apple ml-sharp
Sharp Monocular View Synthesis in Less Than a Second
3D Gaussian Splats
Interactive View Synthesis
How it Works
SHARP is not just a model; it's a highly optimized feedforward pipeline designed for speed and fidelity.
1. Input Image
Any single RGB photo
2. SHARP Network
Depth Pro + Gaussian Regression
3. 3DGS Parameters
Generates thousands of Gaussians
4. Real-time Render
100+ FPS free view
Faster than fast.
The power of feedforward.
Most modern view synthesis methods rely on slow iterative diffusion processes. SHARP changes the game by directly predicting 3D structure in a single feedforward pass, eliminating wait times.
Generation Time (Lower is Better)
* Performance measured on NVIDIA A100 GPU. SHARP achieves a 3 order-of-magnitude speedup.
Zero-shot Generalization
No re-training required for specific scenes. SHARP is pre-trained on massive real-world data and works out of the box.
Metric Scale
Generated scenes have physical metric scale. Moving 1 meter forward in VR actually looks like 1 meter.
Core Metrics
LPIPS Reduction
Significantly reduced perceptual error compared to previous SOTA, meaning closer to human visual realism.
Generation Time
Sub-second generation on standard consumer GPUs, saying goodbye to long rendering waits.
Rendering FPS
Generated .ply files are optimized for smooth real-time performance in existing 3DGS viewers.
Frequently Asked Questions
For running inference (generating the 3D model), any standard GPU setup works. However, to use the real-time visualization features provided in the repository, a CUDA-enabled GPU is currently required.
Diffusion models generate views by iteratively denoising an image, which is slow (minutes) and often inconsistent across frames. SHARP is a feedforward network that predicts the 3D structure (Gaussians) directly in one pass (< 1 second), ensuring consistency and speed.
The code and model weights are released under specific licenses. Please check the LICENSE and LICENSE_MODEL files in the GitHub repository for the most accurate and up-to-date usage terms.
You can run the Python inference scripts on Apple Silicon, but the real-time CUDA-based renderer is not compatible. You would need to use a compatible viewer for the generated .ply files.
Ready to dive in?
Join thousands of developers exploring the future of monocular 3D vision. The code is fully open source.
git clone https://github.com/apple/ml-sharp
cd ml-sharp
conda create -n sharp python=3.13
pip install -r requirements.txt# Automatically downloads model & processes image
sharp predict \
-i ./input_images \
-o ./output_gaussians \
--render