DFX SplatCore decides on the basis of the image count which reconstruction path works — and each path makes different demands on your capture. That is why practice begins not with the camera but with a simple question: what do you have — or what can you obtain? A single image? A handful? A planned series? A video? Each of these starting points is played through here: what happens internally, what your photos must achieve, and which mistakes are typical.
The signpost at import. Video, drone and mobile phone are not paths of their own but suppliers to these three — each with its own pitfalls.
With a single photo triangulation is impossible — a single witness cannot establish depth. What works here is something fundamentally different: a trained model that has learned from millions of examples how things are shaped fills in the unseen. The front comes from your photo — the back is a plausible, learned guess.
This must be honestly placed: the result is a credible showpiece, not a documentation. For a quick product preview, an idea, a decorative object the path is a gift; for archiving, evidence or precision work it is the wrong tool — there no way leads past a real series. To make the guess turn out well, give the model the best possible single image: a free-standing object with a clear outline against a calm background, even light without hard shadows, the whole object sharp, and a viewing angle that reveals the form maximally — usually diagonally from above front, so that three sides are visible at once.
From the second image on the real reconstruction begins — but with few witnesses each single one becomes precious. The basic rule of this path: spread out rather than bundle. Four photos that differ by 90 degrees each tell more than eight photos from almost the same corner — because only the angular difference produces parallax, and only parallax produces depth. Anyone who gives away two nearly identical positions in a small series has effectively one image fewer.
As orientation: two to three images suffice for the flat with little depth (a relief, a façade head-on), six to eight for a compact object diagonally from above on all sides, ten to twelve for a full walk-around. And the honest word to go with it: what no image has seen, no one can establish here either — a back without a photo remains a gap or a guess. The sparing path is built for the quick, light entry; the maximum of fidelity is always delivered by the full series.
From about a dozen images the classical path from Parts 1 and 2 takes over — and now the task reverses: no longer is every vantage point precious, but rather the order of the many. Plan your tracks before you shoot: for an object the circuit in 10–15-degree steps (Part 3.2) plus a second, higher ring; for a room the walk along the walls with regular steps and a view towards the centre. Work calmly, shoot evenly, never let the chain break.
And a number that surprises: more is not automatically better. Three hundred photos, half of which are almost identical, deliver no better model than a hundred well-distributed ones — they only cost computing time and storage. Quality and distribution beat mass. On import Motion Guard checks every image anyway and flags the weak ones — you may (need not) then sort them out yourself; the decision remains, as always, with you.
A slow pan around the object, and the video delivers thousands of single frames — tempting, and DFX SplatCore accepts videos too. But the physics sets conditions one should know. A video at 25 frames per second typically exposes each frame for about 1/50 second — in motion that is an invitation to motion blur, and blurred witnesses you know from Part 3.3. Add to that the rolling-shutter effect (the sensor reads out line by line; fast movement bends verticals) and video compression, which irons away fine details before anyone can measure them.
If video, then like this: move slowly (noticeably slower than feels right), plenty of light (so the automatic can expose briefly — or force the shutter speed manually to 1/250 s and shorter), the highest available resolution, and rather two calm circuits than one hectic one. The honest ranking remains all the same: a deliberately photographed series beats a video almost every time. The video is the convenient path — not the best.
The drone solves the oldest problem of surveying: the unreachable position. Roofs, façades from the second floor up, terrain, monuments up high — what forced Meydenbauer onto the rope, an aircraft handles today. For the capture tracks the same six needs apply, in two proven patterns: the grid (parallel tracks with 70–80 percent overlap, camera vertical or slightly tilted) for terrain and roofs, and the orbit (circular flight around the object, camera on the target) for buildings, towers, sculptures. An often overlooked gain: for three-dimensional façades the oblique view at 45 degrees is usually more valuable than the vertical view from above, because it bears witness to walls and roof at once.
Technically the video warnings apply with added force (rolling shutter, movement): fly slowly, or better still: stop briefly for every photo. Single photos while hovering clearly beat worn-out video frames. The GPS data in the image files does no harm and helps the matching. And a marginal note without legal advice: where flying is permitted is governed by local regulations — clearing that is part of the preparation, like the battery.
AI image
First the gift, which hardly anyone recognises as such: the small sensor of a phone produces a huge depth of field — at typical shooting distances practically everything is sharp, front to back. That is exactly the ideal from Part 3.3 that owners of large cameras have to stop down laboriously to achieve. On top of that the phone is always with you. For many objects it is a fully-fledged, in part ideal, capture device.
Then the trap, and it is treacherous because it is invisible: computational photography. A modern phone rarely photographs one image — it computes several shots into one in a flash (HDR, night mode, sharpening and beauty filters). Every photo is then a different, invisible collage — and the consistency from Part 3.4 breaks, without it showing on any one image. From this follow the phone rules:
