Photogrammetry, Gaussian Splatting, DFX SplatCore

Anyone who engages with 3D from photos stumbles at once over concepts that seem to mean the same thing and do not. Photogrammetry? Gaussian Splatting? NeRF? And where does Splat Studio stand? The confusion is understandable — the concepts come from two different worlds that found one another only recently. This chapter sorts them out, tells the short, wild story of the newcomer — and ends with the question that really counts here: what does it mean for your photos?

2.1  Photogrammetry — the art of measuring

Photogrammetry is the direct continuation of the story from Part 1. Its workflow, roughly sketched: from many photos the camera positions are computed back (the resection), then for countless millions of image points their location in space is determined by triangulation — a point cloud arises, over which a wire mesh (the “mesh”) is stretched, onto which finally the photos are pulled as a surface image.

The result is true to scale — one can measure within it. That is why photogrammetry remains to this day the tool of choice wherever numbers count: surveying, construction, cartography, heritage conservation. Its weaknesses lie on the other side: it needs surfaces with structure (smooth, glossy or transparent things fail), fine details such as hair or grilles fall apart, and the photorealism has a limit — the texture is glued onto the geometry. From close up, in raking light, at soft transitions you can see it.

2.2  NeRF — the scene as an answering machine

In 2020 a research group presented a method that tackled the task in a completely different way: NeRF (“Neural Radiance Fields”). No point cloud, no wire mesh — instead a neural network is trained to answer a single question: “I stand here and look over there — what do I see?” For each desired image this question is asked a million times along lines of sight, and from the answers the view arises.

The results were a sensation: photorealistic free views, soft materials, moods of light — things at which photogrammetry's glued-on texture fails. The price, however, was brutal: training over hours to days, and even the finished model needed seconds to minutes per image. Beautiful — and for practical work unusable.

2.3  The race — and its surprising lesson

What followed was a two-year race for speed, and in the middle of it fell an insight that cleared the way for everything else. Methods such as Plenoxels (2022) threw the neural network out entirely and reached nearly the same quality with a simple grid — in a fraction of the time. Instant-NGP (2022) pushed training from days down to seconds. The lesson from this: the magic was never in the neural network. It is in the idea of refining a scene as a whole until it explains all the photos at once. If that works without a network too — then the representation may be simple, fast and directly editable.

2.4  Gaussian Splatting — two old threads unite

In 2023 a paper from France brought together two traditions that had run side by side for twenty years. One comes from computer graphics: as early as 2001 it was known how surfaces can be cleanly represented from soft, elliptical dabs of colour (“splats”). The other is the young radiance-field idea from NeRF. The result is called 3D Gaussian Splatting — and it won the race in both directions at once: NeRF quality at real-time playback.

The idea behind it is gratifyingly vivid. A scene consists of millions of soft dabs of colour in space. Each dab has a location, a shape (from perfectly round to needle-fine and stretched — this way they nestle against edges and surfaces), a transparency and a colour that may change slightly depending on viewing direction. The training works like sculpting in fast-forward: the dabs are drawn, compared with the real photos, shifted, split, deleted, recoloured — tens of thousands of times, until the rendered scene explains the photos.

And here the circle closes back to Part 1, in a sentence that explains the whole architecture of DFX SplatCore: the training starts with the points and camera positions from the old measuring world. First the classical photogrammetry machine computes the skeleton — where the cameras stood, where the first points lie — then the splatting training lays the photorealistic skin over it. The two-thousand-year measuring tradition was not replaced. It became the foundation.

The confluence of 2023. On the right the family branches that have arisen since — Section 2.6 shows which of them works where in DFX SplatCore.

AI image

2.5  The branching — a family arises

After 2023 research exploded, and the one method became a family. Each branch solves a particular weakness of the original: methods for few images (the original needed many), for generation from a single image, for clean meshes and exports into classical 3D programs, for robustness against blurred or difficult shots, for reflective surfaces and for computing the light out of the shots. The names of these methods — InstantSplat, BAD-GS, SpikeGS, FlashGS and many more — no one needs to memorise. But it is worth knowing that they exist. Because:

2.6  … and DFX SplatCore — the orchestra

DFX SplatCore is none of these methods. It is the orchestra in which selected family members play together — each where its strength lies, and without your having to worry about the line-up:

Two things about this table are remarkable. First: the tolerance table of Motion Guard, which judges your images on import, does not consist of invented numbers — behind it stand real methods with real, measured error tolerance. Second: the last row is deliberately honest. Where something is not yet solved, it says so here — and nowhere will you read it more plainly than in Part 7.

2.7  What this means for your photos

From the double nature — measuring skeleton plus reproducing skin — follows an insight that orders all the capture knowledge and that old guidebooks often overlook:

• What stays: Everything the skeleton needs. Parallax, overlap, sharpness — the camera computation is and remains photogrammetry, and it makes the same demands as ever.
• What became milder: The demands on surfaces. Where the old world knew only glued-on texture, in splatting the colour of a point may be view-dependent — silky sheen, soft materials and fine structures succeed considerably better.
• What became stricter: Consistency. The training takes every pixel of your photos seriously — the light included. A shadow that wanders during the series, an exposure that jumps: what once blurred into the texture becomes today part of the model.

Which puts the decisive question on the table: what exactly does such a reconstruction need from your photos? That is Part 3 — the shortest chapter with the greatest effect.