MIT's InstructMesh Fixes AI-Generated 3D Models So You Can Actually Print Them
MIT News AI · rss · 2026-10-02
- MIT CSAIL, Google, and Northeastern unveil InstructMesh, a WYSIWYG tool that lets users prompt-generate 3D designs, highlight specific parts to refine, and 3D-print working objects.
- The core problem: AI models know how objects look, not how they work. Rebuilding popular Thingiverse models with TRELLIS, researchers found nearly 80% were structurally flawed; novices using InstructMesh identified and fixed issues about 90% of the time.
- Approach: pairs Microsoft's TRELLIS 3D generator with GPT-4's reasoning, letting users describe issues in natural language and approve geometry edits in the model's latent space.
- Demo builds include a mug wrapped by a dragon (tail as handle), an octopus drink dispenser, a denim-look knee brace, and a shrimp-shaped bristle bot enclosure.
- Lead author Faraz Faruqi, now at Google, plans AR integration (e.g., prompt a phone case matching your wallet), physics simulation, and the newer TRELLIS.2. The paper appears at ACM UIST in November.
More from Embodied
- Kaigen dev confirms its AI 3D demos run entirely inside the Unity engine — gdechichi · 2026-10-02
- Cartwheel teases kinematics data breakthrough with a wink — andrew_n_carr · 2026-10-02
- Teleop a humanoid just by wearing glasses, and it sees through them too — danfei_xu · 2026-10-02
- One policy for millions of robots: team builds generalist motor control policy trained on 200+ real robot models — breadli428 · 2026-10-02
- Clone Robotics shows off Torso 3 humanoid hand, betting on an unconventional design path — MarwaEldiwiny · 2026-10-02
- Developer shows AI wearable dot proactively pinging him about schedule choices — dkundel · 2026-10-02