---
license: other
license_name: apple-ml-research-model-license
license_link: LICENSE
library_name: onnx
pipeline_tag: image-to-3d
tags:
  - onnx
  - onnxruntime-web
  - webgpu
  - gaussian-splatting
  - single-image-3d
---

# SHARP (ONNX, WebGPU)

An ONNX export of [SHARP](https://github.com/apple/ml-sharp) ("Sharp Monocular View Synthesis in Less Than a Second"), prepared for in-browser inference with ONNX Runtime Web on WebGPU. It is used by [SharpRig](https://github.com/sm079/sharp-rig), which turns a single photo into a camera-move video in the browser.

> Apple Machine Learning Research Model is licensed under the Apple Machine Learning Research Model License Agreement.

This is a **model derivative**, not an official Apple release, and it is not endorsed by Apple. Like the original, it may be used for **non-commercial research purposes only**; see [LICENSE](LICENSE).

## Files

| File | Size | |
|---|---|---|
| `sharp.onnx` | 4 MB | graph, with the weights stored as external data (`sharp.onnx.data`) |
| `sharp.int8.bin` | 0.66 GB | the external weights as int8 + per-output-channel scales, expanded back to fp16 on load |

## Modifications from the original checkpoint

- **Export:** the network was traced in fp16 (fp32 inputs and outputs) from `sharp_2572gikvuh.pt` and exported to ONNX.
- **Unprojection moved out of the graph:** the final unprojection from NDC to metric space, which needs an SVD that ONNX can't express, is left to the caller as a per-axis scale.
- **Int8 packing:** the fp16 weights are packed as int8 with one float32 scale per output channel (round-to-nearest, weight-only). This halves the download. The pack is dequantised back to fp16 before inference, so the numerics are those of an fp16 model with int8-rounded weights.
- **Accuracy:** measured against the fp32 checkpoint on a test photo, relative depth error is 0.19% (median) and 1.1% (p95).
- **Code:** the export and packing are done by [`tools/export_sharp_onnx.py`](https://github.com/sm079/sharp-rig/blob/main/tools/export_sharp_onnx.py). No retraining or fine-tuning was done.

## Inputs and outputs

- **Inputs:** `image` float32 `[1, 3, 1536, 1536]` (RGB in [0, 1]); `disparity_factor` float32 `[1]` (focal length in px / image width).
- **Outputs** (N = 2 × 768 × 768):
  - `mean_vectors` `[1, N, 3]`
  - `singular_values` `[1, N, 3]`
  - `quaternions` `[1, N, 4]` (w, x, y, z)
  - `colors` `[1, N, 3]` (linear RGB)
  - `opacities` `[1, N]`

  All are Gaussians in SHARP's NDC space.