← Back to Index Scene_01 · SHARP ● live
Scene 01 · Monocular Synthesis · 2025

SHARP

Apple ML · SHARP · Single Image → 3DGS

SHARP starts with a single photograph and ends with a full 3D Gaussian Splat. The scene established the reconstruction branch of Spatial Index by testing how far monocular inference could go without multi-view capture or photogrammetry.

Python Gradio 3DGS .ply
Toronto, ON
2025
OUTPUT
.PLY
Gaussian Splat
01 · Problem

3D capture usually asks for too much setup.

Photogrammetry and multi-view reconstruction can produce rich scenes, but they require deliberate capture workflows, overlap discipline, and more input than many quick experiments can afford.

SHARP asks what happens when that barrier collapses to a single image. If one photograph can become a Gaussian Splat in under a second, then spatial reconstruction becomes lightweight enough to use as a routine creative primitive.

SHARP single image to Gaussian Splat overview
SHARP · monocular view synthesis into a 3D Gaussian Splat
02 · Context

Inference replacing capture overhead.

This project sits at the beginning of the Spatial Index stack because it reframes reconstruction as an inference problem. Instead of carefully collecting a scene, SHARP tests how much spatial plausibility can be recovered from monocular input alone.

That matters beyond convenience. It opens a path where everyday images become candidate spatial assets, not just references, and where 3DGS workflows can start from much looser source material.

03 · Approach

One image in, full splat out.

A single photograph is passed through Apple ML’s SHARP pipeline, which produces a full 3D Gaussian Splat without any multi-view capture stage. The output is exported as `.ply` and surfaced through a lightweight interface for inspection and sharing.

The implementation follows the model pipeline closely rather than reinventing it, with the main goal being operational fluency: understand the data flow, make it accessible, and verify that the reconstruction could be used as input for later scenes.

04 · Stack

A minimal path into 3D Gaussian Splats.

Rendering
Gradio interface · splat inspection viewer
Backend / Data
Python ML pipeline · Apple ML SHARP model
Pipeline
Single image → monocular inference → 3D Gaussian Splat → `.ply` export
Deploy
Hugging Face Spaces · sharpview.spatial-index.xyz
05 · Reflections

Fast reconstruction changes the cadence of experimentation.

SHARP made the broader project feasible by lowering the cost of entry into spatial output. Once splats can be created from single images, reconstruction stops being a special event and becomes something you can iterate on rapidly.

It also exposed a practical limitation: using a model well is not the same as understanding it deeply. That tension shaped later scenes, where the goal shifted from just running pipelines to bending them into new interfaces and experiences.

06 · Build Log

From first image to first splat.

2025
Pipeline adoption
Set up and validated the SHARP monocular reconstruction workflow as the first working 3DGS path in the project.
2025
Interface wrapper
Exposed the model through a simple Gradio surface so single-image tests could be run and inspected quickly.
2025
PLY verification
Confirmed that the generated splats could be exported as `.ply` and reused as inputs for downstream experiments.
2025
Launch
Published the Hugging Face Space and established Scene 01 as the baseline reconstruction experiment for Spatial Index.