This is a simplified guide to an AI model called Sharp-Ml maintained by Kfarr. If you like these kinds of analysis, you should join AImodels.fyi or follow us on Twitter.
Overview
sharp-ml is Apple's SHARP model, which converts a single photograph into 3D Gaussian splats in less than a second via a single neural network feedforward pass. Built and maintained by kfarr, this model produces photorealistic 3D scene representations that can be rendered in real time from nearby viewpoints with metric scale support for accurate camera movements. The approach achieves state-of-the-art results, reducing perceptual loss (LPIPS) by 25–34% and structural dissimilarity (DISTS) by 21–43% compared to prior methods while delivering inference that is three orders of magnitude faster than competing approaches. The model demonstrates robust zero-shot generalization across different datasets without task-specific fine-tuning.
Best use cases
3D product visualization for e-commerce. Product photographers can capture a single high-quality image and immediately generate interactive 3D models for websites, allowing customers to explore items from multiple angles without requiring expensive multi-camera capture rigs or manual 3D modeling. SHARP's metric scale output ensures that spatial relationships remain physically accurate.
Architectural preview and real estate walkthroughs. Real estate agents and architects can take single photographs of interior or exterior spaces and convert them into navigable 3D environments, enabling potential buyers or stakeholders to virtually explore properties from viewpoints not captured in the original image. The metric representation supports consistent spatial understanding across a property.
Game asset prototyping from concept art. Game developers can transform 2D concept artwork or reference photographs into initial 3D geometry for further refinement, accelerating the asset creation pipeline by eliminating manual blockout modeling for scenes where directional variation is needed.
VR environment creation from single images. Content creators can generate stereo-compatible 3D representations from single photographs for VR applications, enabling immersive experiences from existing image assets without requiring depth sensors or multiple camera angles.
Visual effects and video compositing preparation. VFX artists can quickly generate 3D geometry from reference photographs to support compositing workflows, particularly for background plates where depth information needs to be estimated and rendered from novel angles to match camera movements.
Limitations
The model requires single-image inputs only; it cannot accept multiple images of the same scene or video sequences to refine the reconstruction. Rendering trajectories (video generation from the Gaussian splats) requires a CUDA-capable GPU; CPU and MPS (Apple Metal Performance Shaders) can generate the splats but cannot render video output. The model produces 3D Gaussian representations rather than traditional mesh geometry, which may require compatible third-party rendering software and limits interoperability with standard 3D modeling tools. View synthesis quality degrades significantly for large camera movements outside the nearby view range; the model is optimized for modest baseline displacements, not 360-degree rotations or extreme extrapolation. Complex occlusions, thin structures, and textureless regions may produce artifacts or missing geometry. The output Gaussian splats follow OpenCV coordinate conventions (x right, y down, z forward) with scene centering around approximately (0, 0, +z), requiring manual transformation when integrating with renderers using different conventions. Input and output specifications are minimal in the Replicate schema, providing limited visibility into preprocessing requirements or maximum image dimensions during inference.
How it compares
TripoSplat also converts 2D images into 3D Gaussians using a learned density-control approach. Choose SHARP if you prioritize inference speed (sub-second) and metric-scale accuracy with absolute camera movements; choose TripoSplat if you need access through a different platform or have specific density refinement requirements. SHARP's three orders of magnitude faster inference makes it more suitable for real-time interactive applications.
Kandinsky 5.0 Pro Image-to-Video generates videos directly from images using diffusion, while SHARP generates 3D Gaussian geometry that you render. Choose SHARP for interactive 3D exploration and precise geometric control; choose Kandinsky 5 for quick video generation where the camera movement style matters less than production speed.
Kling Native 4K Image-to-Video produces 4K video output directly without upscaling. Choose SHARP if you need true 3D geometry, metric accuracy, and interactive rendering; choose Kling if you require high-resolution video output and do not need geometric precision or viewpoint control.
Stable Diffusion Videos (Mo-Di) interpolates Stable Diffusion's latent space for video generation. Choose SHARP for physically grounded 3D representations with metric scale; choose Stable Diffusion Videos for generative stylization and maximum flexibility in motion patterns.
Technical specifications
SHARP operates as a single feedforward neural network that regresses 3D Gaussian representation parameters from a single image. The model checkpoint (sharp_2572gikvuh.pt) is automatically downloaded and cached locally at ~/.cache/torch/hub/checkpoints/ on first run or can be manually downloaded and specified via command-line flags. The implementation uses PyTorch and supports inference on CPU, CUDA GPUs, and Apple MPS devices. Gaussian splat rendering via the included gsplat renderer requires CUDA and initializes on first launch with non-trivial startup overhead.
The output format is OpenGL PLY files containing 3D Gaussian splat parameters compatible with various public 3DGS renderers. Scene geometry is centered approximately at world coordinates (0, 0, +z) using OpenCV conventions (x-axis right, y-axis down, z-axis forward). Metric scale is preserved throughout the representation, enabling accurate camera movements with absolute spatial relationships maintained. The model demonstrates zero-shot generalization across diverse datasets without task-specific training.
Key technical constraints and features:
- Input: single 2D image (URI format via Replicate API)
- Output: 3D Gaussian splat representation in PLY format
- Inference time: sub-second on standard GPU
- Video rendering: CUDA GPU required; CPU/MPS support only for Gaussian generation
- Coordinate system: OpenCV convention with metric scale
- Generalization: zero-shot across datasets
- Performance vs. prior work: 25–34% LPIPS improvement, 21–43% DISTS improvement, three orders of magnitude faster inference
Model inputs and outputs
Inputs
- image (URI string, required): Single input image for view synthesis. Format must be a valid image file accessible via the provided URI.
Outputs
- Output (URI string): Uniform resource identifier pointing to the generated 3D Gaussian splat representation in PLY format, ready for rendering with compatible tools.
Getting started
import replicate
client = replicate.Replicate()
# Prediction with a single input image
output = client.run(
"kfarr/sharp-ml:2be5eda5c0ed5eee96e94f7391fe407d070be0d0aa567437fe7940133c789c21",
input={
"image": "https://example.com/path/to/input/image.jpg"
}
)
print(output)
The image parameter accepts a URI pointing to your input photograph. The model returns a URI to the generated Gaussian splat file in PLY format, which you can download and render using compatible third-party tools or visualization software.
Frequently asked questions
Q: What image formats and resolutions does SHARP accept?
A: The Replicate API accepts image URIs; the underlying model details regarding supported formats (JPEG, PNG, etc.) and maximum resolution are not explicitly documented in the available schema. Standard web image formats are recommended.
Q: Can I use the output Gaussian splats in standard 3D software like Blender or Maya?
A: SHARP exports 3D Gaussian splats in PLY format, which is a point cloud format rather than traditional mesh geometry. Direct import into Blender or Maya requires specialized plugins or conversion tools; however, the splats are compatible with various public 3DGS renderers specifically designed for this representation.
Q: What is the commercial license for SHARP, and can I use it in production?
A: The model is released under Apple's specific license (see LICENSE_MODEL in the GitHub repository). Review the license file before deploying in commercial applications, as restrictions may apply depending on use case and jurisdiction.
Q: Why does video rendering require CUDA when Gaussian splat generation supports CPU and MPS?
A: The gsplat renderer (used for video trajectory generation) has compute-intensive requirements optimized for CUDA architecture. CPU and MPS support for Gaussian generation is sufficient for inference; rendering the splats into video frames leverages GPU-accelerated rasterization available only through CUDA at this time.
Q: How does SHARP handle scenes with significant occlusions or thin structures?
A: SHARP may produce artifacts or missing geometry for complex occlusions and thin structures because it infers 3D geometry from a single 2D image without explicit depth information or multi-view consistency. Scenes with simple geometry and clear texturing produce the best results.
Q: What coordinate transformations are required if I render SHARP's output in a non-OpenCV engine?
A: SHARP uses OpenCV conventions (x right, y down, z forward) with the scene centered around (0, 0, +z). If your renderer uses a different convention (e.g., z-up in game engines), you must manually rotate and scale the Gaussian splats before rendering to achieve correct alignment.
Q: Is SHARP still actively maintained, and when was the latest version released?
A: The latest version was deployed on February 24, 2026, and the model remains publicly available on Replicate. Active maintenance status depends on Apple's ongoing research priorities; refer to the GitHub repository for the current development state.
Top comments (0)