ThreePod: A Verification-First System for Prompt- and Image-Driven 3D Modelling and Printing

Abstract: Generative 3D systems can produce a plausible object from a sentence or a single photograph, but plausibility is not the same as coherence, editability, correct scale, printability, or honest provenance. This paper presents ThreePod, a local-first, verification-first pipeline that treats every generative proposal as an untrusted candidate. The text route plans atomic actions, generates editable Three.js blocks, and commits a block only after schema, grounding, syntax, live- browser, and geometric checks. The image route segments a subject, reconstructs a shape, measures visible-side silhouette agreement, separates observed from inferred surface, and exports a part-labelled GLB. The print route converts an approved result into a simulated, layer-locked toolhead path with material and time budgets. We report an offline suite of 766 tests (723 passed, 2 skipped, 41 deselected, 0 failed), a ten-prompt scene benchmark with 401 geometry checks and zero failures, and four research image assets with valid GLB exports and visible-side IoU between 0.6016 and 0.9640. A CPU-only comparison of 3B and 7B local code models shows the expected quality–latency trade-off. The contribution is not a better-looking generator; it is an auditable pipeline that makes generation inspectable, measurable, and print- aware.

Keywords: 3D generation, text-to-3D, image-to-3D, verification, printability, LLM code generation, provenance, digital manufacturing.