How Apple's SHARP turns one photograph into a moving 3D view

Apple's SHARP makes a photograph behave like a window into a scene: shift the viewpoint and the picture gains a sense of depth. Introduced in December 2025, the research uses a neural network to predict a renderable 3D representation from one image. The attraction is being able to create spatial photographs from pictures that were taken with an ordinary single viewpoint.[1]

The paper first appeared on 11 December 2025. Its 27 February 2026 revision identifies publication at ICLR 2026. This is an explanation of that research and its available software.[2]

Why a scene is more useful than a zoom

Consider a plant standing in front of a window. Zooming into its photograph enlarges the whole image. Moving a camera sideways would make the nearby plant shift against the more distant window. That relative movement is parallax, a depth cue. This simple example explains why an image needs an estimate of spatial structure to respond convincingly to a change of viewpoint.

Gaussian splatting provides a way to draw that structure. Its building blocks are soft-edged blobs positioned in three-dimensional space. Their shape, orientation and opacity control how they contribute to the picture; rendering projects them onto the image and blends their overlapping colour contributions in visibility order. The original 2023 method established a fast approach to representing and rendering scenes from multiple photographs. The representation gives a renderer something explicit to work with when the camera moves.[6]

SHARP predicts two depth layers, then refines the positions, shapes, colours and opacity of roughly 1.2 million Gaussians. It builds on Depth Pro features. This happens in one network pass; subsequent viewpoints reuse the resulting scene.[2]

That separation is the practical attraction. Scene creation is a preparation step, while the renderer answers each new camera position. A viewer can therefore interact with the resulting spatial representation instead of receiving only one predetermined animation. Apple describes the output as having metric scale, allowing virtual camera movements to be expressed in physical units.[1]

The speed claim has two separate measurements

The researchers measured scene creation and rendering separately on an NVIDIA A100 GPU. Rendering used each dataset’s native resolution.[2]

These figures describe different workloads. A fast initial conversion does not imply that every subsequent image, at every output resolution, has the same rendering cost.

What a researcher can run today

Apple provides code and downloadable weights. The weights use Apple's research licence, which permits scientific and academic investigation without commercial exploitation. It explicitly excludes product development and use within commercial products or services. That scope matters when choosing what to build with the release.[4]

The code has separate Apple software terms allowing modification and redistribution subject to their conditions. Those permissions do not expand the model licence.[5][4]

For an eligible research exercise, the repository documents a route from a folder of images to saved Gaussian scenes:[3]

1. Obtain the repository, create and activate the recommended Python 3.13 environment, then install its requirements.txt dependencies.
2. Run sharp --help to check that the command is available.
3. Place input photographs in a folder and run sharp predict -i /path/to/input/images -o /path/to/output/gaussians. The checkpoint downloads automatically on first use.
4. Open the resulting .ply Gaussian files in a compatible Gaussian-splat renderer. The repository notes that another renderer may require recentering or rotating the scene.

Prediction supports CPU, CUDA and MPS. The supplied video-trajectory path has a narrower requirement: a CUDA GPU. On that hardware, adding --render to prediction creates trajectory videos; sharp render can also use previously generated Gaussian files. The first renderer launch includes initialization work.[3]

The revealing test is a small camera movement

The intended experience is nearby views, not a complete reconstruction of hidden geometry. The paper targets movements around half a metre and documents depth failures involving macro photography, stars and reflections.[2]

The project page offers comparison videos as an accessible starting point. A useful way to inspect this type of result is to follow one foreground edge as the camera moves, then examine the background appearing beside it. Checking those edges reveals whether the scene remains coherent through the movement. The suggested exercise makes the central technical question visible without requiring the reader to install anything.[7]

Report an error or suggest a correction