AI smile design: 478-point 3D cephalometry, constrained canvas preview and 441-form tooth selection

1. Constrained Canvas and deterministic 3D library vs unconstrained diffusion filters

Conventional generative smile filters distort lip contours to hallucinate non-manufacturable teeth whose 2D pixels have no 3D counterpart in CAD/CAM software (ExoCAD, 3Shape). AtlasMaster resolves both flaws through a Constrained Canvas coupled to a deterministic 3D morphology pool:

• 4,686-pixel lip lock on de-identified mouth crops: Full facial identity is never ingested into the preview canvas; only a de-identified mouth crop is processed with a 4,686-pixel inner lip lock keeping contour deviation within p95 ≤ 0.2 mm.
• Locked anatomical guide rails: Facial midline, incisal edge curvature, golden-proportion corridor, and gingival zenith points are geometrically locked before rendering.
• 441 tooth forms and 476/477 STL/PLY renders: Every silhouette shown in the 2D preview maps one-to-one to a real 3D mesh inside our 441-form tooth library (476/477 verified STL/PLY renders), enabling laboratories in Sakarya (Serdivan, Adapazarı) and nationwide to move directly to 3D ExoCAD fabrication.

2. From 2D photo to 3D cephalometric morphology selection: locked n = 32 test set

MediaPipe FaceMesh extracts 478 3D landmarks to compute bizygomatic width (ZY-ZY, #234–#454) and bigonial mandibular width (Go-Go, #172–#397). Trained on rules distilled from 113 peer-reviewed papers and evaluated via a 4-arm ablation study, the Qwen2.5-3B-Instruct-4bit (MLX QLoRA, 1,000 iterations) model achieved the following verified results on the locked n = 32 test set:

Model output is a computational morphology recommendation and rationale, not a clinical decision; all metrics derive from a 100% synthetic/locked test set, and clinical judgment belongs solely to the dentist. Following our zero-concealment rule for negative results, our 800-step v6 DoRA experiment’s 35% empty-output rate and 43.2–46.3 s latency boundary were published openly while keeping the deterministic 3D cephalometric QLoRA layer in production (Full Experiment & Ablation Record →).

3. Joint decision with the dentist, intraoral mock-up and conservative design

AI smile design preview translates aesthetic expectations into a measurable 2D/3D engineering blueprint before a bur touches the tooth:

• Visual and physical preview first (mock-up): The Constrained Canvas preview is translated into a 3D-printed model and silicone index for an intraoral temporary trial (mock-up), enabling a shared decision with the dentist.
• Numerical guidance for minimal preparation: Knowing target volume upfront supports additive composite injection (injection moulding), prepless veneers, or conservative ceramic designs while protecting enamel.
• Context-blind Reaper verification: Across our synthesis, 360/394 external citations and 26/26 author sources (386/420) were verified against Crossref, PubMed, and DOI records by Reaper, isolating 25 fabricated citations (Reaper scorecard).

4. Frequently asked questions (AI smile design)

How does AI smile design differ from unconstrained photo filters?

Unconstrained diffusion filters alter lip contours and tooth count arbitrarily. AtlasMaster uses a Constrained Canvas: on a de-identified mouth crop, a 4,686-pixel lip lock keeps contour deviation at p95 ≤ 0.2 mm, locks the midline and gingival zeniths, and previews only candidate forms mapped to a real 441-form STL/PLY 3D tooth library (476/477 renders).

How does the AI select tooth morphology from facial shape, and what is its measured accuracy?

Using 478 3D cephalometric landmarks from MediaPipe FaceMesh, the bizygomatic (ZY-ZY) to bigonial (Go-Go) ratio is computed. Fine-tuned on rules from 113 peer-reviewed papers, the Qwen2.5-3B-Instruct-4bit MLX QLoRA (1,000 iterations) model raised facial shape accuracy on the locked n = 32 test set from 0.0% (0/32) to 90.6% (29/32), while mandibular width, target tooth line, and SPA JSON reached 100.0% (32/32).

How is an AI photo preview transferred into ExoCAD and laboratory production?

Every tooth silhouette previewed in 2D corresponds directly to a 3D STL/PLY mesh inside the 441-form library. Once the dentist and individual agree on a candidate form, the morphology ID is exported into the ExoCAD or chairside CAD/CAM package for mock-up printing, silicone-index composite injection moulding, or ceramic milling.

Does this page include real patient data, clinical promises, or pricing?

No. All experimental metrics derive from the locked n = 32 test set and a 100% synthetic face, and 360/394 external citations plus 26/26 author sources (386/420) are verified by Reaper. This page describes a digital design and visual preview service only, with zero pricing or clinical promises; clinical decisions belong solely to the dentist.