Modelr turns images into 3D on Apple Silicon in under 20 seconds

Developer ZimengXiong announces Modelr, an open-source macOS and iOS app that converts 2D images into 3D meshes entirely on-device — a first for Apple Silicon. The app is a Swift-MLX/Python MLX port of Tencent's Hunyuan3D-Paint and Hunyuan3D-Shape models, eliminating the overhead of PyTorch. On a M4 Max in FP16, the small shape model finishes in 20.9 seconds using just 5.6 GB of peak memory; the large shape model takes 22.3 seconds at 7.3 GB. The texturing pipelines (Hy3D Paint) require more resources (231–344 seconds, ~38–39 GB), but the core shape generation is impressively fast and memory-efficient.

Modelr also supports iPhones thanks to quantized 4-bit and 8-bit weights, making it the only image-to-3D app running entirely on Apple's mobile ecosystem. Users import an image, automatically remove backgrounds with SwiftVision, and watch the diffusion process stream in real time. The developer admits the target use case is still exploratory — likely simple 3D assets for apps or previewers — but the technical achievement is notable. The full source code and model weights are available on GitHub, enabling other developers to integrate low-memory, fast 3D generation into their own Swift apps. Modelr represents a significant step toward democratizing 3D content creation on consumer hardware.