Constrained canvas smile preview: how photo-to-smile mockup works with no clinical claim

1. Why unconstrained generative AI fails and what the Constrained Canvas locks

Passing a full facial photograph to an unconstrained image generator alters lip vermilion borders, invents non-existent tooth counts, and produces geometries that cannot be milled in CAD/CAM. The Constrained Canvas methodology eliminates hallucinated drift through four mechanical constraints:

• De-identified mouth crop and 4,686-px lip lock (K023): Full facial identity never enters the external visual model. Only a de-identified mouth crop (excluding eyes, nose, forehead, and full facial biometrics), the selected 3D tooth layer, and a 4,686-pixel lip boundary mask are processed, keeping all 4,686 preserved lip, beard, gingival, and lower-tooth pixels identical across outputs within p95 ≤ 0.2 mm contour fidelity.
• Projected 3D STL/PLY morphology layer (K051–K052): Rather than synthesizing arbitrary teeth, the canvas projects a 3D mesh selected via 478-landmark cephalometry from the 441-form library (476/477 verified STL/PLY renders), locked at the cervical boundary (0x800) within p95 ≤ 0.2 mm (41/41 self-tests).
• Transparent retirement of failing branches (K055): During R&D, the 800-step (9.5 min, 15.93 GB RAM) v6 MLX DoRA (gemma-4-E4B) branch (which produced 35% empty outputs, 14/40, with 43.2–46.3 s latency) and the v4 unconstrained generative arm (11/24 quality failures) were retired in favor of the v7 mechanical lighting and Constrained Canvas pipeline (6/36 failures, retiring 76.5 GB of generative weights).
• Direct bridge to intraoral mock-up: Because the 2D preview is anchored to an actual 3D STL/PLY shell, the identical geometry exports to ExoCAD to print a model and silicone index for a non-prep intraoral mock-up.

2. Four-arm ablation benchmark (8/8 runs) and Constrained Canvas verification table

On the 100% synthetic brachycephalic face (n = 1, ZY-ZY = 629.7 px, Go-Go = 526.9 px, Go-Go / ZY-ZY = 0.837, N-Gn / ZY-ZY = 1.226, SAM 2D W/H = 0.77, verified at 0 / 0 / 0 real-person similarity), the 4-arm ablation experiment (ornek/sentetik-kare/ablasyon-4kol.json, 4 arms × 2 repeats = 8/8 calls, 494 s) and Constrained Canvas pipeline verified the following metrics:

3. Why “digital smile design with no clinical claim” is a core engineering rule

Both healthcare communication regulations (RG 33075) and scientific falsifiability require a clear separation between visual simulation and clinical dentistry:

• Optical and geometric simulation only: A Constrained Canvas mockup illustrates how a specific 3D tooth morphology fits within the lip frame; it does not predict biological tissue response or clinical outcomes.
• Dentist-led conservative planning: Visual alignment before tooth preparation helps dentist and individual evaluate enamel-preserving options such as silicone-index composite injection moulding or prepless laminate veneers.
• Zero online patient intake (O-012): No patient photographs or case files are accepted through this site or by email; communication is restricted to technical and professional collaboration.

4. Frequently asked questions (Constrained Canvas & photo-to-smile mockup)

What is a constrained canvas smile preview, and how does photo to smile mockup work?

Instead of feeding a full facial photograph into an unconstrained generative model, the Constrained Canvas methodology supplies only a de-identified mouth crop (excluding eyes, nose, forehead, and full facial biometrics), a projected 3D tooth morphology layer from the 441-form library, and a 4,686-pixel lip boundary lock—keeping all 4,686 preserved lip, beard, gingival, and lower-tooth pixels bit-for-bit identical within p95 ≤ 0.2 mm contour fidelity.

Why were the v4 unconstrained generative arm and v6 DoRA retired in favor of the v7 mechanical lighting and Constrained Canvas pipeline?

Applying an unconstrained generative model to retractor frames distorted tooth contours (11/24 failed checks in the v4 generative arm), while the 800-step v6 MLX DoRA (gemma-4-E4B) open-ended test produced 35% empty outputs (14/40) with 43.2–46.3 s latency. Switching to the Constrained Canvas and v7 mechanical lighting + specular + incisal shadow pipeline reduced failed checks to 6/36 and allowed 76.5 GB of generative weights to be retired.

How does the 2D Constrained Canvas preview connect to 3D ExoCAD fabrication?

The teeth previewed inside the canvas are not hallucinated pixels; they are calibrated projections from 441 unique 3D tooth forms (476/477 verified STL/PLY renders) locked at the cervical boundary (0x800 flag) with p95 ≤ 0.2 mm geometric fidelity (41/41 self-tests).

Why is this workflow defined as digital smile design with no clinical claim?

In compliance with healthcare communication regulations (RG 33075) and scientific rigor, a visual preview is strictly an optical and geometric simulation—not a clinical promise or diagnosis. No patient files are accepted online, no pricing is listed, and all clinical decisions belong solely to the dentist.