Marble archival-photo pilot — Bowling Green, 1941

Conclusion

None of four Marble 1.1 generations met TimeWalk’s historical-accuracy requirement. Cleaner inputs improved visual coherence, but did not preserve building geometry, park layout, skyline, or period details reliably. Stop this pilot; retain the sources and outputs as a benchmark for a future Atlas evaluation. Atlas was not tested, and no claim is made that it will solve these problems.

Goal and method

Create an explorable reconstruction of Bowling Green and lowest Broadway circa 1941 from freely accessible archival photographs. Work used the World Labs Marble API, image and multi-image reconstruction, model marble-1.1, and seed 194112. Outputs included panoramas, Gaussian splats (SPZ), and collision meshes. No Unreal import or surveyed geometric registration was performed.

The tests were iterative troubleshooting, not a controlled benchmark: source count, preparation, prompts and coverage changed between runs. The request was for 1941; the expanded set also included explicitly dated 1936 and 1938 references.

Sources

• Arthur Rothstein, December 1941: park/Broadway overview, LOC 2017774895 and overlapping street view, LOC 2017774901. These are northward views, not opposite directions.
• Rothstein, December 1941: Battery Park oblique, LOC 2017774897, looking northeast toward the western-side buildings. This is not a west-facing photograph from inside Bowling Green.
• Suydam, September 1, 1938: elevated southward view of the park and Custom House, NYC Municipal Archives WPA image wpa_3306, reproduced in DORIS’s Custom House article.
• Berenice Abbott, July 23, 1936: Custom House sculpture with Produce Exchange behind, NYPL image 482769. Date/title from an attributed mirror; direct catalogue access was blocked. A conflicting mirror link to another photograph was rejected after inspection.

LOC FSA/OWI photographs were treated as public-domain references. Free web access to the other items is not a blanket reuse-license determination; their item-specific rights remain to be checked before redistribution. Originals and source manifests are retained locally.

Four tests

Test	Inputs and change	Observed result	Credits
V1	Two overlapping December 1941 north-facing photographs	Repeated curved facade, altered plaza, distorted statue, modern-looking tower and vehicles	1,600
V2	One overview photograph, monochrome, tighter composition prompt	More recognizable source-facing view; invented surrounding geometry, altered rooflines and later-looking cars	1,580
V3	Five photographs spanning 1936–1941; broader but mismatched viewpoints; equal square canvases with pale padding	Severe white gaps and framed-image-like planes, duplicated/distorted buildings and anachronisms	1,600
V4	Two December 1941 overlapping photos; film borders removed, border-free 768×1024 crops; scope narrowed to north end	White-gap failure absent in inspected primary view; still incorrect facade proportions, compressed park, altered skyline and statue/landscaping	1,600

Total: 6,380 credits = $5.104(about$5.10) at 1,250 credits per dollar, within the $10 pilot ceiling. No paid high-quality mesh exports were purchased. Generation charges consumed prepaid credits; this is their dollar equivalent.

What went wrong—and what remains uncertain

1. Coverage was not a calibrated capture. The five-photo set mixed street-level, elevated, distant and detail viewpoints. It lacked a true west-facing park view. More photographs did not automatically provide usable geometric constraints.
2. Input preparation introduced a preventable problem. Pale letterboxing in V3 likely contributed to the white-gap/frame failure. V4 removed it, but also changed photo selection and prompt scope, so padding’s causal contribution was not isolated. This preparation was TimeWalker’s responsibility, not evidence by itself of a model limitation.
3. Generated plausibility was not historical fidelity. Even the cleaner run altered directly photographed buildings and the park. Black-and-white appearance did not make invented details accurate.
4. Higher splat count did not repair wrong geometry. Both lightweight and full-resolution viewing were supplied. Ted inspected the full-resolution latest result and rejected its accuracy.
5. This does not establish that every Marble multi-image workflow fails. We did not test a well-covered, calibrated same-era survey, a map-constrained blockout, or Atlas.

World Labs’ multi-image guidance recommends matching dimensions/aspect ratios, nearby viewpoints, shared space and overlap. It also explains that unseen areas are generated plausibly, not recovered as a verified floor plan.

Delivery and verification

Private Marble web links produced a permission-denied response. The supplied API key was used; a different agent-owned account was not intentionally substituted. Exact web/API identity matching remained unresolved. No world was made public.

Actual SPZ assets were successfully rendered in the official Spark viewer through direct CDN links. Such links expose the referenced asset to anyone possessing the link; they do not fix Marble account ownership. V4 full-resolution output contains 1.92 million splats. Gzip/SPZ headers and collider GLB length/magic were checked. Collision geometry is not a textured production mesh. No Unreal compatibility or metric-accuracy claim is made.

PC navigation was a separate issue: Chrome’s active GPU engine mapped to Microsoft Basic Render Driver rather than the RTX 5090; two TimeWalk processes also loaded the NVIDIA GPU. The 100K-splat viewer was faster for Ted. Rendering performance does not explain inaccurate scene content.

Recommendation

• Stop further Marble rerolls for this pilot. Keep the photographs as references for map- and footprint-constrained modeling.
• Revisit when Atlas access arrives, without assuming it guarantees fidelity.
• Use a fixed benchmark: verified building footprints and street widths, matching camera views, held-out photographs, and explicit checks for facade proportions, skyline, park boundary, statue location and period vehicles.
• Separate source-supported geometry from invented surroundings. Require historical review before production use or paid mesh export.

Evidence and related research

Experiment receipts, prompts, source manifests, costs, QA and outputs: TimeWalker project artifact directory artifacts/bowling-green-1941/, including v3/ and v4/. Earlier two-run QA notes are superseded by this four-run summary.

• World Labs API pricing
• Bowling Green
• Modern NYC building footprints — useful methodological reference, not evidence of 1941 footprints.