Six needs. That is all. Every capture rule, every piece of advice, every myth you meet in forums can be measured against these six needs — and that is exactly what the later chapters will do. Anyone who understands them no longer has to learn a list of rules by heart: they can derive every rule for themselves and, more importantly, decide for themselves when it does not apply.
One image beforehand that carries all the rest: a reconstruction is a trial by evidence. Every photo is a witness, the software is the court. No witness has seen the whole truth — but when the statements overlap, are free of contradiction and sharply worded, the truth follows inevitably. A witness who lies or mumbles does not just render their own testimony worthless: they damage the whole trial.
Do the oldest experiment in the world: hold your thumb in front of you, close your left and right eye in turn. The thumb jumps back and forth against the background — and all the more strongly the closer it is to your face. This jumping is called parallax, and it is the only source from which the reconstruction gains depth: what is near shifts strongly between two vantage points, what is far hardly at all. From the difference the distance is computed — that is the triangulation from Part 1, applied to every image point.
From this follows perhaps the single most important rule of all, and it contradicts a deep photographic instinct: move — do not zoom. Whoever stays on the same spot and zooms in produces larger images of the same vantage point, but no parallax whatsoever. And whoever merely turns on the same spot produces a panorama — pleasant to look at, but worthless for the depth computation, because here too nothing jumps. Depth arises solely when the camera changes its position: sideways, in an arc, around the object.
Both vantage points see the same near point — but in front of completely different parts of the distant backdrop. From exactly this offset the depth is computed.
For the court to be able to compare statements, the witnesses must have seen one another: two neighbouring photos must share the same slice of the world, otherwise there is nothing to compare. As a rule of thumb, 60 to 80 percent overlap between neighbouring images has proven itself. Why not more? With nearly identical images the parallax is missing — the witnesses all say the same thing, the court learns nothing new. Why not less? Because then the matching breaks off.
For a photo series is a chain: image 7 hangs on image 6, image 8 on image 7. If a link is missing — a 90-degree gap in the walk around the object, a forgotten section of the room — the series falls apart into two islands that the software can no longer relate to one another. The result is not a somewhat worse model, but a half one, a distorted one, or none at all. In numbers, for the most common case: a full walk around an object in steps of 10 to 15 degrees gives 24 to 36 photos per circuit — plus a second, higher circuit for the top-down view where the form demands it.
The circuit seen from above. Each camera position shares its view with the neighbours — if the chain breaks anywhere, the model falls apart into islands.
The matching between the images runs over recognisable features: corners, edges, grains, spots — everywhere the image differs locally and unambiguously from its surroundings. Blur smears exactly this unambiguity. And so a blurred image is not merely useless but actively harmful: it delivers to the court few, ambiguous statements that can be wrongly matched — and a wrong match falsifies the computed camera position, which in turn poisons all statements of that witness. A single mumbling witness can bend a clean chain of evidence.
Making matters worse: blur is not curable after the fact — no software in the world reconstructs detail that was never recorded. That is why DFX SplatCore checks every image already on import (Motion Guard, with its own manual in this library) — and why the ideal for the shot is continuous sharpness: as much as possible the whole object within the zone of sharpness, not just a narrow plane. How to set that on the camera is covered by Part 6 — here the why is enough: every blurred area of a photo is an area about which that witness can say nothing.
Recall Alberti's window from Part 1: the camera is a plane with exactly describable properties — focal length, image centre, distortion. The reconstruction treats your entire series as views through one and the same window. Any change to this window in the middle of the series is as if the witness had swapped glasses partway through without telling the court: all their angle statements no longer fit together.
Concretely this means — and each of these rules follows inevitably from the one image:
Beneath everything lies an assumption so self-evident that it is rarely spoken aloud: a point of the world looks (nearly) the same from every vantage point. The brown grain on the wooden handle is brown grain — no matter whether you look from left or right. Only under this assumption may the court compare statements at all.
Matte surfaces obey it obediently. Three things break it: reflection (the highlight travels with the viewer — every witness sees it elsewhere), transparency (one sees through the point onto changing backgrounds) and movement (the point is simply no longer there for the next witness). The consequences in the model: ghost points, mush, artefacts floating in the air. An honest nuance belongs here: splatting can model mild view-dependence — silky sheen, half-matte and shimmer succeed astonishingly well. The hard mirror and clear glass remain the open opponents; Part 7 treats them without varnish.
Why the reflective is hard, in one image: the highlight travels with the vantage point. Two witnesses, two statements — and both tell the truth.
The training takes every pixel seriously — the lighting included. This has two consequences one must know. First: light becomes part of the model. A hard cast shadow on the object is understood not as a shadow but as a dark spot of the surface — it is afterwards baked in and travels along in the finished 3D model, no matter how you turn it. Soft, even light is therefore not a matter of taste but a matter of model quality: an overcast sky is a gift for this work, a softbox its substitute in the studio.
Second: the light must not change during the series. If the sun emerges from behind clouds, the auto-exposure jumps, the shadows wander — then the witnesses contradict one another, for the same spot is bright on image 4 and dark on image 19. The special case that follows and surprises many: the on-camera flash is poison for this work. It travels with the camera — every photo is lit differently, every photo contradicts every other. Continuous light instead of flash, diffuse instead of hard, constant instead of changing: that is the whole doctrine of light in a single line. (DFX SplatCore can compute lighting out after the fact — the de-lighting in the Orbit module — but the old rule holds here too: capturing cleanly beats repairing cleanly.)
AI image
With that the foundation is laid: parallax, overlap, sharpness, consistency, multi-view consistency, light. Everything that comes now — the starting points, the myths, the camera settings — is the application of these six needs to practice. You have the equipment to test every claim for yourself.
