Anyone who searches forums for capture tips for 3D finds much self-confidence and little reasoning. Some of it is true, some was true once, some was never right — and all of it is passed on equally. This chapter takes on the eight most stubborn claims. The yardstick is always the same: the six needs from Part 3. No authority, no tradition — only the question of what the reconstruction really needs.
This sentence comes from a different discipline. In portrait and available-light photography the large sensor has real advantages: creamy bokeh, less noise in the dark. Only: for the reconstruction the criteria reverse. The desired image from Part 3 is continuous depth of field — and there the smaller sensor is physically ahead: at the same f-number it delivers more depth of field. A Micro Four Thirds camera at f/4 renders as deeply sharp as a full-frame camera only at f/8. What counts as a weakness in the portrait forum (“no subject isolation possible”) is a built-in advantage here.
The honesty belongs with it, in both directions: camera shake weighs more heavily on the small sensor (the same tremor fills a larger share of the frame — Motion Guard factors this in correctly with the crop factor and judges small sensors handheld correspondingly more strictly), and in low light it noises earlier. Both are answered with light and technique, not with a different sensor. There is a reason why compact formats such as Micro Four Thirds are taken for granted in industrial and surveillance systems: they deliver robust, deeply sharp, repeatable images. The reconstruction does not ask which sensor made the photo — it asks whether it is sharp, distributed and free of contradiction.
AI image
Resolution helps — up to a point. More pixels mean more fine features that can be recognised. But: the training computes internally with a limited image edge anyway (also to keep memory and time within bounds), and beyond that limit every further megapixel fizzles out. What is decisive is not the quantity of pixels but the sharpness per pixel: a shaken 60-megapixel image is a worse witness than a crisply sharp one with 20. Whoever wants to upgrade upgrades on light, stability and care — that is where model quality lives.
What an expensive lens buys — wide-open sharpness, beautiful bokeh, low distortion — is surprisingly unimportant for the reconstruction. Bokeh is unwanted here (Part 3.3), distortion is corrected computationally (DFX SplatCore knows the correction data of thousands of lenses and hands it to the camera computation as prior knowledge), and stopped down to middle apertures even modest lenses are respectably sharp. What really counts costs nothing: the same lens for the whole series, a clean front element, a sensible working aperture. The budget is better spent on continuous light or a stable tripod than in the lens cabinet.
Seductive, and wrong three times over. First, the object becomes smaller in the wide image — fewer pixels per object detail, weaker witness testimony. Second, extreme wide-angle distorts in the corners and up close. Third, and most importantly: fewer photos was never the goal — variety of vantage points is (Part 3.1). The useful range lies in the middle: moderate wide-angle to slight tele, and more positions instead.
The special case fisheye deserves precision, because it is often banned outright: the raw fisheye image does indeed break the window model from Part 3.4 — its projection bends straight lines, whereas the camera computation assumes straight lines. But: some systems (for example at Olympus/OM System) can convert the fisheye projection into a normal, rectilinear ultra-wide-angle already in the camera — and such a corrected image is again a valid window and thus usable. New since July 2026: DFX SplatCore recognises a fisheye lens from the device database and says so, instead of quietly computing with the wrong camera model. Two caveats: the outermost corners lose some resolution in the correction, and the correction must remain identical across the whole series. So no ban — a transformation with conditions.
Both camps are wrong, because both talk about the furniture instead of the physics. The only question is: does the camera move while the shutter is open? Handheld with a short shutter speed in good light: perfectly fine — and for coverage even an advantage, because you reach many positions faster. Long shutter speeds, stopping down, interiors, low ISO: tripod. It is an arithmetic problem, not a matter of faith — and Part 6.4 works it through with you.
For DFX SplatCore users a reassurance comes with it: Motion Guard judges tripod shots fairly. A 30-second exposure looks hopelessly shaken by the pure formula alone — the image check therefore measures back against the pixels, recognises the actual sharpness and treats the shot as what it is: stabilised. Only one thing more, anticipating Part 6.2: the image stabiliser belongs switched off on the tripod.
RAW is indeed the right choice (more on this in Part 6.6): more reserves in shadows and highlights, lossless white balance, full control over the development — modern converters pull visibly more detail from a clean RAW series. But “save everything” is a lie with three names: blur (detail never recorded stays never recorded), missing parallax (no slider creates new vantage points after the fact) and broken chains (the missing back is in no RAW bit). And the consistency rule travels along into the development: the same settings for the whole series, no geometry interventions whatsoever.
The most understandable fallacy on the list, for geometrically it seems to come out the same: whether the camera turns around the object or the object in front of the camera, the angle changes just the same, doesn't it? For the reconstruction it does not — and Part 3 says why. For the witnesses see everything: if only the object turns, then the entire background on every photo bears witness that the camera did not move — while the object claims the opposite. Two statements, one contradiction. Worse still: light and shadow do not turn with the object — the shadowed side wanders across the surface, and multi-view consistency (3.5) breaks along with it.
Still, the turntable is not banned outright — it merely demands professional conditions: a completely structureless, even background (which the software can exclude as a witness — DFX SplatCore masks cut-out objects), plus even, shadow-free light all around. Whoever cannot reliably create these conditions chooses the dependable path: the object stands, you move. The world in the background then turns from a nuisance into an additional witness.
The most convenient myth of all, and it deserves a precise answer in two lists. What the software actually compensates for: moderate exposure differences, lens distortion, individual weak images in a strong series (Motion Guard finds and flags them), baked-in light mood within limits (de-lighting). That is a lot — more than any generation of tools before could. What physically no one compensates for, not today and not tomorrow: detail that was never recorded; vantage points that were never taken; backs that no photo ever saw; chains that have broken. Software amplifies information — it does not create it.
Exactly along this boundary runs the division of labour on which these pages rest: everything on this side of the boundary is handled by DFX SplatCore. Everything beyond it is your craft — and it is learnable, plannable and, with every object anew, satisfying. How to set it concretely on the device is the next part.
