SHARP demo: explore the results

SHARP · IMAGE TO 3D

See Apple’s single-image view synthesis examples, then generate a 3D Gaussian representation from your own photograph locally.

Explore the official comparison gallery

This gallery contains pre-rendered research examples with interactive scene and method comparisons. It does not upload your photo or run SHARP inference in your browser.

Open gallery in a new tab ↗

Generate a scene from your own image

Install SHARP

Python 3.13 · official SHARP CLI

conda create -n sharp python=3.13 -y
conda activate sharp
git clone https://github.com/apple-aiml-research/ml-sharp.git
cd ml-sharp
python -m pip install -r requirements.txt
sharp --help

Put one or more photos in an input directory, then run prediction. The checkpoint downloads automatically on the first run.

Predict

sharp predict -i ./input -o ./output

Successful prediction writes 3D Gaussian .ply files to the output directory. Open them in a compatible Gaussian-splat viewer. The result supports nearby viewpoints; it is not a complete reconstruction of unseen parts of a scene.

Choose the right device

Gaussian prediction

CPU, CUDA or Apple MPS

Render a camera-path video

CUDA GPU required by the official renderer

Optional: render a video on CUDA

CUDA

sharp predict -i ./input -o ./output --render
# Or render previously generated Gaussians:
sharp render -i ./output -o ./renderings

If something goes wrong

• First run appears slow: allow time for checkpoint download and initialization; inspect the terminal output.
• Video rendering fails on Mac: prediction supports MPS, but the official --render path requires CUDA.
• Scene orientation looks wrong in another viewer: SHARP uses OpenCV coordinates; check scale and rotation in that viewer.

Official code and examples · checked September 12, 2026

Apple GitHub ↗ · Official examples ↗ · Paper ↗

Image to 3D: one photo, a 3D Gaussian splat

IMAGE TO 3D · GAUSSIAN SPLATS

Apple SHARP predicts a 3D Gaussian representation from a single photograph in under a second on a GPU and renders nearby viewpoints in real time. This page is the task-first guide: what you get, how to run it locally on Mac or CUDA, and how to view the result.

See official SHARP examples · GitHub

What you get

• A .ply file of 3D Gaussians with metric scale, ready for splat viewers.
• Real-time parallax around the original viewpoint—think 3D photos, not a full model.
• Optional rendered camera-path videos on CUDA GPUs.

Run SHARP locally

1. Create a Python 3.13 environment and install the official repository.
2. Put one or more photos in an input folder.
3. Run sharp predict; the checkpoint downloads on first use.
4. Open the output .ply in a Gaussian-splat viewer.

Python 3.13 · official SHARP CLI

conda create -n sharp python=3.13 -y
conda activate sharp
git clone https://github.com/apple-aiml-research/ml-sharp.git
cd ml-sharp
python -m pip install -r requirements.txt
sharp predict -i ./input -o ./output        # CPU, CUDA or MPS
sharp predict -i ./input -o ./output --render   # video rendering needs CUDA

Viewing the result

SHARP writes standard 3D Gaussian .ply files. Web-based splat viewers open them directly; SHARP uses OpenCV camera conventions, so check scale and orientation if a viewer shows the scene flipped.

Need a relightable asset instead?

Luce (Apple, August 2026) targets objects rather than scenes and predicts PBR materials—albedo, metallic-roughness and normals—so the asset can be relit and exported as a mesh. Code was not confirmed as public at the time of checking.

Luce → · Luce vs SHARP →

Questions

Is this the same as photogrammetry?

No. Photogrammetry needs many photos and reconstructs the full object. SHARP infers a plausible 3D scene from one photo and is designed for nearby views only.

Can it run in the browser?

Generation runs locally with the official CLI; there is no in-browser inference. Viewing the .ply output in a web splat viewer works fine.

Does it work on iPhone?

Not as shipped; the reference implementation is a Python package for Mac and Linux. Render the result on a computer, then share the video or splat file.

Hardware

• Prediction: CPU, CUDA or Apple Silicon (MPS).

• Video rendering (--render): CUDA GPU required by the official renderer.

• A recent Mac produces a scene in seconds; unseen regions stay unreconstructed.

• SHARP model page

• Depth Pro

• More tools