From Prompt to Watertight STL: Neural Mesh Generation for Additive 3D Printing

Published Jul 24, 2026 • By 3D & AI Research Division • Vector Studios Research

Autonomous 3D Printing and Neural Slicing Engine

Standard generative 3D models produce chaotic, self-intersecting triangle meshes. While an unclosed mesh with missing faces might look acceptable in a video game shader with backface culling enabled, modern slicing engines (Bambu Studio, PrusaSlicer, Cura) will immediately reject it or generate catastrophic layer voids. Inside JARVIS 3D Creator, we pioneered an autonomous dual-pass neural pipeline designed specifically for physical additive manufacturing.

Engineering Takeaways

Implicit Volumetric Modeling: SDF representations guarantee watertight, manifold solid geometry from first principles.
Automated Pre-Flight Slicing: Catching thin walls and structural weaknesses in software saves hours of failed physical printing.
Quad-Dominant Edge Flow: Clean topology allows seamless transitions from physical 3D printing STL files to game-engine ready FBX/GLTF assets.
Micron-Level Verification: Additive manufacturing requires geometric tolerance validation, not just visual aesthetics.

Generate Watertight 3D Assets with JARVIS

Vector Studios builds neural geometry pipelines that convert prompts and concept art into production-grade CAD and game meshes.

Connect with 3D AI Research →

1. The Slicing Problem: Why Game Meshes Fail 3D Printers

A 3D printer does not see triangles; it evaluates planar slices of solid mathematical volumes. Traditional text-to-3D models suffer from three fatal flaws:

Non-Manifold Edges: Edges shared by three or more faces, confusing slicing boolean operations.
Inverted Surface Normals: Causing slicers to interpret outer walls as empty interior air.
Internal Floating Polygons: Creating hidden toolpaths that drag the print nozzle through partially cured plastic or metal powder.

2. Continuous Signed Distance Fields (SDF) Representation

To solve this fundamentally, JARVIS does not directly generate polygon vertex coordinates. Instead, our neural model synthesizes an implicit continuous Signed Distance Field (SDF). The function f(x, y, z) defines the exact distance from any coordinate in 3D space to the closest surface boundary, with the mathematical sign indicating strictly whether the point is inside or outside the solid body. Because the isosurface f(x,y,z) = 0 is continuous, it is mathematically impossible to produce non-manifold boundaries or unclosed seams.

3. Dual Marching Tetrahedra & Quad Retopology

Once the implicit SDF is synthesized, an adaptive Dual Marching Tetrahedra algorithm extracts the polygonal skin. Unlike naive marching cubes that produce jagged stair-stepping, our extractor aligns polygon edges along sharp mechanical surface normals. An automated second pass performs quad-dominant retopology, ensuring clean edge-flow and seamless subdivision for gaming and CAD export.

4. Embedded Headless Slicing Verification

Before an asset is delivered to the user or queued for production, JARVIS runs an embedded headless slicer instance in an isolated container. The automated check verifies:

Minimum Wall Thickness: Ensuring no structural feature is thinner than 0.4mm (FDM) or 0.15mm (SLA).
Overhang Angles: Calculating support volume requirements for overhangs exceeding 45 degrees.
First-Layer Bed Contact Area: Predicting print bed adhesion stability and warping risks.

5. Real-World Metallurgy: From Resin to Titanium EBM

We rigorously test JARVIS 3D output on real workshop machinery. From functional PLA/PETG prototypes on Bambu Lab hardware to high-precision dental resin and industrial titanium Electron Beam Melting (EBM) aerospace components, our generated meshes achieve verified dimensional tolerances within ±0.02mm.