Parallax Metrology
Track 1 image set
All images: MidJourney

Track 1 · NCD × VTL Atlas

Where literal substrate disagrees with library-free structure-space — and which dimensions hold any two apart.
Every image is a displacement vector from a fixed geometric origin, across the full VCLIG battery, scaled by each axis's geometric range — no learned population. NCD asks whether two images share compressible substrate. VTL asks whether they organize space similarly. The useful cases are often the disagreements.
showing inlined sample · 7 images
Executive finding
Track 1 is a structurally coherent portrait family: the images are not literal duplicates, but they repeatedly instantiate the same portrait-composition grammar.

NCD mostly rejects byte-level sameness across the set, while VTL finds many near relationships in spatial structure. The finding is not “these are the same images.” It is stronger and more useful: many different MidJourney substrates converge on the same compositional machine.

54
Structural kinship pairs
High relative substrate distance with low relative VTL distance: different pixels, same structure.
325
Pairwise comparisons
The atlas compares every image against every other image across substrate and structural layers.
L1
Dominant driver layer
Most separations are driven by spatial placement, dispersion, centroid, packing, and peripheral pull.
NCD ≈ 1.0
Substrate behavior
The files mostly do not compress together, consistent with separate generated PNGs rather than duplicates.

Collection summaryhow the full set divides across the NCD / VTL crossing

Distance matrixdarker green = structurally closer · darker red = farther · click a cell for drivers

structural kinship case — far by substrate (NCD), near by structure (VTL)

Disagreement mapx = substrate farness rank · y = VTL farness rank · lower right is structural kinship

Top structural kinship

Highest disagreement

Closest VTL pairs

Farthest VTL pairs

Displacement from originmagnitude of each image's position vector — how far from the structural zero

Nearest neighborwho each image is structurally closest to — the cluster structure

Domain findingsinterpretation of the Track 1 portrait field

Reading the output
The set behaves less like a pile of separate prompts and more like a narrow portrait structure manifold.

The atlas separates literal substrate from structural organization. At the substrate layer, most images remain far apart: the PNGs do not share much compressible byte structure. At the VTL layer, many pairs move close together because the images reuse a common arrangement of figure, void, shadow, center of mass, and vertical portrait staging.

What the structural-kinship cases mean

  • Different pixels, same portrait grammar. The 54 structural-kinship pairs are the key signal: NCD says the files are materially different, while VTL says their spatial skeletons are close.
  • Prompt families are not the only explanation. Some strong kinship pairs cross subject labels, such as courtier, burgher, clergyman, militia officer, and guild master. The shared structure lives in the image architecture, not just the words.
  • The images converge on a central bust template. The repeated grammar is dark-field portraiture with a central mass, controlled peripheral pull, similar dispersion, and comparable vertical staging.

What drives separation

  • L1 spatial structure dominates. The strongest aggregate layer is spatial position and distribution: dispersion, vertical centroid, peripheral pull, cohesion, and packing.
  • Topology and tone refine the split. Void count, mass islands, shadow mass, and gestural thrust determine whether two otherwise similar portraits separate structurally.
  • Altered-structure pairs mark composition changes. The 54 altered-structure pairs often preserve a similar substrate family but shift crop, figure scale, void topology, shadow field, or head/body placement.

What NCD contributes

  • NCD acts as a negative control. It confirms these are not simply byte-near duplicates or trivial copies.
  • The useful result is disagreement. When NCD is far and VTL is near, the instrument identifies structural reuse beneath pixel-level difference.
  • Here, NCD’s range is tight. Most substrate distances sit near 1.0, so the NCD high/low split should be read relatively within Track 1, not as an absolute universal threshold.

Bottom line

  • Track 1 is visually diverse at the surface, but structurally constrained underneath.
  • The generator varies costume, label, era, and rendering details while preserving a stable portrait-composition scaffold.
  • The atlas makes that scaffold measurable. It turns a visual hunch into pairwise evidence: which images share structure, which only share substrate, and which axes explain the difference.