SHARP — Single-Image 3D Gaussian View Synthesis

Mirror of Apple's SHARP model weights, converted to safetensors format for use with ComfyUI-FFMPEGA.

Model Description

SHARP predicts 3D Gaussian Splat (3DGS) parameters from a single photograph in under 1 second, then renders camera trajectory videos using gsplat.

• Input: Single RGB image
• Output: 3D Gaussian splat representation → camera trajectory video or .ply export
• Speed: <1s prediction on GPU
• Resolution: 1536×1536 internal processing

Usage in ComfyUI-FFMPEGA

1. Set llm_model to none
2. Set no_llm_mode to sharp
3. Connect an image to image_a or image_path_a
4. Run — the model will auto-download on first use

Parameters

Parameter	Default	Description
sharp_trajectory	rotate_forward	Camera motion: rotate_forward, swipe, shake, rotate
sharp_num_frames	60	Number of video frames (10–300)
sharp_max_disparity	0.08	Lateral camera range
sharp_max_zoom	0.15	Zoom intensity
sharp_save_ply	false	Export .ply Gaussian splat file

License

⚠️ Research Use Only — Non-Commercial

This model is licensed under the Apple Machine Learning Research License. Model weights are restricted to Research Purposes only — non-commercial scientific research and academic development.

See LICENSE for the full terms.

Citation

@article{stier2025sharp,
  title={SHARP: Synthesizing 3D Gaussians from a Single Monocular Image with High Fidelity and Accurate Geometry},
  author={Stier, Nikolai and Wadhwa, Neal and Szeliski, Richard},
  year={2025},
  url={https://github.com/apple/ml-sharp}
}

Original Repository

• Code: github.com/apple/ml-sharp (BSD-like license)
• Paper: See repository for links
• Copyright: © 2025 Apple Inc. All Rights Reserved.