InstructMesh separates intent, spatial grounding, and geometry editing. The user selects a flawed region. An LLM maps natural language to an operation and parameters from a fixed vocabulary; it does not edit 3D geometry directly. The system modifies a coarse latent voxel representation, previews additive or subtractive changes, and decodes an updated mesh after user confirmation.[1][2]
The reported numbers have bounded tasks. In 120 TRELLIS reconstructions, 94 models had more than one fabrication-related issue. Study 1 recruited 12 novices to inspect and repair five flawed models; within that study they identified 90.4% of annotated flaws and repaired 89.7%.[2]
The current implementation targets visually identifiable static geometry. A 64^3 grid limits feature granularity, users must diagnose the problem region, fabricated examples are applications rather than systematic printability validation, and dynamic mechanisms are out of scope. Physics simulation is described as future work.[1][2]
Sources:
- https://www.csail.mit.edu/news/new-tool-lets-users-repair-ai-generated-3d-models-then-fabricate-them-just-way-they-want
- https://arxiv.org/html/2608.28534
Reader links:
Top comments (0)