Local Image-to-3D: Runs on Mac and iPhone
arduinoRPi4 · reddit · 2026-07-12
The author completed Swift/MLX Python ports for Hunyuan3D-Paint and Hunyuan3D-Shape, sharing local performance metrics on an M4 Max:
- hy3d shape small model: 20.9s, peak 5.6GB
- hy3d shape large model: 22.3s, peak 7.3GB
- hy3d paint rgb: 231s, peak 38GB
- hy3d paint pbr: 344s, peak 39GB
This enables the models to run on recent Mac generations, and even on iPhones under Q4/Q8 conditions. The companion app, Modelr, supports image import, background removal via SwiftVision, and real-time viewing of the diffusion and texture generation processes. The author also open-sourced the iOS/macOS app and model weights to help others integrate fast, low-memory image-to-3D capabilities into their own Swift apps.
More from Embodied
- CoRL paper: flow-matching VLAs can run 1-step Euler inference with no retraining — kaixhin · 2026-10-26
- Robotic dexterity should be judged on speed, force control and tool use, not copying humans — DJiafei · 2026-10-26
- Shenzhen company combines humanoid robots with industrial robot arms — RoboBalaji · 2026-10-26
- Realbotix, a sex-doll maker, unveils Echo humanoid with lifelike face — CyberRobooo · 2026-10-26
- Track bolt robot tightens railway fasteners in 6–7 seconds with sub-mm precision — RoboBalaji · 2026-10-26
- Tesla FSD (Supervised) Approved in Latvia, Its 10th European Country — XFreeze · 2026-10-26