RealityCapture review 2018: photogrammetry workflow and results

Summary

This RealityCapture review 2018 is an archival refresh of 3D Mag’s hands-on article. The original page was published on March 15, 2018, but it now also shows an Updated July 28, 2026 timestamp. That newer page date can make an old review look current at a glance, even when its tests, pricing, and software context are historical. [1]

Read in that historical frame, the review still gives a clear picture of what impressed the author at the time. RealityCapture felt fast, produced strong textured results on some subjects, and benefited from a more polished workflow in the reviewed update. But one caveat has to stay attached to every result claim: the article’s Results section says the photo sets were not reprocessed for the 2018 update, and the models shown there were made with the June 2017 version of RealityCapture. This is useful as a dated workflow and output review, not as a clean 2018 re-benchmark and certainly not as a 2026 buying guide. [1]

What this refresh is (and isn’t)

This refresh keeps the original review’s test envelope visible, fixes timeline confusion, and uses current RealityScan documentation only where it helps explain general photogrammetry concepts or 2026 status. It does not treat a 2018 article as evidence for present-day features, licensing, or competitive standing.

• Is: an archival interpretation of a 2018 review with corrected dates, naming, and source matching.
• Is: a photogrammetry-focused explanation of what the review observed versus what is being explained conceptually.
• Is: a short 2026 status note on Epic ownership, naming, and licensing context.
• Is not: a new 2026 RealityScan feature review.
• Is not: a current pricing guide or current product ranking.
• Is not: a deep dive into neural reconstruction, NeRFs, or Gaussian splatting.

Timeline: key dates you must keep straight

• February 4, 2016 — CG Channel reports RealityCapture’s commercial release. [3]
• June 2017 — the models shown in the review’s Results section were made with a June 2017 version of RealityCapture. [1]
• March 15, 2018 — 3D Mag publishes the review page. [1]
• October 6, 2018 — Agisoft’s Metashape 1.5.0 pre-release thread appears, marking the late-2018 rename transition from PhotoScan. [12]
• December 31, 2018 — Agisoft Metashape 1.5.0 build 7492 appears in the changelog. [13]
• March 9, 2021 — Epic Games announces that Capturing Reality has joined Epic. [15]
• June 17, 2025 — RealityCapture is rebranded under the RealityScan name for desktop. [16]
• May 12, 2026 — RealityScan publishes the legacy licensing server notice, which states a shutdown date of August 3, 2026. [17]

RealityCapture’s “public release” context (2016) — without relying on 3D Mag alone

3D Mag was not reviewing an unreleased prototype in 2018. CG Channel’s February 4, 2016 update reported RealityCapture’s commercial release and described the software as a tool for building meshes from photographs and laser scans, with export in formats such as PLY and OBJ. That independent 2016 context matters because it places the later review in a market where RealityCapture had already moved beyond open beta and was being discussed as a practical production tool rather than a lab demo. [3]

The 3D Mag article picks up the story from a later point in the product’s history. Its value is not that it establishes the software’s first public appearance, but that it records how the software felt in a specific hands-on environment tied to a particular machine, a particular update, and a result set that partly predated the article itself. That is the right frame for reading a RealityCapture photogrammetry software review from 2018 without treating it as a timeless verdict. [1]

RealityCapture review 2018: what was actually tested (and what wasn’t)

The review’s test envelope was specific. 3D Mag says the software was run on an Alienware Aurora R5 with an i5 dual-core CPU, 32 GB of RAM, and an Nvidia GTX 1070. It also describes a workflow that could be driven through RealityCapture’s Start button for an automated or semi-automated pass through alignment and reconstruction. The same review ties its 2018 update to RealityCapture 1.0.3.3939, and highlights that version as part of an improved editing and selection experience. [1]

That does not mean every result on the page was freshly regenerated under that exact 2018 update. The article’s note in the Results section says the photo sets were not reprocessed for the 2018 update and that the models shown there were made with the June 2017 version of RealityCapture. Readers therefore need to keep two layers separate: the article was updated around 1.0.3.3939, but the illustrated result set was not fully rerun for that update. The official release-notes history for 1.0.3.3939 RC is still useful context because it documents highlights such as the RealityCapture PGM perpetual photogrammetry license, a new selections experience, a new map view experience, and a new control points experience. Those release-note items help explain why the refreshed review talks about workflow improvements, but they do not turn the page into a controlled regression test of that version. [1] [2]

Facts that must travel with any performance/result claim

• The page was published on March 15, 2018 and now shows Updated July 28, 2026. [1]
• The review states that the photo sets were not reprocessed for the 2018 update and that the displayed result models were made with the June 2017 version. [1]
• The update is tied to RealityCapture 1.0.3.3939, while the official release-note history documents features such as RealityCapture PGM, new selections, map view, and control points. [1] [2]
• The hardware was an i5 dual-core CPU, 32 GB RAM, and GTX 1070. [1]
• The pricing quoted on the page was archival: RealityCapture PGM $2,800 until April 30, then $4,000; full RealityCapture $15,000; Promo $99 for 3 months with a 2,500-photo limit; Steam $39.99 for 1 month. [1]

The photogrammetry pipeline (concepts, not UI click-path)

NIST’s ASTM-derived glossary defines photogrammetry as the art, science, and technology of obtaining reliable information about physical objects and environments through recording, measuring, and interpreting photographic images. That definition is useful here because it separates the topic of this review from active range-sensing systems: the geometry is inferred from overlapping photographs and camera relationships, not directly measured by a projected-light or laser-ranging device. [14]

In general photogrammetry terms, the pipeline starts with image alignment, where the software matches features across photographs, estimates camera positions, and refines camera calibration. In PhotoScan’s 1.4 manual, that stage produces a sparse point cloud and a set of camera poses. After that comes a denser reconstruction stage, which different software families describe with terms such as dense point cloud or depth-map-based reconstruction, followed by mesh reconstruction, cleanup or simplification, UV unwrap, texture generation or reprojection, and export. Current RealityScan help pages are useful for modern terminology around alignment settings, unwrapping, and export controls, but here they serve only as concept support, not as proof that the 2018 RealityCapture interface exposed the same controls under the same names. [10] [5] [8] [7]

Accuracy, precision, resolution: what the 2018 review did not measure

The 2018 article is mainly a workflow-and-output review. It does not present an external metrology test, a checkpoint report, or a standards-style validation of dimensional accuracy. That distinction matters because reprojection error is an internal image-space diagnostic, not direct proof of real-world measurement accuracy. PhotoScan’s manual defines reprojection error as the distance, in pixels, between where a reconstructed 3D point projects onto an image and the original detected image point used to reconstruct it. Current RealityScan guidance likewise treats mean and median pixel errors as alignment diagnostics, with values ideally under 0.5 px, but still only as diagnostics. If scale or geometry must be trusted as measurements, independent controls such as GCPs, known distances, or test/check points are needed rather than a single low pixel-error number. [10] [6] [9]

Capture discipline still matters

The 2018 review makes more sense when you remember that photogrammetry quality is constrained by the photo set long before it is constrained by UI polish. Current RealityScan help says each point on the surface should be visible in at least two images, advises keeping viewpoint changes under about 30 degrees, and recommends overlap high enough that neighboring images share substantial coverage. RealityScan’s alignment settings page adds that, for best quality, neighboring images should ideally overlap by more than 60%. PhotoScan’s 1.4 manual states the same basic constraint in older language: minimize blind zones, use sharp images, and remember the software can reconstruct only geometry visible from at least two cameras. [4] [5] [10]

• Keep neighboring-image overlap comfortably above 60% when you can. [5]
• Make sure each surface point appears in at least two photos. [4] [10]
• Keep viewpoint jumps under roughly 30 degrees between shots where possible. [4]
• Use sharp photos and avoid blur from slow shutter speeds or shaky capture. [10]

Mesh reconstruction + cleanup: what the review emphasized

The 3D Mag review repeatedly distinguishes between a model that looks good once textured and a model that is actually clean as geometry. On statue-like and soft organic subjects, the author saw high-detail results arrive quickly, but also notes that MatCap or untextured inspection was necessary to judge the real mesh quality. The review further says that RealityCapture could generate very high-poly outputs, and its verdict recommends simplification and, in many cases, use of the built-in smoothing tool before treating the result as a solid downstream asset. [1]

The second recurring point was control. The review says hole filling still happened automatically, with no control over which holes were filled, and the verdict repeats that more control over hole filling remained on the author’s wish list. That is best treated as a review observation about the software behavior seen in that period, not as a universal statement about all later versions. The broader lesson still holds: a fast mesh is not the same thing as a trusted mesh, especially when the subject includes smooth patches, thin edges, or partially occluded areas. [1]

Workflow checkpoints before trusting a mesh

• Make sure all aligned cameras end up in one component if the scene is supposed to be one coherent reconstruction. [6]
• Inspect the model in an untextured or MatCap-style view, not only with texture enabled. [1]
• Judge texture quality separately from geometry quality, because one can flatter the other. [1]
• If scale or accuracy matters, use control points, GCPs, and test/check points rather than visual inspection alone. [9] [10]

Texture mapping quality (and why it can hide geometry issues)

One of the strongest historical takeaways from the review is that RealityCapture texture mapping quality could make a reconstruction look more complete than its raw surface deserved. The bust examples on 3D Mag make the point clearly: the textured result can look convincing, while a switch to MatCap exposes weaker geometric detail or noisier surfaces. The same pattern appears again in the camera and teddy examples, where texture and dense detail can be visually impressive, but the untextured mesh reveals surface noise, chaotic topology, or places where smoothing still helps. For readers interested in a RealityCapture mesh reconstruction workflow, that distinction is central: texture is a presentation layer over geometry, not proof that the underlying surface is already clean enough for measurement, retouching, or game-asset preparation. [1]

A later peer-reviewed ISPRS study gives one narrow but useful caution on color fidelity. In an orthophoto and color-checker test using Agisoft PhotoScan 1.4.3 (2018) and Reality Capture 1.0.3 (2018), the authors report that RealityCapture produced 8-bit output from 16-bit input in that specific workflow, with mean channel errors of R 0.105, G 0.085, and B 0.107 tones in the reported table. The same paper concludes that this should be understood within its orthophoto and color-processing setup, not as a blanket statement about every textured mesh export. Still, it is a reminder that texture fidelity has technical layers of its own, including bit depth and internal conversion behavior. [18]

RealityCapture vs Agisoft PhotoScan/Metashape in 2018

For historical accuracy, a 2018 comparison is mainly RealityCapture vs Agisoft PhotoScan, not RealityCapture vs Metashape. Agisoft’s own pre-release thread announces Metashape 1.5.0 (former PhotoScan) in October 2018, and the official changelog records Version 1.5.0 build 7492 on December 31, 2018. So when the 3D Mag review was published on March 15, 2018, the correct competitor label was still PhotoScan. If someone searches RealityCapture vs Agisoft Metashape today, the historically careful answer is that Metashape is the later name for the same Agisoft product line, but the 2018 review itself belongs to the PhotoScan era. [1] [12] [13]

That naming discipline prevents two different errors. First, it stops readers from back-projecting later branding into an earlier review. Second, it keeps the comparison narrow: 3D Mag was comparing outputs and workflow behavior on its own test scenes, not issuing a universal market ranking. In that test envelope, RealityCapture looked especially strong on statues and organic subjects, while flatter or more technical objects pushed the reviewer toward preferring PhotoScan’s geometry even when PhotoScan took longer. [1]

Export formats: what’s confirmed for the 2018 review vs what’s only in current docs

For the 2018 review layer, keep the claim narrow and concrete. 3D Mag says meshes could be exported as .FBX, .OBJ, or .PLY, and that point-cloud data could be exported as .XYZ through the mesh export path. Those are the formats confirmed in the review context, and they are enough to show that the software could move results into wider DCC, scan-processing, or inspection pipelines without locking the user into a closed format. [1]

Current RealityScan documentation is broader, but it belongs in a separate bucket. Today’s export help documents options such as single-texture maximal sides from 512 up to 65536 pixels and texture pixel formats including 24-bit BGR, 32-bit BGRA, and 64-bit RGBA. Those details are useful if you are checking modern export behavior, but they are not evidence that the same options existed in the 2018 review build. For this archival piece, the rule is simple: cite the review for 2018 export claims and cite the current docs only for present-day documentation. [7]

Limitations & failure modes (photogrammetry realities, not just RC flaws)

Many of the review’s weaker outcomes are better understood as photogrammetry constraints than as evidence that RealityCapture was inherently poor. PhotoScan’s 1.4 manual explicitly advises avoiding not textured, shiny, highly reflective, transparent, and absolutely flat objects or scenes. It also warns that the software can reconstruct only geometry visible from at least two cameras. Those warnings map cleanly onto what 3D Mag observed: statue-like and organic subjects behaved better, while flatter technical objects looked less convincing in the reviewer’s tests. [10] [1]

Thin and closely spaced structures are awkward for the same broader reason. PhotoScan’s manual uses the space between fingers as an example of where masking may be needed to suppress noise, which is a good reminder that occlusion, background contamination, and narrow gaps can all confuse reconstruction and texturing. So “fast” never meant “hands-off.” Even when alignment succeeds, the downstream job can still involve cleanup, smoothing, masking choices, hole judgment, and, if measurement matters, independent validation rather than visual confidence alone. [10] [1]

Current status note (dated, minimal, no buyer’s guide)

The company context changed after the review era. Epic Games announced the Capturing Reality acquisition on March 9, 2021, and the desktop product was rebranded on June 17, 2025, when RealityScan 2.0 was introduced as the product formerly known as RealityCapture. That means a reader in 2026 is looking back at a review written under an older product name, older licensing models, and older version boundaries. [15] [16]

Legacy licensing note: RealityScan published a notice on May 12, 2026 stating that the legacy RealityCapture licensing server would close on August 3, 2026. As of this article’s access date, August 7, 2026, that stated shutdown date is already in the past. Read that as a dated notice, not as a substitute for checking the current access and migration requirements directly in the live vendor documentation. [17]

Verdict: how to read this RealityCapture review 2018 in 2026

The most useful way to read this RealityCapture review 2018 is as a historically bounded test of workflow feel and output character. In that frame, the review still says something real: on the author’s machine and datasets, RealityCapture came across as fast, strong on textured organic subjects, and sometimes less convincing on flatter technical objects where the reviewer preferred PhotoScan’s geometry. It also shows why the software drew attention at the time: the combination of fast reconstruction, high-detail textured output, and a more capable 2018-era workflow update gave it a clear identity. But those observations remain tied to the review’s hardware, version, and June 2017 results caveat. [1]

What a RealityCapture review 2018 cannot do for a 2026 reader is settle the modern choice between current RealityScan and current Metashape. For that, you would need present-day documentation on alignment settings, control points, unwrap behavior, export controls, hardware support, and licensing, plus a recent side-by-side evaluation run on current versions. This archival article is still useful, but only if it stays in its lane. [5] [7] [8] [9] [16]

FAQ

What is the short verdict of this RealityCapture review 2018?
The short verdict of this RealityCapture review 2018 is that 3D Mag found RealityCapture impressively fast and especially strong on organic or statue-like subjects, but less convincing on flatter technical objects where PhotoScan’s geometry looked better to the reviewer. That verdict is still historically useful, but it must be read with the article’s own caveat attached: the result models shown in the Results section were not reprocessed for the 2018 update and were made with a June 2017 version. [1]

Was RealityCapture “better” than Agisoft PhotoScan in 2018?
Only in a limited, test-specific sense. The review judged RealityCapture very favorably on speed and on some organic subjects, but it did not claim universal superiority. On flatter technical objects, the same review leaned toward PhotoScan’s geometry. Also, the historically correct 2018 comparison label is PhotoScan, not Metashape: Agisoft’s Metashape 1.5.0 naming transition appears in an October 2018 pre-release thread, with a December 31, 2018 changelog entry for build 7492. [1] [12] [13]

Why do photogrammetry meshes look great textured but messy in MatCap or untextured view?
Because texture can visually mask geometric problems. The 3D Mag examples show that a reconstruction may look convincing once the surface is textured, while MatCap or untextured shading reveals noisy geometry, weak smooth areas, or topology that still benefits from smoothing and cleanup. Texture answers “does the surface look plausible from this view,” while raw mesh inspection answers “is the geometry itself stable, clean, and usable.” That is why textured beauty shots should never be the only evidence in a review. [1]

What does reprojection error mean — and what does it not prove?
Reprojection error is an internal image-space diagnostic. PhotoScan’s manual defines it as the pixel distance between the projected position of a reconstructed 3D point and the original detected image point used in reconstruction. Current RealityScan guidance says mean and median alignment errors are ideally under 0.5 px, and the alignment settings page recommends keeping maximum feature reprojection error around 3 px, but those are still alignment diagnostics. They do not by themselves prove external dimensional accuracy, survey accuracy, or scale correctness without independent controls or checkpoints. [10] [6] [5]

Why do shiny, reflective, transparent, or flat objects fail in photogrammetry?
Because the solver needs stable, repeatable image features across overlapping views. PhotoScan’s manual explicitly says to avoid not-textured, shiny, highly reflective, transparent, and absolutely flat objects or scenes. Those surfaces either lack reliable visual detail or change appearance too much from one angle to the next, which weakens matching, camera estimation, and dense reconstruction. Thin gaps and occluded spaces can create similar trouble, which is why masking and careful capture planning often matter as much as the software choice. [10]

What export formats were confirmed in the 2018 3D Mag test, and what’s only in current RealityScan docs?
For the 2018 test layer, the confirmed exports were .FBX, .OBJ, and .PLY for meshes, plus .XYZ for point-cloud output through the mesh export path. Those are review-confirmed claims. Current RealityScan docs describe broader modern export controls, including single-texture limits from 512 to 65536 pixels and additional texture pixel-format options, but those current-doc details should not be back-ported into the 2018 review unless independently verified for that older build. [1] [7]

Is RealityCapture still called RealityCapture today (as of August 2026)?
Not as the main current desktop brand. On June 17, 2025, RealityScan announced that the product formerly known as RealityCapture was now RealityScan for desktop. Separately, RealityScan’s May 12, 2026 notice stated that the legacy RealityCapture licensing server would close on August 3, 2026. Since this article’s access date is August 7, 2026, that shutdown date is already past, so readers should verify the current migration and access rules in the live documentation rather than relying on an old licensing assumption. [16] [17]

Sources

1. 3D Mag — “Updated for 2018: RealityCapture Review”
2. Capturing Reality — Release Notes PDF (includes historical Version 1.0.3.3939 RC entry)
3. CG Channel — “Capturing Reality releases RealityCapture”
4. RealityScan Help — How to Take Photographs
5. RealityScan Help — Alignment Settings
6. RealityScan Help — Alignment Quality
7. RealityScan Help — Model Export
8. RealityScan Help — Unwrap Tool
9. RealityScan Help — Control Points
10. Agisoft — PhotoScan Professional Edition User Manual v1.4 (PDF)
11. Agisoft — PhotoScan Standard Edition User Manual v1.4 (PDF)
12. Agisoft Forum — Metashape 1.5.0 pre-release thread (former PhotoScan; dates)
13. Agisoft — Metashape changelog (PDF)
14. NIST Glossary — Photogrammetry definition (ASTM-derived)
15. Epic Games — Capturing Reality acquisition announcement
16. RealityScan — RealityScan 2.0 rebrand announcement (RealityCapture → RealityScan)
17. RealityScan — Legacy licensing server shutdown notice
18. ISPRS Archives (2019) — bit-depth / orthophoto color fidelity study including RealityCapture 1.0.3
19. ISPRS Archives (2017) — critical review of automated photogrammetric processing