Printability methodology
Version 1.0. Six dimensions. Five auto-scored (out of 75); the sixth — visual fidelity — is graded manually with the rater under each result card and bumps the total to /100. Every dimension is reproducible from the input mesh.
Can you go from a single photo to a printable STL — without a CAD operator — and how good is the result?
The 6 dimensions
Manifold / watertight
25 ptsbinaryDoes the mesh pass a strict manifold check (every edge bordered by exactly 2 faces, no holes)? If not, the slicer rejects the file. This is the foundational gate.
Vertex count sanity
10 ptsbandedPolygon count appropriate for the bounding box. 5,000–100,000 verts = 10pts. Below 500 or above 1M = 0pts.
Wall thickness
20 ptsbandedMinimum wall thickness, normalized to a 100mm bounding box. ≥1.2mm = 20pts. <0.4mm = 0pts. Estimated via voxel-grid sampling.
Overhang risk
15 ptsbandedPercentage of faces with normals more than 45° from vertical. ≤15% = 15pts. >50% = 0pts. Scored as-delivered, no auto-orient.
File-size efficiency
5 ptsbandedSTL bytes per 1000 vertices. Catches outputs that bloat normals/UVs without adding geometry.
Visual fidelity
25 ptsmanual rater · auto coming v0.2Does the output look like the source image? Excluded from the score by default — auto-only totals are out of 75. Use the rater under each result card to grade Poor (0) / Mid (12) / Good (25), and the score extends to /100. CLIP-similarity automation planned for v0.2.
Wall thickness is estimated via raycast sampling, not full voxelization — quick but approximate. Visual fidelity is scored manually at v0; the CLIP-similarity automation is planned for v0.2. Manifold check is heuristic (edge-count) at v0; strict manifold-3d WASM check planned. Scores ending in “partial auto” mean at least one dimension was scored by hand or with a placeholder.
Methodology source: github.com/forkable-factory. Disagree with a score? Open an issue with the input image and the alternative score.