Project

ml-sharp

active

Sharp monocular view synthesis for Apple Silicon -- depth-aware novel view generation on-device via MPS.

Overview

Forked from apple/ml-sharp. Adds Apple Silicon MPS support and Gradio demo to Apple's official Sharp depth model.

This fork adds an Apple Silicon fallback for the included demo video generator. Feed it a single image and it uses the Sharp monocular-depth network plus a depth-parallax warp to produce a swipe video locally through PyTorch MPS, without an NVIDIA GPU or cloud API.

How it works

On macOS, the demo uses Sharp's monocular-depth sub-network and a depth-parallax warp rather than the CUDA rasteriser used by the full 3D Gaussian splatting pipeline. The fork also adds a Gradio interface and can fall back to CPU when MPS is unavailable.

Features
🍎
Apple Silicon native

Uses PyTorch MPS for the fallback demo path on Apple Silicon, with CPU fallback and no CUDA requirement.

🔭
Novel view synthesis

Generate a depth-parallax swipe video from a single image for previews on macOS.

🗺️
Depth-driven fallback

The underlying Sharp pipeline produces depth data and 3D Gaussian splats; the macOS fallback uses its depth output for parallax rendering.

🖥️
Gradio demo

Interactive web UI for uploading an image and previewing depth and synthesised views in the browser.

📦
pip-installable

Standard Python package with pinned dependencies -- pip install -r requirements.txt and it runs.

Architecture
Input image single RGB frame SHARP weights Apple checkpoint macOS fallback PyTorch + MPS SHARP depth monocular map Parallax warp camera sweep Swipe video depth parallax SHARP depth drives a local parallax-video fallback on Apple Silicon
Posts
AI Image Generation