Apple has released SHARP, a model that produces a high-quality 3D Gaussian Splatting reconstruction from a single photograph in under a second. The standard workflow for this kind of capture normally requires dozens of images shot from carefully planned angles. SHARP asks for one.
Points of light instead of polygons
3D Gaussian Splatting, introduced by researchers at Inria in 2023, reconstructs a scene not as a mesh with textures but as millions of tiny coloured points, each with its own shape, opacity and orientation. A renderer blends these points together to produce a photorealistic view that can be explored in real time, at frame rates high enough to run in a browser rather than a dedicated application. Compared with photogrammetry, it skips the laborious construction of triangles and UV maps. Compared with neural radiance fields, it renders far faster, because there is no per-pixel ray marching to compute. That combination — visual quality close to NeRF, speed close to a video game — is why the technique moved out of research labs and into production pipelines within roughly two years.
From research curiosity to standard toolkit
Adoption across 2025 and into 2026 has been broad rather than niche. In film, the production of Superman shipped with dynamic Gaussian Splatting, OTOY has built full path-traced splat support into OctaneRender, and volumetric video specialists such as 4DViews now offer it as an alternative that handles hair, fur and reflective surfaces better than mesh-based capture. In real estate, Zillow was the first major platform to deploy the technique through its SkyTours tool, and CoStar Group answered with exterior splatting support in Matterport 3D Exteriors — a shift that lets a smartphone video replace what used to require a dedicated 360-degree camera rig. Game engines have followed: Unreal Engine now has both free and commercial splat plugins, and Unity integration is maturing on a similar track. Perhaps most relevant for public-facing projects, browser-based renderers such as Spark and Babylon.js 8.0 mean a splat scene can be shared as a plain URL, with no app install and no plugin. According to the industry guide published by Utsubo, professional capture services currently run from roughly $500 to $5,000 or more depending on scope, with custom web integration reaching $5,000 to $50,000, and turnaround from same-day for a single object to one to two weeks for a large space.
One photo, one second: what SHARP changes for archives
SHARP sits apart from all of this because it removes the multi-angle capture step entirely. Feed it one image and it outputs a splat in under a second, according to Apple. For an institution sitting on decades of catalogue photography — a single frontal shot of an object, taken for inventory purposes rather than 3D reconstruction — that photograph suddenly becomes a plausible starting point for a three-dimensional asset, rather than a dead end.
That said, a single image cannot describe what it does not show. It has no information about the back of an object, no true depth measurement, no way to resolve occlusions. What SHARP generates from one photo is a plausible reconstruction, not a measured one. Photogrammetry and multi-angle video capture remain the only route to a geometrically accurate model — the kind a conservator, a structural engineer or an exhibit designer working to real dimensions would need to rely on. Generating from a single image is a different exercise from surveying an object in the field, and the two should not be priced, or specified, as if they were interchangeable.
DIVE’s take
This matters most at the point where a museum or public body writes a tender. Until now, a digitisation brief for 3D content almost always started from a capture requirement: send a photographer or a scanning team, book time with the object, budget for the equipment and the hours on site. If a single archival photograph can now produce a usable splat in seconds, that assumption no longer holds automatically, and commissioning bodies will start asking why a full capture campaign is needed at all when they already have a photo library.
Our answer is that it depends entirely on what the output is for. For rapid previsualisation — testing how an object might sit in a virtual exhibition layout, generating a browsable catalogue for a website, giving a curator something to look at before committing to a full digitisation budget — a SHARP-style single-image reconstruction is a genuinely useful new layer, and a cheap one. It lowers the barrier for smaller institutions that have photographic archives but no budget for a scanning campaign. For anything where accuracy matters — restoration documentation, scientific study, exhibits where visitors will view an object from angles the original photograph never covered — it is not a substitute for on-site capture, and specifying it as one in a procurement document would be a mistake that surfaces later, usually as a redo.
The practical consequence for anyone drafting a capitolato or a project brief is to separate these two use cases explicitly, rather than asking generically for “3D from our archive”. State whether the deliverable needs to withstand close inspection from multiple viewpoints, or whether it exists to give an audience a sense of an object online. The two questions have different answers, different costs, and — as things stand — different techniques behind them. The wider trend toward browser-based splat viewing, visible across real estate and entertainment, reinforces this: public institutions can now plan for web delivery as the default distribution channel for 3D content, without the accessibility barrier that app-based VR used to impose. That part of the shift is straightforward good news. The single-photo generation part is promising, but it needs to be scoped correctly, not oversold.