How does 478-landmark cephalometric smile analysis work in digital smile design?

Using MediaPipe FaceMesh, 478 3D anatomical coordinates [x, y, z] are extracted from the face. The ratio of bigonial mandibular width (Go-Go, #172-#397) to bizygomatic cheekbone width (ZY-ZY, #234-#454) is computed in 3D space and combined with SAM 2D central incisor segmentation to rank candidate morphologies from a 441-form tooth library.

Why do untrained vision-language models misread square jaws as oval, and how did QLoRA fix it?

Due to training prior bias, untrained vision-language models label brachycephalic (square/broad mandible, Go-Go / ZY-ZY >= 0.82) faces as oval (0/32 accuracy on the locked n = 32 set: 18 oval, 14 uncertain). Fine-tuning Qwen2.5-3B-Instruct-4bit via MLX QLoRA for 1,000 iterations on rules distilled from 113 peer-reviewed papers across 576 cases (519 train + 57 validation) raised facial shape accuracy on the locked n = 32 set to 90.6% (29/32) and jaw_width / target tooth line accuracy to 100.0% (32/32).

Are any real patient photographs used in the 3D cephalometric benchmark?

No. The brachycephalic showcase face (n = 1, seed = 42) is 100% synthetic. A three-tier similarity audit (lookalike-tarama.json) verified 0 Google Lens web exact matches, 0 local perceptual hash matches (dHash <= 10 across 1,515 images), and 0 exact matches across the 576-case cephalometric dataset (0 / 0 / 0).

Does 3D cephalometric smile analysis replace clinical diagnosis?

No. The algorithmic output is not a medical or radiographic diagnosis; it provides anthropometric pre-screening across 441 tooth forms for the dentist and laboratory. Clinical judgment belongs solely to the dentist.

✓ Verified by Reaper · O-008 & O-009 · 21 Oct 2026

3D cephalometric smile analysis: 478-landmark face mesh for digital smile design

Q: Does 3D cephalometric smile analysis replace clinical diagnosis?

No. The algorithmic output is not a medical or radiographic diagnosis; it provides anthropometric pre-screening across 441 tooth forms for the dentist and laboratory. Clinical judgment belongs solely to the dentist.

Written by: Atlas (Gemini 4 Argon) | Published by: Yusuf Aziz · Atlas Studio | Reaper: October 21, 2026, 0 errors

For dentists and digital dental laboratories, cephalometric smile analysis replaces subjective template browsing with 3D geometric measurement of facial proportions. Developed within the Sakarya-based AtlasMaster R&D pipeline led by Yusuf, the 478-landmark 3D face mesh workflow combines MediaPipe FaceMesh coordinates, anthropometric rules distilled from 113 peer-reviewed papers, and a 1,000-iteration Qwen2.5-3B-Instruct-4bit QLoRA model—raising facial shape accuracy on the locked n = 32 test set from 0.0% (0/32) to 90.6% (29/32) and target tooth line accuracy to 100.0% (32/32).

Scope and privacy boundary (RG 33075 & KVKK): This page summarizes the methodology documented in From Face to Tooth: 3D Cephalometric QLoRA. It contains no real patient photographs, no clinical promises, and no pricing; the synthetic showcase face passed a 0 / 0 / 0 real-person similarity audit.

1. How the 478-landmark 3D face mesh and cephalometric ratios are measured

On a single 2D photograph, slight head pitch or focal-length distortion can make a broad mandible appear narrow. Three-dimensional cephalometric smile analysis resolves perspective distortion through four computational layers:

478-landmark 3D coordinate mesh (landmarks-3d.json): MediaPipe FaceMesh extracts 478 anatomical landmarks with [x, y, z] coordinates. Bizygomatic cheekbone width (ZY-ZY, landmarks #234–#454) and bigonial mandibular width (Go-Go, landmarks #172–#397) are computed in 3D Euclidean space.
Anthropometric thresholds from 113 peer-reviewed papers: Rules synthesized from 113 publications—including Farkas (1994) craniofacial indices, Williams (1914) geometric classification, Frush & Fisher (1958) SPA morphology, and Lombardi (1973) visual proportion principles—separate brachycephalic (square/broad jaw, Go-Go / ZY-ZY ≥ 0.82), mesocephalic, and dolichocephalic archetypes.
Synthetic square-face (n = 1) reference measurement: On the 100% synthetic brachycephalic reference face (seed = 42), the 3D ratio is measured at Go-Go / ZY-ZY = 0.837 (with interactive module indices jaw_ratio = 0.9236, face_ratio = 0.8892), yielding Jaw Width: wide, Target Line: slightly curved, and SPA: masculine — square/angular central incisors.
SAM 2D incisor segmentation and 441-form matching: Central incisor width-to-height ratio is segmented via SAM 2D to rank 5 priority and 3 alternative tooth forms from the 441-form library (476/477 verified STL/PLY renders).

2. Untrained VLM vs 3D cephalometric QLoRA: locked `n = 32` test set results

Trained across 576 cases (519 training + 57 validation) over 1,000 iterations on 16 transformer layers, Qwen2.5-3B-Instruct-4bit reduced training loss from 2.728 to 0.136 and validation loss from 2.718 to 0.213. Performance on the unseen locked n = 32 test set (ornek/sentetik-kare/n32-kilitli-test.json) is shown below:

Cephalometric Task / Metric	Untrained Model	1,000-Step QLoRA	Primary Evidence File
Facial shape (`face_shape`) accuracy (`K008`)	0.0% (0/32 · 18 oval, 14 uncertain)	90.6% (29/32)	bot/olgular.json (O-008)
Mandibular width (`jaw_width`) (`K009`)	0.0% (0/32)	100.0% (32/32)	bot/olgular.json (O-009)
Target arch line (`target_line`) (`K009`)	0.0% (0/32)	100.0% (32/32)	n32-kilitli-test.json
SPA morphology & valid JSON (`K009`)	3.1% (1/32)	100.0% (32/32)	face-to-tooth-cephalometric-lora.html
Train / validation loss (`K053`)	2.728 / 2.718 (initial)	0.136 / 0.213 (step 1,000)	face-to-tooth-cephalometric-lora.html
Real-person similarity audit (`K054`)	3-tier verification	0 / 0 / 0 matches	lookalike-tarama.json

3. The `0 / 0 / 0` similarity audit and de-identified processing

Personal data privacy in cephalometric smile analysis is enforced through three deterministic safeguards:

Three-tier similarity scan (0 / 0 / 0): On the synthetic square face (n = 1), Google Lens exact web matches measured 0, perceptual hash matches (dHash ≤ 10) across 1,515 local images measured 0, and exact landmark matches across the 576-case cephalometric dataset measured 0.
De-identified mouth crop: Full facial photographs are never sent to external image generators; only a de-identified mouth crop bounded by a 4,686-pixel lip lock is processed.
Transparent negative-result logging: During experimentation, the v6 MLX DoRA (gemma-4-E4B) branch produced 35% empty outputs and was retired in favor of the verified 3D cephalometric QLoRA pipeline.

4. Frequently asked questions (Cephalometric smile analysis & 3D face mesh)

How does 478-landmark cephalometric smile analysis work in digital smile design?: Using MediaPipe FaceMesh, 478 3D anatomical coordinates [x, y, z] are extracted from the face. The ratio of bigonial mandibular width (Go-Go, #172–#397) to bizygomatic cheekbone width (ZY-ZY, #234–#454) is computed in 3D space and combined with SAM 2D central incisor segmentation to rank candidate morphologies from a 441-form tooth library.
Why do untrained vision-language models misread square jaws as oval, and how did QLoRA fix it?: Due to training prior bias, untrained vision-language models label brachycephalic (square/broad mandible, Go-Go / ZY-ZY ≥ 0.82) faces as oval (0/32 accuracy on the locked n = 32 set: 18 oval, 14 uncertain). Fine-tuning Qwen2.5-3B-Instruct-4bit via MLX QLoRA for 1,000 iterations on rules distilled from 113 peer-reviewed papers across 576 cases (519 train + 57 validation) raised facial shape accuracy on the locked n = 32 set to 90.6% (29/32) and jaw_width / target tooth line accuracy to 100.0% (32/32).
Are any real patient photographs used in the 3D cephalometric benchmark?: No. The brachycephalic showcase face (n = 1, seed = 42) is 100% synthetic. A three-tier similarity audit (lookalike-tarama.json) verified 0 Google Lens web exact matches, 0 local perceptual hash matches (dHash ≤ 10 across 1,515 images), and 0 exact matches across the 576-case cephalometric dataset (0 / 0 / 0).
Does 3D cephalometric smile analysis replace clinical diagnosis?: No. The algorithmic output is not a medical or radiographic diagnosis; it provides anthropometric pre-screening across 441 tooth forms for the dentist and laboratory. Clinical judgment belongs solely to the dentist.

5. Primary sources, 3D data files, and contact

Links to the full 3D cephalometric QLoRA study, JSON coordinate sets, and Sakarya CAD/CAM collaboration:

Primary study and data: From Face to Tooth: 3D Cephalometric QLoRA · n32-kilitli-test.json · landmarks-3d.json
Related technical pages: CAD/CAM Dentistry Sakarya · Constrained Canvas Mockup · Digital Smile Design · yusuf@atlasmaster.app

Falsification gate: If 478-landmark 3D QLoRA facial shape accuracy drops below 80% on an independent set of n ≥ 100 synthetic or consented faces, or if single-frame 2D template selection matches its accuracy, this cephalometric thesis is falsified.