buildoak/sharp-onnx

SHARP ONNX — Apple's Single-Image 3D Gaussian Splatting

ONNX export of Apple's SHARP model for use with tortuise, a terminal-native 3D Gaussian Splatting viewer.

Files

File	Size	Description
sharp.onnx	1.9 MB	Model structure (ONNX graph)
sharp.onnx.data	2.6 GB	Model weights (external data)

Both files are required. The model exceeds protobuf's 2GB limit, so weights are stored separately.

Usage

These files are automatically downloaded by tortuise when you run:

cargo install tortuise --features sharp
tortuise photo.jpg

Or manually place both files in ~/.tortuise/models/.

Model Details

• Architecture: DINOv2 ViT-Large encoder + Sliding Pyramid Network + DPT decoders
• Parameters: 702M (340M trainable)
• Input: Single RGB image (resized to 1536×1536 internally)
• Output: ~1.2M 3D Gaussians (positions, scales, rotations, colors, opacities)
• ONNX opset: 17
• Source checkpoint: sharp_2572gikvuh.pt from apple/Sharp

License

The model weights are licensed under the Apple Machine Learning Research Model License. This is a research-only, non-commercial license. See the LICENSE file for full terms.

This ONNX conversion is a format transformation of Apple's original PyTorch checkpoint. No architectural modifications were made.

Attribution

Based on Apple SHARP model. Copyright (C) 2025 Apple Inc. Licensed under the Apple Machine Learning Research Model License Agreement.

Paper: SHARP: Monocular View Synthesis in Less Than a Second