Sharp

Sharp

Turns a single photo into a 3D Gaussian scene in under a second on a standard GPU, ready to render in real time for nearby views. For researchers and graphics developers.

Made by apple on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Sharp is an Apple research model that builds a 3D scene from one photograph. A single pass through the network predicts a set of 3D Gaussians, in under a second on a standard GPU. You can then render that scene in real time from nearby camera positions.

Who it suits: Researchers and graphics developers who want quick 3D views from a single image. The scene has absolute scale, so metric camera movements make sense. The license is apple-amlr, so read its terms before using it.

Title: Who uses it, and how
Industry | Who uses it | How they use it | Example
Graphics research | Researchers | Study single-image view synthesis and compare against it | Run it on a set of photos and review the Gaussian output
XR and AR | Prototype developers | Turn a photo into a scene to look around in | Move a camera slightly around a photographed room
Visual effects | 3D artists | Make a quick 3D base from a still image | Open the output file in a Gaussian splat viewer
Tooling | Developers | Feed its output into splat renderers | Load the generated .ply file into a viewer

Skip it if you need views far from the original camera angle, because the card describes nearby views.

Where to use it:

• A terminal with the sharp command, set up from the project's code repository
• Prediction on a CPU, a CUDA GPU or Apple MPS
• Video rendering on a CUDA GPU
• Any public 3D Gaussian splat renderer, using the output .ply files

Use case | Who | What they get back
Predict a scene | Researchers | 3D Gaussian splat files in an output folder
Render a video | Developers | A video along a camera path, on a CUDA GPU
Inspect in a viewer | Artists | A scene to look around in a 3DGS renderer

How to set it up: You need Python and the environment from the project's code repository. Then you download the checkpoint file.

• Python and pip
• The environment from the code repository
• A CPU, CUDA GPU or MPS device for prediction, and a CUDA GPU if you want videos

1. Set up your environment with the repository's steps.
2. Download the checkpoint from this page, or with the Hugging Face command line.
3. Run the predict command on your images.

pip install huggingface-hub
huggingface-cli download --include sharp_2572gikvuh.pt --local-dir . apple/Sharp

How to use it: Run sharp predict with an input folder of images, an output folder and the checkpoint. Add the render option to make videos along a camera path, which needs a CUDA GPU.

sharp predict -i /path/to/input/images -o /path/to/output/gaussians -c sharp_2572gikvuh.pt

sharp predict -i /path/to/input/images -o /path/to/output/gaussians --render -c sharp_2572gikvuh.pt

# Or from the intermediate gaussians:
sharp render -i /path/to/output/gaussians -o /path/to/output/renderings  -c sharp_2572gikvuh.pt

Tool | What it does
sharp predict | Turns the photos in a folder into 3D Gaussian files
--render | Also renders videos along a camera path, on a CUDA GPU only
sharp render | Renders videos from Gaussian files you already made
-c | Points to the checkpoint file

Example: You run sharp predict on a folder that holds one photo of a room. The output folder fills with a .ply file of 3D Gaussians, which you open in a splat viewer and nudge the camera around.

For newcomers and experts: If you are new, think of it as a tool that guesses the 3D shape of a scene from one photo. Start with a single photo and open the result in a splat viewer.

For people in the field, a single network pass outputs every parameter of the Gaussian scene. The card says it cut LPIPS by 25 to 34 percent and DISTS by 21 to 43 percent against the best earlier model, with synthesis three orders of magnitude faster. It follows the OpenCV coordinate convention, with the scene center near the positive z axis.

Photograph goes in
One network pass predicts Gaussians
Save 3D Gaussian .ply files
Render nearby views in real time

Good to know:

• License: apple-amlr, so check its terms before use
• Cost to run: none named, you use your own computer
• Data it reaches: it downloads the checkpoint from Hugging Face, then reads your images and writes files to the folder you choose
• Not for: far-away viewpoints, or video rendering without a CUDA GPU
• Third-party renderers may need you to scale and rotate the scene to re-center it
• The gsplat renderer takes a while to start the first time

On Hugging Face, 13 November 2026

• Task: image to 3d
• Library: ml-sharp
• Downloads in the last 30 days: 4,150
• Likes: 423
• License: apple-amlr
• Quantized versions: 9
• Fine-tunes: 2
• Tags: ml-sharp, image-to-3d

Links

• Model: https://huggingface.co/apple/Sharp
• Files: https://huggingface.co/apple/Sharp/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download apple/Sharp

Then follow the usage in its model card. Nothing in this kit runs by itself.