Apple SHARP in ComfyUI: 3D Gaussian Splatting from a Single Image

Apple just dropped SHARP in December 2025, and it's absurd. One image goes in. A complete 3D Gaussian splat comes out. In under one second. On a standard GPU.

I've been obsessively testing 3D reconstruction tools, and nothing comes close to this speed-to-quality ratio. Most methods need dozens of images and minutes of processing. SHARP needs one image and one second.

Quick Answer: SHARP (SHarp monoculAR sPlatting) is Apple's open-source model that converts a single photograph into a photorealistic 3D Gaussian representation in under one second. ComfyUI-Sharp provides native integration with automatic model downloading and EXIF-based focal length calculation.

Why SHARP Is Different

Traditional 3D Gaussian Splatting requires multiple viewpoints. You take 20, 50, sometimes 100 photos of a scene from different angles. The algorithm triangulates 3D points from the overlapping views.

SHARP needs one photo.

The model was trained to infer depth, geometry, and appearance from monocular cues alone—the same visual understanding humans use to perceive 3D from 2D photographs. Shadows imply shape. Perspective implies distance. Texture gradients imply surface orientation.

Apple's paper showed dramatic improvements:

• LPIPS reduced 25-34% versus best prior model
• DISTS reduced 21-43%
• Generation time: three orders of magnitude faster

That's not incremental. That's generational.

ComfyUI-Sharp Setup

The ComfyUI wrapper by PozzettiAndrea makes integration smooth.

Installation

cd ComfyUI/custom_nodes
git clone https://github.com/PozzettiAndrea/ComfyUI-Sharp
pip install -r ComfyUI-Sharp/requirements.txt

Restart ComfyUI after installation.

Dependencies

You'll also need ComfyUI-GeometryPack for the Gaussian Viewer node:

git clone https://github.com/WASasquatch/ComfyUI-GeometryPack

Model Downloads

Models auto-download on first run. For offline use, manually place sharp_2572gikvuh.pt in ComfyUI/models/sharp/.

SHARP infers 3D Gaussian parameters directly from single-image features

Understanding the Workflows

ComfyUI-Sharp provides two primary workflows:

Workflow 1: Standard/User Input Focal Length

Use this when you know the camera's focal length or want to specify it manually. Connect:

1. Load Image → Image input
2. SHARP CheckpointLoader → Loads the model
3. SHARP Inference → Generates Gaussians
4. Gaussian Viewer (optional) → Preview in ComfyUI

Set focal length manually in the SHARP Inference node. For most smartphone photos, 26-28mm equivalent is reasonable. DSLR varies by lens.

Workflow 2: EXIF Focal Length Extraction

The smarter approach. Use the "Load Image with EXIF" node, which automatically extracts focal length from image metadata.

Most modern cameras embed focal length in EXIF. This gives the model accurate geometric information without manual input.

Connection flow:

1. Load Image with EXIF → Extracts image + focal length
2. SHARP CheckpointLoader → Model
3. SHARP Inference → Uses extracted focal length automatically
4. Gaussian Viewer → Preview

This is my default workflow. More accurate, less guessing.

Output Formats

SHARP outputs 3D Gaussian splats in .ply format. These are compatible with:

• Standard 3DGS renderers: SuperSplat, GaussianSplats3D
• Unity/Unreal plugins: Various community integrations
• Web viewers: Browser-based splat viewers
• NeRF Studio: Import and refine

The output folder contains the generated PLY files. Import them into your preferred 3D viewer or game engine.

Generated 3DGS can be viewed from novel angles and integrated into 3D applications

Practical Applications

Product Photography to 3D

Turn existing product photos into 3D assets. One good product shot becomes an explorable 3D model. Perfect for e-commerce, AR previews, or 3D catalogs.

Scene Reconstruction for Games

Convert concept art or reference photos into rough 3D scenes. SHARP outputs aren't game-ready geometry, but they provide excellent spatial reference for 3D artists.

Archival Photography

Historical photos of buildings, interiors, or objects can be reconstructed in 3D. The single-image requirement means existing archives become usable.

Combine with Trellis 2

SHARP creates 3DGS; Trellis creates meshes. Different output formats for different purposes. Use both depending on whether you need splats (visualization) or meshes (geometry).

Quality Factors

What affects SHARP output quality:

Image Quality: Higher resolution input = more detail in output. Start with at least 1024px on the short edge.

Focal Length Accuracy: Wrong focal length = wrong geometry. Use EXIF extraction when possible.

Scene Complexity: Single objects work best. Complex multi-object scenes are harder to reconstruct accurately.

Lighting Conditions: Even lighting produces cleaner results. Strong shadows can confuse depth estimation.

Occlusion: SHARP infers hidden geometry, but heavy occlusion limits accuracy. What the camera can't see, the model can only guess.

Hardware Requirements

Gaussian Prediction: Works on CUDA, MPS (Apple Silicon), and CPU. The model is efficient enough for all platforms.

Video Rendering: Currently requires CUDA GPU. The --render option for generating novel view videos needs NVIDIA hardware.

For Mac users: prediction works on M1/M2/M3/M4, but video output needs a CUDA box or cloud GPU.

Comparison to Other Methods

SHARP dominates the single-image speed category. If you need maximum fidelity and have multiple views, traditional multi-view methods still win. But for speed and convenience, SHARP is unmatched.

Integration Ideas

With AI Image Generation

Generate an image with FLUX or Stable Diffusion, then immediately convert to 3D with SHARP. One workflow from prompt to 3D scene.

With ControlNet

Use SHARP output as ControlNet depth reference. The 3D structure provides accurate depth maps for consistent multi-view generation.

With Video Generation

Generate a base image, create 3DGS with SHARP, render camera paths, use frames as references for video AI like WAN 2.2.

iOS Integration: Spatial Scenes

Apple commercialized related technology in iOS 26 as "Spatial Scenes." The research that produced SHARP is the foundation for consumer 3D photo features.

This means SHARP-style reconstruction is heading to mainstream devices. Understanding the technology now prepares you for where consumer tools are going.

Troubleshooting

Model not loading: Ensure you have enough VRAM (4GB+ recommended). Check model file isn't corrupted.

Wrong geometry: Focal length is probably incorrect. Try EXIF extraction or adjust manually.

Missing GeometryPack: The Gaussian Viewer requires ComfyUI-GeometryPack. Install it separately.

CUDA errors on render: Video rendering needs CUDA. Use CUDA machine or skip video output.

Low detail output: Input image resolution matters. Use higher resolution source images.

FAQ

Does SHARP work on CPU? Prediction works on CPU but is significantly slower. GPU recommended.

Can I use SHARP for animation? The output is static 3DGS. You can animate the camera around it, not the scene itself.

Is SHARP open source? Yes, Apple released it under standard research license. Check GitHub for specifics.

How does this compare to Apple's Vision Pro spatial photos? Vision Pro uses stereo cameras. SHARP works with single images—fundamentally different approaches.

Can I import SHARP output into Blender? Yes, via community 3DGS import plugins. The PLY format is standard.

What about video input? SHARP is image-only. For video reconstruction, look at DUS4 or similar temporal methods.

Does it work with AI-generated images? Yes, and results are often excellent since AI images have clean, predictable structure.

The Bigger Picture

SHARP represents a shift in how we think about 3D creation. The bottleneck has always been capture—getting enough angles, enough lighting, enough data.

Single-image reconstruction removes the capture bottleneck. Any existing photograph becomes a potential 3D asset. Every AI-generated image can be spatially explored.

For Apatero.com, we're exploring how this changes creative workflows. When 3D is as easy as 2D, new applications become practical. AR product previews, immersive storytelling, spatial content at scale.

Apple open-sourcing this signals their belief in democratizing 3D. The tools are here. The question is what we build with them.