Digital smile design and AI smile design: 3D cephalometry, injection moulding and measurement

1. 3D cephalometry: a measured face instead of a stock template

In conventional workflows, anterior tooth forms are often picked by eye from a generic library template. Yet across a century of dentofacial morphology literature (Williams 1914; Frush & Fisher 1958 SPA factor; Lombardi 1973; Paolucci et al. 2012), the ratio between bizygomatic facial width (ZY-ZY) and bigonial mandibular width (Go-Go) directly governs line angles, incisal curvature and tooth silhouette.

• 478-point 3D face mesh: From 3D landmarks extracted via MediaPipe FaceMesh, bizygomatic width (#234–#454) and bigonial width (#172–#397) are measured and classified against anthropometric thresholds (Farkas 1994) via the Go-Go / ZY-ZY ratio.
• Incisor aspect ratio and 441-form library: Upper central incisor width-to-height ratio is segmented via SAM 2D and combined with facial class, lip line and SPA parameters to select candidate tooth morphologies from a 441-form library ready for ExoCAD design.

2. Facial shape classification and locked n = 32 test set measurement

Zero-shot vision-language models systematically misclassify broad/square-jawed (brachycephalic, Go-Go / ZY-ZY ≥ 0.82) faces as “oval” and default to rounded ovoid teeth. To eliminate this bias, rules distilled from 113 peer-reviewed papers and 3D cephalometric measurements were used to fine-tune Qwen2.5-3B-Instruct-4bit on Apple Silicon (MLX QLoRA) and evaluate it on a locked n = 32 test set never seen during training:

The full experiment on a 100% synthetic brachycephalic face (n = 1, Go-Go / ZY-ZY = 0.837) and the 4-arm ablation study, complete with an interactive 3D wireframe and morphology slider, are available in the technical article: From Face to Tooth: How a 3D-Cephalometry QLoRA Model Beat the Untrained Model (n = 32 + 4-Arm Ablation) →

3. AI visual preview = less tooth preparation thesis (conservative dentistry)

The primary clinical value of AI smile design preview is not marketing, but conservative dentistry. Running a visual simulation on a de-identified mouth crop inside a Constrained Canvas (where midline, incisal curve, gingival zeniths and lip boundaries are locked) produces two direct clinical outcomes:

• Less tooth reduction (minimally invasive boundary): Both dentist and patient evaluate the target morphology before a bur ever touches enamel. Tooth preparation is a last resort; cases that can be resolved via minor additive bonding, orthodontic alignment or prepless veneers avoid unnecessary enamel loss.
• Resolving “natural vs bold” expectations upfront: In clinical practice, aesthetic preference splits along a clear axis: patients asking for teeth “natural like my own” versus those seeking a “bold, bright look.” AI preview followed by an intraoral trial (mock-up) settles this choice before irreversible fabrication; final verification happens in the mouth, not on the screen.

4. Injection moulding, chairside veneers and the material axis

Digital smile design is not restricted to ceramic milling; the 3D digital wax-up guides whichever restorative pathway preserves the most tissue while meeting biomechanical load:

• Injection moulding and composite veneers: Using a transparent silicone index printed from the 3D digital design to deliver non-prep or minimal-prep flowable composite injection moulding and direct composite veneers in a single visit.
• Same-day chairside CAD/CAM: Intraoral scanning, digital design and chairside milling (CEREC / in-house CAD-CAM) enabling preparation and final delivery of single-unit and short-span ceramic veneers within the same day.
• Purpose + span + bar-support material axis: Material selection does not follow a rigid “anterior vs posterior” rule of thumb. Glass ceramics (lithium disilicate / feldspathic) or composite/moulding serve single-unit aesthetic indications, whereas multi-unit bridges require multi-layered zirconia and full-arch implant prostheses rely on titanium-bar-supported zirconia (or fiber-bar support where indicated).

For the full synthesis testing these 20 clinical/sectoral verdicts against 11 independent voices and 100+ verified primary sources, see: Digital Dentistry: What Is and What Should Be — The Equation in Turkey and the World (20 Field Verdicts & Voice × Thesis Matrix) →

5. Frequently asked technical questions (FAQ)

How is digital smile design done with 3D cephalometry?

Facial proportions (478-point 3D cephalometric mesh, bizygomatic ZY-ZY and bigonial Go-Go width ratio), lip dynamics and dental arch geometry are measured numerically to select candidate forms from a 441-form library and generate a constrained AI visual preview before any tooth reduction begins.

Does AI smile design preview reduce tooth preparation?

Yes. Evaluating the target tooth morphology inside a Constrained Canvas (locked midline, incisal curve, gingival zenith and lip boundaries) followed by an intraoral mock-up settles aesthetic expectations upfront so cases treatable via additive bonding, injection moulding or prepless veneers avoid unnecessary enamel preparation.

How did the 3D cephalometric QLoRA model perform on the locked n = 32 test set?

On the locked n = 32 test set never seen during training, fine-tuning Qwen2.5-3B-Instruct-4bit via MLX QLoRA raised facial shape threshold classification accuracy from 0.0% (0/32) to 90.6% (29/32), while mandibular width, target tooth line and valid SPA JSON reached 100.0% (32/32).

6. Verification records and internal links

• Industry article (Spark verdict · EN / TR / AR): Digital Dentistry: What Is and What Should Be — The Equation in Turkey and the World · TR · AR
• 3D Tooth Library & Rendering Pipeline (441 tooth forms · 476/477 STL/PLY renders): STL/PLY 3D Tooth Library, 478-Landmark Cephalometry, Constrained Canvas & ExoCAD Handoff (TR · AR)
• Experimental Spark (3D cephalometry & QLoRA · n = 32): From Face to Tooth: How a 3D-Cephalometry QLoRA Model Beat the Untrained Model
• Synaptic Harvest (N = 96 experiments): Who Trains Whom? The Mount Changes, the Law Remains
• Multi-agent architecture & Spark protocol: Autonomy & live measurements · Spark (Kıvılcım) 7-step hypothesis protocol · Applied Spark & LLM Red Team (420 Citations) · Identity Freeze & Three-Family Architecture · QAJAN 6-Gate Hypothesis Factory
• Reaper citation verification method and live scorecard (360/394): reaper-en.html#karne · data/reaper_karne.json
• Machine-readable measurement records and Maya 20-question debate: /bot/ (16 measurement records) · /bot/maya/ (20 Q&A)
• About the author: Yusuf Aziz — Dental CAD/CAM & digital morphology

Falsification test: If cases planned with Constrained Canvas + 3D cephalometric preview show no meaningful reduction in enamel preparation depth or remake rate compared to unpreviewed workflows, or if 3D QLoRA threshold accuracy drops below 80% on an independent set of n ≥ 100 faces, the method claims on this page are revised.