At Elf 3D we occupy a very narrow slice of the technology world: we write small, highly-specific geometric algorithms for teams that the big vendors overlook. Our latest collaboration was with an orthotics workshop that prints personalised foam insoles and prosthetic covers. Every model they produce starts as a hand-held 3D scan of a real limb, usually captured in a clinic or even in a living-room. This freedom makes life easier for patients, but it fills every mesh with gaps, folds and random noise the commercial repair plug-ins can’t handle.
The workshop was clear: “Please clean our meshes—without sanding away the subtle contours that keep our clients comfortable.” They didn’t need AI hype, cloud subscriptions or yet another licence fee. They needed a dependable algorithm that ran on the same workstation as their slicer. That is exactly the kind of brief we live for.
Generic “one-click fix” tools either smoothed the model into a potato or kept the defects intact. We needed surgical precision, not blunt force.
We wrote the repair code from a blank repo, so every line targets the client’s scans and nothing is weighed down by third-party baggage. First, the raw “triangle soup” is rebuilt into our own volumetric half-edge graph, where every vertex, edge, and face remembers where it came from; if a piece later vanishes, we know exactly why. We detect self-intersections with exact arithmetic to avoid the tolerance tricks that usually blur fine anatomical details.
Next, we decide what belongs inside the model and what doesn’t. Four rules—Even-Odd, Oriented-Sum, Union, and Intersection—cover messy walk-around scans, asymmetric arches, merged partial scans, and quality checks, respectively. A technician picks the rule with a single JSON flag, no menus required.
Finally, an adaptive patcher closes only the gaps the printer can resolve. It inserts a Delaunay surface, projects it to follow the local curve, and merges any sliver thinner than the nozzle width, so edges don’t bulk up. Normals are unified automatically, because the slicer rejects mixed orientations. The whole repair finishes in under a second on an ordinary desktop.
Detail loss in shallow wrinkles: Our first prototype mistook genuine plantar wrinkles for noise, erasing subtle pressure-relief channels. We rewrote the classifier to accept a per-file minimum-thickness hint supplied by the workshop.
False positives on fabric: A sock cuff hugging an ankle still fools the system; a technician currently sketches a “keep” mask before repair.
UI freezes on huge scan: Anything above eight million triangles locked the preview. We’re experimenting with live decimation and GPU streaming, but that remains on the to-do list.
Today the workshop exports a raw scan, ticks one checkbox, and receives a printer-ready STL without opening Blender. 3D-printing starts earlier in the day, so most clients now pick up their insoles the same afternoon instead of “sometime tomorrow.” The change is modest, but for a sports rehab patient or an elderly customer, a single day matters.
We’re extending the engine with selective remeshing: densifying the plantar arch while coarsening flat support zones to cut print time further. We also plan to embed a cloth-detector that separates woven texture from skin without manual masks. And yes—live previews that never freeze, even on monster scans. The algorithm is a living thing; as the client’s needs evolve, the code evolves with them.
About Elf.3D: We are a small team of three engineers and one project manager. Rather than selling the same software to everyone, we create new code for each partner and their specific data. If your 3-D workflow suffers from geometry issues that mainstream tools overlook, contact us at info@elf3d.com. We will start with a blank repository and craft a solution around your needs.