
PhotoSplat3D
PhotoSplat3D turns a single photo into a 3D Gaussian Splat entirely on device, on Apple Vision Pro and on Apple Silicon Macs.
Location
Hangzhou, China
Size
Total raised
Hiring?
Platform
Apple Vision Pro app on visionOS 26 or later, and a Mac app for Apple Silicon M2 or later, with a Python desktop GUI and CLI for macOS and Linux
GPU
No discrete GPU required. Inference runs on CPU, on Apple Silicon through MPS, or on CUDA, and on Vision Pro it runs on the headset's own silicon with no cloud step. Trajectory video rendering from the CLI is CUDA only
Pricing
Free. The App Store listing carries no purchase and the developer declares no data collection
License
MIT for the fork's own contributions, with Apple's upstream license on ml-sharp derived code and a separate license on the model weights
Best for
Turning one photograph into a Gaussian Splat without a capture rig or a cloud service
Updated
July 2026
PhotoSplat3D is a single image Gaussian Splatting application from Hangzhou Ninglin Cultural Innovation Co., Ltd. It ships in two forms, an Apple Vision Pro app on the App Store and a Mac app for Apple Silicon, and its source sits on GitHub as a desktop friendly fork of Apple's ml-sharp project. The underlying method is SHARP, published as Sharp Monocular View Synthesis in Less Than a Second. The App Store version is 2.4 and the app is free at 2.7 GB.
Reconstruction runs locally in every configuration. On Vision Pro a photo becomes an explorable model in seconds with no upload and no network requirement, and the developer declares that the app collects no data. On the Mac the same conversion runs in one click and adds batch processing over folders of images. The command line accepts a device flag of cpu, mps or cuda, so the model runs on Apple Silicon through Metal Performance Shaders or on an NVIDIA card, though the separate trajectory video renderer is CUDA only. The model checkpoint is roughly 300 MB and downloads on first inference.
Panoramas are handled as a first class input rather than as a special case. The predict command detects 180 degree and 360 degree equirectangular images automatically and reprojects them before reconstruction, with a choice of five strategies named cube6, ring8, ring12, front4 and front6 that trade the number of synthesised views against face resolution. That path is what lets a single 360 frame from a consumer camera become a splat without a multi image dataset behind it.
A separate import route brings in captures the app did not train. PhotoSplat3D reads public Insta360 Spatial Capture share links pasted into the app, downloading a compact spatial model of roughly 10 to 20 MB by default or the original PLY at 100 to 250 MB, which puts an Insta360 X6 Spatial Capture scene into an immersive Vision Pro viewer. The developer states plainly that this parses preview assets from public share pages, that it is not an official Insta360 API, and that PhotoSplat3D is not affiliated with Insta360, so the path depends on Insta360 leaving those pages as they are.
Beyond generation the Vision Pro app behaves as a viewer and a light editor. It imports PLY, Splat, SPZ and SOG, and rendering is credited to two open source projects, SuperSplat and MetalSplatter, the Metal renderer that reached 1.0 in February 2026. Spatial brush editing paints and erases directly on a model in space with 48 colour presets across six palettes and support for the Logitech MX Ink stylus. An orbit mode, gesture controls for move, rotate and scale, exposure adjustment, half and full immersion toggles and an in app community feed for sharing and remixing models round out the app. Text to image generation sits in front of the pipeline, so a prompt can become an image and then a splat without a camera at any stage.
Licensing is layered because the lineage is. The fork's own contributions are MIT, Apple derived code stays under Apple's upstream license with its copyright headers preserved on redistribution, and the model weights carry a third license of their own. Anyone redistributing a build needs to read all three rather than assuming the MIT notice covers the whole repository.
Frequently Asked Questions
Does PhotoSplat3D need a cloud service to make a Gaussian Splat?
PhotoSplat3D performs reconstruction entirely on device. On Apple Vision Pro the conversion runs on the headset with no upload and no network connection required, and the developer states that the app collects no data.
What hardware does PhotoSplat3D require?
PhotoSplat3D requires an Apple Vision Pro running visionOS 26 or later for the headset app, and a Mac with Apple Silicon M2 or later for the Mac app. The Python source additionally runs on macOS or Linux with CPU, MPS or CUDA inference.
How many photos does PhotoSplat3D need?
PhotoSplat3D reconstructs a Gaussian Splat from a single image. It is a fork of Apple's ml-sharp, which implements the SHARP monocular view synthesis method, so there is no multi image dataset and no structure from motion step.
Can PhotoSplat3D open Insta360 Spatial Capture scenes?
PhotoSplat3D imports public Insta360 Spatial Capture share links, pulling a compact spatial model of roughly 10 to 20 MB or the original PLY at 100 to 250 MB. The developer states this parses preview assets from public share pages rather than an official Insta360 API, and that PhotoSplat3D is not affiliated with Insta360.
Which Gaussian Splat formats can PhotoSplat3D open?
PhotoSplat3D imports and views PLY, Splat, SPZ and SOG models, and it writes its own output as .ply. Rendering is credited to the SuperSplat and MetalSplatter open source projects.
Is PhotoSplat3D free and open source?
PhotoSplat3D is free on the App Store and its source is published on GitHub. The repository is multi licensed, with MIT covering the fork's own contributions, Apple's upstream license covering ml-sharp derived code, and a separate license covering the model weights.
Alternatives
Recent Headlines
PointCloudSnapPro Adds Gaussian Splats to Maya 2026
Foundry Ships Nuke 17.1v1 With Dynamic Gaussian Splat and Relighting Support
NavVis Adds Gaussian Splats to IVION 12
LocalMesh Opens Its Beta, Turning One Photo Into a Gaussian Splat and a Textured Mesh on a Local GPU
Streamed Reality Generates Gaussian Splats Inside the Unity Editor
PlayCanvas Opens SuperSplat to Third Party Tools With a REST Publishing API
I Tried to Make Gaussian Splatting 1,000× Faster
GaussianSplatting.jl v2.0.0 Adds MCMC and Sky Dome Pass
Laskos Releases 360 Gaussian Pro With a Rebuilt COLMAP
TACTYC Ships Onesplat, Generating Gaussian Splats From a Single Image Inside Houdini

