The two working checklists stand with their chapters: the settings checklist at the end of Part 6, the “no-second-attempt” checklist in Part 8.2. Here in addition the third — the quick check, before you even set up:
| Parallax | The “jumping” of near things against a distant background when the vantage point shifts — the only source of depth (Part 3.1). |
| Triangulation | Determining the location of a point from two positions and two angles — the primal principle since antiquity (Part 1.1). |
| Resection | Computing back from an image where the camera stood (Lambert 1759 — Part 1.3). |
| Photogrammetry | Measuring from photographs: camera positions, point clouds, meshes — true to scale (Part 2.1). |
| Structure from Motion (SfM) | The automated process that computes from many photos simultaneously all camera positions and first spatial points (Part 1.5). |
| COLMAP | The free standard software for SfM — works in DFX SplatCore on the many-image path (Part 4.3). |
| NeRF | “Neural Radiance Field” (2020): a scene as a trained answering machine — photorealistic, but slow (Part 2.2). |
| Gaussian Splatting | A scene from millions of soft, stretched dabs of colour (“splats”), displayable in real time (2023 — Part 2.4). |
| Multi-view consistency | The silent basic assumption that a point looks the same from everywhere — violated by reflection, glass and movement (Part 3.5). |
| Mesh | A surface of connected triangles — the format of the classical 3D world and of game engines (Part 10.3). |
| GLB | Widespread 3D file format for web and shops — an export from DFX SplatCore (Part 9.6). |
| PBR / de-lighting | Material description for realistic lighting — and the computing out of the capture light, so that the object fits in any scene (Part 10.3). |
| EXIF | The capture data every camera writes into the image (time, aperture, ISO, focal length, model) — the basis of the image check. |
| Crop factor | By how much a sensor is smaller than full frame (MFT ≈ 2.0; APS-C ≈ 1.5). Multiplies the shake risk — and grants depth of field (Part 5.1). |
| Equivalent aperture | F-number × crop factor: makes depth of field comparable across sensor sizes (f/4 on MFT ≈ f/8 on full frame). |
| Diffraction | Physical softening at very small aperture openings — the reason why f/22 is not sharper than f/8 (Part 6.3). |
| Rolling shutter | Line-by-line readout of the sensor — bends vertical lines under fast movement; important with video and drone (Part 4.4). |
| Computational photography | Phones compute several shots invisibly into one image (HDR, night mode) — breaks the consistency of the series (Part 4.6). |
| Motion Guard | The image check of DFX SplatCore: judges every series on import, discards nothing — developed by Studio Heitzig (Part 6.7). |
| Orbit | The presentation module: hotspots, galleries, audio guide, multilingualism, offline export (Part 8.5). |
| Structured data | Machine-readable object and product description in the export — makes twins understandable to search engines and AI assistants (Part 9.7). |
This work teaches capturing — two sister documents in this library deepen the check: the Motion Guard guide explains in detail how the image check judges — shake factor, tolerances, tripod detection, log. The device overview lists all cameras, lenses and phones of the recognition database, with a search field. Whoever reads on there after this work understands every line of the finding.
The scientific milestones on which Parts 1 and 2 rest are published research: Surface Splatting (Zwicker, Pfister, van Baar, Gross, 2001), Photo Tourism (Snavely, Seitz, Szeliski, 2006), COLMAP (Schönberger, Frahm, 2016), Neural Radiance Fields (Mildenhall et al., 2020), Plenoxels and Instant-NGP (2022), 3D Gaussian Splatting (Kerbl, Kopanas, Leimkühler, Drettakis, 2023) as well as 2D Gaussian Splatting (Huang et al., 2024). Historical statements on Lambert, Laussedat and Meydenbauer follow the photogrammetric literature.
Image labelling: All illustrations marked “AI image” were generated with an image generator and show staged symbolic pictures, not real persons, places or products; any incidentally recognisable brand or product depictions are unintended, the brands belong to their owners. Explanatory graphics are drawn by hand. Measured values, logs and future program screenshots are never AI-generated.
Photography for 3D is part of the DFX SplatCore Library. Concept, text and design: Studio Heitzig.
Labelling of AI images in this work Example images may be AI-generated — such as symbolic pictures or historical illustrations. Every such image visibly carries the badge “AI image” directly on the illustration. Photographs of real objects and all measured values, logs and screenshots from DFX SplatCore are, by contrast, never AI-generated. Explanatory graphics (timeline, diagrams) are drawn by hand.
The quick answers to the questions that run across all the parts. Each points to where it stands in full — for everyone who first wants to solve a concrete problem and read on later at leisure.
It depends on the object, but as rules of thumb: a compact object from all sides needs a circuit in 10–15-degree steps — that is 24 to 36 images — plus a second, higher ring for the top-down view. Flat things (a relief, a façade head-on) go with considerably fewer.
The most common error is to squint at the quantity instead of the distribution: a hundred well-distributed images beat three hundred, half of which are almost the same. In full in Part 4 and Part 3.2.
Best of all the one you already have — in all seriousness. For the reconstruction it is not the sensor that counts but whether the image is sharp, well distributed and free of contradiction. A smaller sensor (such as Micro Four Thirds) is often at an advantage here, because it delivers more continuous depth of field (Part 5.1).
If you do want to invest, put the money into light and stability (continuous light, light tent, a solid tripod) instead of an expensive lens — that is where model quality lives (Part 5.3).
For many objects yes, in part even excellently — the small sensor makes almost everything continuously sharp. Three things you must heed: switch off night mode and HDR (they compute several images invisibly and break the consistency), use RAW if possible, and stick to one lens — 0.5×, 1× and 3× are different cameras. Details in Part 4.6.
Almost always one of the six needs from Part 3. The quick diagnosis: holes or missing areas → no photo looked there, or the chain has broken (more overlap, no gaps, Part 3.2). Wavy or holey smooth surfaces → too little structure, the feature search finds no anchors (Part 7.3). Milky veils or floating shreds → reflection, glass or something moved (Parts 7.1, 7.2, 7.5). Everything slightly washed out → blur in the series (Part 3.3).
Sometimes — it depends on whether what the reconstruction needs happens to be met: were the images taken from different positions (not just zoomed or turned)? Do they overlap? Are they sharp and consistently exposed? A series you photographed by circling can work; six pretty product photos from the same corner cannot. When in doubt: send it through the import and read the Motion Guard finding (Part 6.7).
Exposure, colour, denoising, sharpening — yes, gladly, as long as the same settings act on the whole series. Taboo is everything that changes the geometry: no cropping, no rotation, no perspective correction, and no retouching in single images (that turns a witness into a liar). Distractions are removed on the finished 3D model, not in the source images. In full in Part 6.6.
No. Reflective and transparent objects are a known hard case — for physical reasons, not from clumsiness (the highlight travels with the viewer, Part 3.5). What helps: a light tent that turns hard reflections into soft shimmer, and photographing more densely (Part 7.1). Clear glass is honestly not yet a showcase discipline today. Research is working exactly here — and your cleanly captured series stay usable when the time comes.
Photographing an object is a matter of minutes. How long the reconstruction then computes depends on image count and computer — that is a matter for the tool manuals of this library, not of this work. Just remember the order: the time you put into good capture is the only one that can never be recovered (Part 8.3) — the computing time runs overnight too.
Zooming instead of walking. Whoever stays on the spot and zooms in (or merely turns) produces no parallax — and without parallax no depth (Part 3.1). The second is the broken chain: a gap in the circuit, a skipped angle, and the model falls apart (Part 3.2). Both cost nothing but the knowledge you now have.
