Forked from apple/ml-sharp. Adds Apple Silicon MPS support and Gradio demo to Apple's official Sharp depth model.
ml-sharp runs the Sharp monocular depth estimation and novel view synthesis pipeline natively on Apple Silicon using PyTorch MPS acceleration. Feed it a single image and it produces depth maps and synthesised views from alternative camera angles — no NVIDIA GPU, no cloud API, just the M-series Neural Engine and GPU doing the work locally.
The Sharp model architecture uses a transformer encoder to estimate per-pixel depth from a single RGB image, then applies a point-cloud warp to synthesise views from nearby camera positions. Running on MPS gives 5–10× speedup over CPU on M1/M2/M3 chips. The pipeline includes preprocessing scripts, a Gradio demo interface, and export utilities for depth maps as EXR or PNG.
Uses PyTorch MPS backend — runs on M1 through M4 chips with full GPU acceleration, no CUDA required.
Generate plausible views from camera positions never seen during capture — useful for parallax effects and 3D previews.
Export dense per-pixel depth maps as 16-bit EXR or normalised PNG — feed into compositing tools or 3D pipelines.
Interactive web UI for uploading an image and previewing depth and synthesised views in the browser.
Standard Python package with pinned dependencies — pip install -r requirements.txt and it runs.