4 of the revised manuscript now includes the rating questions and describes the case sampling procedure, evaluator composition, and CAD-based evaluation procedure.

We are preparing a larger follow-up user study on the anatomical plausibility of crowns generated by CrownGen. The study will involve a new group of dental professionals who did not participate in the original evaluation, and we expect to complete it in about one month.

References

  • Wu, J., Huang, W., Gu, J., Hao, J., Fu, Y., Wang, Y., Lai, P., Tsoi, J.K.H. From Geometry to Clinic: A Physics-Aware Evaluation Framework and Benchmark for Automated Dental Crown Design. In: MICCAI 2026, pp. 200–209. Springer, Cham (2027). https://doi.org/10.1007/978-3-032-38236-8_20
  • Ye, J.Z., Yan, X., Hauberg, S., Chang, A.X., Zhang, H., Søndergaard, P.L. CrownFusion: 3D Dental Crown Generation Using Geometry Images and Latent Diffusion. In: MICCAI 2026, LNCS 16889, pp. 464–474. Springer, Cham (2027). https://doi.org/10.1007/978-3-032-38186-6_44

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.