Spatial 3D capability map · Published 2026-07-28
What do spatial 3D models actually measure?
These systems recover geometry from images, and that recovery is measured. Holding a world, reasoning about space, and predicting how a scene changes are different questions, and they have different evidence behind them.
Short answer
Spatial models measurably recover geometry, camera relationships and task-specific representations from images. That is real progress and it is reported per output — point maps, depth, camera pose, correspondence, reconstruction — on named benchmarks. It is not the same as maintaining a persistent world, reasoning causally about space, or predicting how a physical scene will change. Those are separate capabilities, tested by separate protocols, and the sources reviewed here do not establish them.
The capability map
Six things that get collapsed into one word
The phrase "spatial intelligence" is doing a lot of work in public materials. Underneath it sit at least six different capabilities, each with its own measurement: reconstructing a scene's geometry; estimating camera pose and relationships; learning a spatial representation for downstream tasks; answering spatial-reasoning questions; generating a 3D world you can move through; and understanding physical causes well enough to predict an intervention.
A system that does the first well tells you nothing about the last. This article keeps them apart, and it does not treat a model's use of the word "world" as evidence for all six. Where a source measures one of them, we say which one, on what protocol.
Three things make this harder than reading a results table, and each shows up below at the point where it matters. A widely cited cross-paradigm benchmark was written by a group that also entered a model into it. The one comparison here run by a party with no stake in the models, in aerial photogrammetry, finds the accuracy ranking flips depending on how many images you feed it. And the strongest capability wording for a commercial 3D world generator appears on the company's About page rather than in the product announcement itself.
1. Reconstruction and geometry: measured, per output
DUSt3R, MASt3R and VGGT are the reference points for learned 3D geometry from images without the standard calibrated structure-from-motion entry point. Each reports measured results on named benchmarks, and the useful discipline is to keep the outputs separate rather than to average them into a capability word.
DUSt3R reports point-map reconstruction and camera-related geometry, with a global alignment stage and benchmark tables covering visual localization, multi-view pose, monocular depth and multi-view depth. MASt3R reports dense correspondence and localization results on the Map-free dataset's validation and test sets, multi-view pose regression on CO3Dv2 and RealEstate10K, and visual localization on Aachen Day-Night and InLoc. VGGT reports camera pose on RealEstate10K, dense multi-view stereo on DTU, point-map estimation on ETH3D and two-view matching on ScanNet-1500.
We deliberately publish no numbers from these three papers, and we do not rank them against each other. Ordering systems on results we have not checked line by line would be an assertion dressed as a summary. Where the systems are compared, below, the comparison belongs to whoever ran the protocol.
VGGT's own limitations section is worth more than a ranking anyway. The paper states that the model does not support fisheye or panoramic images, that reconstruction performance drops under extreme input rotations, and that while it handles minor non-rigid motion it fails under substantial non-rigid deformation. It then asserts that addressing these "can be straightforwardly achieved by fine-tuning the model on targeted datasets with minimal architectural modifications." No experiment in the paper tests that remedy; it is a developer assertion, not a result.
2. A cross-paradigm benchmark, and who wrote it
SpatialBench asks directly whether spatial foundation models generalize, and it is the broadest cross-paradigm evaluation reviewed here. Its protocol covers 19 datasets, 546 scenes, 5 spatial domains, 41 models across 6 paradigms, 5 task suites and 72,540 evaluation samples, under four deterministic input-density settings: Single Frame, Sparse, Medium and Dense. Its answer is negative: current models are not all-round players.
The specific findings are more useful than the headline. Across those four input-density settings the benchmark reports that performance shifts with paradigm, memory regime and scene domain. Its domain-level analysis reports major out-of-distribution gaps in egocentric and wrist-view domains, meaning head-mounted and hand-mounted camera views. And the Dense setting is a hard wall rather than a score: DUSt3R, MASt3R, VGGT, MapAnything, OmniVGGT, several other entries and the authors' own model all run out of memory above 140 GB, and one system times out beyond four hours per scene. Their reported averages are parenthesized because they cover only the settings that finished. Any sentence citing a SpatialBench result has to name which of the four settings it came from, or it is not saying anything checkable.
There is a source-role fact that the paper is open about and that readers of secondary coverage will usually miss. The same paper introduces the authors' own model, Depth-Anything-Next, and their own training set, and enters that model into the benchmark it defines. Its results there are mixed rather than a sweep: large gains on depth error in the Single-Frame, Sparse and Medium settings, gains on pose accuracy in Sparse and Medium, but worse trajectory error and worse F-score than the best baseline in the Medium setting, and the same out-of-memory failure in Dense.
This bounds one thing and not another. The negative finding stands, and it is if anything strengthened by the authors' own entry failing to be all-round either. What it does not support is treating SpatialBench's cross-model ordering as an independent third-party evaluation of DUSt3R, MASt3R or VGGT. It is a comparison published by a participant. That is a normal and disclosed arrangement in machine-learning benchmarking; it is still not third-party verification. Our general framework for reading evaluation claims is in How do you evaluate a world model?, and this is a concrete instance of it rather than a restatement.
3. Where the ranking flips: an independent comparison
The most informative source here was not written by anyone with a stake in these models. A photogrammetry group evaluated DUSt3R, MASt3R and VGGT against COLMAP, the standard classical pipeline, on the UseGeo aerial dataset, with input counts from 1 to 191 images and overlap from roughly 10% to 70%.
The finding is that the accuracy ranking reverses with configuration, and it is worth being precise about which quantity reverses, because three different kinds of percentage appear in this study and they are easy to confuse. Accuracy is a mean vertical distance in metres. Completeness is the share of ground-truth points recovered within a 1 m threshold, so an absolute percentage. And several figures are relative improvements, which are neither.
| Configuration | What the study reports |
|---|---|
| 1 image | The learned methods reconstruct dense urban point clouds; COLMAP fails outright, because viewing angles are insufficient for triangulation. The learned reconstructions still show holes around buildings and failures on small towers. |
| 2 images, large viewpoint difference | COLMAP often fails or yields sparse low-quality models, accuracy up to 2.3 m. MASt3R and VGGT reach accuracy up to 0.4 m, with a completeness gain of +10%. Across these comparisons they beat COLMAP on completeness in 87% of instances, by up to an additional 19% completeness. |
| 38 images, standard overlap | COLMAP at high resolution reaches 0.2 m accuracy — a 92% relative error reduction against roughly 2.0 m for the others. At low resolution COLMAP fluctuates and can reach 4 m where MASt3R reaches 0.4 m. |
| 38 images, low overlap | COLMAP is more accurate in 93% of cases, by up to 80% (relative accuracy improvement). MASt3R and VGGT are more complete in 80% of cases, with completeness gains of up to +50%. As overlap falls to 10%, COLMAP's completeness drops to about 8% (absolute), which the authors call insufficient for practical use. |
| High resolution, typical 70% overlap | The authors report that the learned methods' "advantage diminishes" here: similar or moderately higher completeness, an advantage of up to 20% (completeness gain), while COLMAP is substantially more accurate, reducing errors by up to 9% (relative accuracy improvement). |
So the reversal is real and it is specific to accuracy. It does not extend to completeness. At the typical 70% overlap that ordinary photogrammetric survey work uses, the learned methods may still hold a completeness advantage of up to 20%, and the authors describe COLMAP's completeness edge over VGGT as "not substantial". A version of this article that said the learned models simply lose on the classical pipeline's home ground would be sharper than the study's own analysis section.
One number in this study circulates without its context and should not: the headline comparison of up to 74% completeness (absolute) for COLMAP against 36% (absolute) is a comparison with DUSt3R, while the same paper's analysis section says the COLMAP-versus-VGGT completeness gap is not substantial. Two further cautions. Every one of the learned methods reports a 100% image-reconstruction success rate at every overlap level from 70% down to 10%, while COLMAP at low resolution falls from 75% to 13% — but success here means output was produced, not that the output was accurate, and it must never be read next to the accuracy rows as though it were a quality result. And this is one aerial protocol on one dataset. It is not a universal verdict on spatial models, and it does not show the learned methods are weak or that COLMAP is obsolete. The authors' own conclusion is that VGGT cannot serve as a replacement for traditional structure-from-motion and multi-view-stereo pipelines in typical photogrammetric applications, while being valuable in sparse or difficult cases.
4. Spatial reasoning is a different test
Answering questions about space is measured separately from reconstructing it. Spa3R learns a view-invariant spatial representation from unposed multi-view images and reports results on the VSI-Bench spatial-reasoning benchmark and on a set of other benchmarks.
Those are results about those named tasks. This is one system, in a preprint, evaluated by its own authors, and a single system cannot define a whole capability category. It is evidence that a 3D visual-reasoning evaluation layer exists and can be scored, not evidence about general or human-like spatial reasoning, and not evidence that the system holds a scene or acts in one.
5. Generating a world is not reconstructing one
The commercial edge of this field is generation rather than measurement. World Labs describes Marble, announced 12 November 2025, as "a first-in-class generative multimodal world model". The workflow it describes is create, edit, expand, combine, export: create 3D worlds from text, images, video or coarse 3D layouts, interactively edit, expand and combine them, then export as Gaussian splats, meshes or videos.
That is a creation workflow claim, and it is well documented as one. What it is not is a measurement. The announcement reports no reconstruction accuracy, and its only reference to physics is a collider mesh that the page itself calls a "low-fidelity" mesh "intended for coarse physics".
There is a wording detail worth stating precisely, because it is the kind of thing that turns into a capability claim after two rounds of paraphrase. The word "persistent" does not appear in the Marble announcement. It appears on the World Labs About page, which says the product "enables anyone to create spatially cohesive, high-fidelity, and persistent 3D worlds from just a single image, video or text prompt." So the company does describe the worlds as persistent — on its company page, not in the technical announcement. Attribute it to the page it appears on. Neither page publishes a test of persistence.
We record this asymmetry because we walked into it ourselves: until this month this site's own source table described the announcement as supporting "persistent, editable and exportable 3D world generation", attributing to World Labs a phrase that page does not use. That is corrected, and it is a fair illustration of how the strongest wording migrates toward the least technical document.
What these sources establish, and what they do not
| Capability | Status against the sources reviewed here |
|---|---|
| Metric geometry from images | Established, per output and per benchmark, in developer-authored evaluations. |
| Cross-domain generalization | SpatialBench reports major out-of-distribution gaps in egocentric and wrist-view domains, on a benchmark its own authors built. |
| Large-scale reliability | Contested and configuration-dependent: the accuracy ranking reverses with image count and overlap in aerial photogrammetry. |
| Named 3D visual reasoning | Established for the named benchmarks only, by a single system's authors. |
| Persistent world state | Not established by these sources. The word appears on a company About page, not in the product announcement, and neither publishes a test. |
| Physical or causal understanding | Not established by these sources. None of them tests intervention or causation. |
| Independent third-party validation | Of the seven sources in this review, only the aerial photogrammetry study is independent of the systems it evaluates, and it covers aerial photogrammetry only. |
Every line above is a statement about these sources, not about the world. "Not established by these sources" is not the same as "does not exist", and we did not run a documented search wide enough to say how much independent evaluation of these systems exists elsewhere.
The defensible summary is narrower than the marketing and more useful than a dismissal. Geometry recovery is measured and real. Everything past geometry is a separate question with separate evidence, and for the strongest-sounding capabilities, none of the sources reviewed here provides it.
Evidence map
| Evidence | Sources | Boundary |
|---|---|---|
| Learned systems report measured 3D geometry outputs on named benchmarks. | DUSt3R; MASt3R; VGGT | Developer evaluations. Outputs reported separately; no ranking among them is published here, and no cell-level values were verified. |
| A cross-paradigm benchmark reports that current models are not all-round players. | SpatialBench v2, Table 1 and the OOD analysis | Protocol-bound; every citation names one of four input-density settings. The authors also enter their own model, so the cross-model ordering is not third-party evaluation. |
| An independent aerial comparison finds the accuracy ranking reverses with image count and overlap. | UseGeo photogrammetry evaluation, §4.1 | Accuracy-specific. No completeness reversal at typical 70% overlap; one aerial protocol, not a universal verdict. |
| 3D visual reasoning is scored on its own benchmarks. | Spa3R, VSI-Bench results | Single system, preprint, author-evaluated; cannot define the capability category. |
| Marble is publicly positioned as generative 3D world creation with an edit, expand, combine and export workflow. | World Labs Marble announcement; World Labs About page | Developer positioning. "Persistent" appears only on the About page; neither page publishes a persistence test, reconstruction accuracy or physical simulation evidence. |
| Persistence and physical or causal understanding are not established by these sources. | All seven reviewed sources | A statement about this source set, not proof that such evidence does not exist elsewhere. |
Sources
- DUSt3R: Geometric 3D Vision Made Easy (arXiv:2312.14132v3) — pose-free point-map reconstruction and geometry estimation.
- Grounding Image Matching in 3D with MASt3R (arXiv:2406.09756v1) — dense correspondence, localization and reconstruction.
- VGGT: Visual Geometry Grounded Transformer (arXiv:2503.11651v1) — feed-forward geometry prediction; §5 states the fisheye, panoramic, extreme-rotation and non-rigid limits.
- SpatialBench: Is Your Spatial Foundation Model an All-Round Player? (arXiv:2605.27367v2) — cross-paradigm benchmark; the authors also enter their own model.
- An Evaluation of DUSt3R/MASt3R/VGGT 3D Reconstruction on Photogrammetric Aerial Blocks (arXiv:2507.14798v2; DOI 10.1080/10095020.2025.2597491) — the one independent comparative evaluation reviewed here.
- Spa3R: Predictive Spatial Field Modeling for 3D Visual Reasoning (arXiv:2602.21186v1) — 3D visual-reasoning benchmark results.
- World Labs: Marble: A Multimodal World Model, 12 November 2025 (archived 2026-07-27) and the World Labs About page — developer positioning, retrieved and archived 2026-07-27.
Technical claim mapping
Structured claim and source IDs are available in claims.json, sources.json and articles.json. Fixed paper versions, exact page and table locators, the source-role audit and the non-conclusions for this article are recorded in the research package for spatial-3d-capability-boundaries.
Verification scope for DUSt3R, MASt3R and VGGT: this article checked which results each paper reports and where those tables sit, not the individual table cells. No value from those three papers is published here, and no ordering among them is asserted. Comparisons involving them come from SpatialBench and from the aerial photogrammetry evaluation, each named with its own protocol.