A browser-based Gaussian splat generator built on top of Apple SHARP. ✨
This project lets you:
upload one image
generate Gaussian splats in the browser
preview the result
download a .ply file
.ply
Links
Project repo: bring-shrubbery/ml-sharp-web
Follow the author on X: @bringshrubberyy
Upstream SHARP repo (Apple): apple/ml-sharp
SHARP project page: apple.github.io/ml-sharp
SHARP paper: arXiv:2512.10685
sharp_web_predictor.onnx
sharp_web_predictor.onnx.data
sharp_web_predictor.onnx.data
Both files must be served together from the same folder (for example public/models/).
public/models/
Why this matters:
The .onnx file is only the graph and metadata.
.onnx
The .onnx.data file contains most of the model weights.
.onnx.data
For that reason, the app uses the hosted model by default. Uploading only the .onnx file directly in the browser usually will not work because the .onnx.data sidecar is separate.
.onnx
.onnx.data
Export the SHARP model to ONNX (beginner-friendly steps)
Everything runs in the browser, but you still need an exported SHARP ONNX model.
In-page preview with @mkkellogg/gaussian-splats-3d
@mkkellogg/gaussian-splats-3d
Troubleshooting 🛠️
"expected magic word ... found 3c 21 64 6f" (WASM error)
This means a WASM file request returned HTML instead.
Try:
run the app with bun dev (not file://...)
bun dev
file://...
restart the dev server after bun install
bun install
verify these load in your browser: /ort/ort-wasm-simd-threaded.asyncify.mjs /ort/ort-wasm-simd-threaded.asyncify.wasm
/ort/ort-wasm-simd-threaded.asyncify.mjs
/ort/ort-wasm-simd-threaded.asyncify.mjs
/ort/ort-wasm-simd-threaded.asyncify.wasm
/ort/ort-wasm-simd-threaded.asyncify.wasm
"Failed to load external data file ... sharp_web_predictor.onnx.data"
This means the ONNX sidecar file is missing or not served correctly.
Check:
public/models/sharp_web_predictor.onnx
public/models/sharp_web_predictor.onnx
public/models/sharp_web_predictor.onnx.data
public/models/sharp_web_predictor.onnx.data
The app can reach the hosted model files in your deployment/browser environment
The app runs, but generation is very slow or crashes
SHARP is large and browser inference is heavy.
Try:
Chrome or Edge (desktop)
smaller Max gaussians in the UI
Max gaussians
closing other memory-heavy tabs/apps
waiting longer on first run (model + runtime initialization can take time)
Tech stack
Bun
React
TypeScript
Vite
ONNX Runtime Web
GaussianSplats3D viewer
Project status
Working prototype / experimental. 🧪
The app runs end-to-end in the browser, but performance and compatibility depend heavily on browser WebGPU/WASM support and your machine's available memory.
Before you start (important license note)
Apple's SHARP repository has separate licenses for code and model weights.
SHARP code license: LICENSE
SHARP model license: LICENSE_MODEL
If you use Apple's released SHARP checkpoint/weights, you must follow LICENSE_MODEL (research-use restrictions apply).
LICENSE_MODEL
What you need
Bun installed
A modern desktop browser (Chrome or Edge recommended)
Enough disk space and RAM for the SHARP model (the exported ONNX sidecar is large, ~2.4 GB)
Quick start (run the app) 🚀
Star this repo 🤩
If this project helps you, please star it:
bring-shrubbery/ml-sharp-web
Install dependencies
bun install
This also copies ONNX Runtime Web WASM assets into public/ort/ automatically.
public/ort/
Start the app
bun dev
Open the URL shown by Vite (usually http://localhost:5173).
http://localhost:5173
Use the app
Upload an image.
Click Generate Splat.
Generate Splat
Preview the result and download the .ply file.
.ply
Important model file note (.onnx + .onnx.data)
.onnx
.onnx.data
SHARP exports usually produce two files:
sharp_web_predictor.onnx
sharp_web_predictor.onnx
sharp_web_predictor.onnx.data
sharp_web_predictor.onnx.data
Both files must be served together from the same folder (for example public/models/).
public/models/
Why this matters:
The .onnx file is only the graph and metadata.
.onnx
The .onnx.data file contains most of the model weights.
.onnx.data
For that reason, the app uses the hosted model by default. Uploading only the .onnx file directly in the browser usually will not work because the .onnx.data sidecar is separate.
.onnx
.onnx.data
Export the SHARP model to ONNX (beginner-friendly steps)
Everything runs in the browser, but you still need an exported SHARP ONNX model.
In-page preview with @mkkellogg/gaussian-splats-3d
@mkkellogg/gaussian-splats-3d
Troubleshooting 🛠️
"expected magic word ... found 3c 21 64 6f" (WASM error)
This means a WASM file request returned HTML instead.
Try:
run the app with bun dev (not file://...)
bun dev
file://...
restart the dev server after bun install
bun install
verify these load in your browser: /ort/ort-wasm-simd-threaded.asyncify.mjs /ort/ort-wasm-simd-threaded.asyncify.wasm
/ort/ort-wasm-simd-threaded.asyncify.mjs
/ort/ort-wasm-simd-threaded.asyncify.mjs
/ort/ort-wasm-simd-threaded.asyncify.wasm
/ort/ort-wasm-simd-threaded.asyncify.wasm
"Failed to load external data file ... sharp_web_predictor.onnx.data"
This means the ONNX sidecar file is missing or not served correctly.
Check:
public/models/sharp_web_predictor.onnx
public/models/sharp_web_predictor.onnx
public/models/sharp_web_predictor.onnx.data
public/models/sharp_web_predictor.onnx.data
The app can reach the hosted model files in your deployment/browser environment
The app runs, but generation is very slow or crashes
SHARP is large and browser inference is heavy.
Try:
Chrome or Edge (desktop)
smaller Max gaussians in the UI
Max gaussians
closing other memory-heavy tabs/apps
waiting longer on first run (model + runtime initialization can take time)
Tech stack
Bun
React
TypeScript
Vite
ONNX Runtime Web
GaussianSplats3D viewer
Project status
Working prototype / experimental. 🧪
The app runs end-to-end in the browser, but performance and compatibility depend heavily on browser WebGPU/WASM support and your machine's available memory.