A.rtist I.nfluencer  ·  Visual Thinking Lens  ·  Parallax Portrait Basin

A Deterministic Measurement of Image-System Independence
in Authored and Generative Portraiture

The Geometry of Authored Portraiture: Image-System Independence

1,341 Images  ·  4 Classes  ·  8 Traditions  ·  A 3-Axis Treatment Basin

Russell Parrish  ·  Parallax  ·  Portrait Basin  ·  VTL Structural + Color Kernel  ·  June 2026

Sample of the 1,341-image corpus: pre-modern oil portraits, Midjourney default, and Midjourney steered

Executive Summary

This report measures what separates an authored portrait from a generative one, using the deterministic VTL structural + color kernel. The corpus is a matched design: 446 pre-modern European and American oil portraits (CC0, museum open-access) against 447 Midjourney portraits from naive prompts, plus 448 composition-steered and 32 color-steered Midjourney portraits as two same-render controls. Every measurement is deterministic and reproducible from the source image; every headline statistic below carries a bootstrapped 95% confidence interval (Appendix B) and a permutation test.

1. Overview & Dataset Structure

The VTL kernel is a deterministic image-measurement device. It converts a rendered 2-D image into a fixed vector of structural and chromatic coordinates — gradient mass, void ratio, cohesion, dispersion, OKLab color, and color–structure coupling — using classical computer vision (Sobel gradients, connected-component labelling, percentile thresholds). It ranks nothing; it locates an image in coordinate space, the way CIE L*a*b* locates a color or GPS locates a point. Its diagnostic power comes from distributions across a dataset: when images that should differ cluster tightly, that reveals a default; when they spread, that reveals range.

This study points the kernel at one coherent basin and its generative shadow: pre-modern (≤1900) European and American oil portraiture vs the Midjourney portrait default. The corpus is fully matched across four classes.

Six research questions structure the report. (1) Can the two classes be told apart by where the visual mass sits? (2) If not, which treatment axes separate them? (3) How is the difference physically constructed, image by image? (4) How much of the coordinate space does each class occupy? (5) Does naming a historical tradition move the engine, or does it default regardless? (6) Which of these behaviors is reachable by prompt, and which is locked in the weights?

Two senses of "distance." The engine reports a deviation-Ω, inverse-variance-weighted distance from the authored reference distribution, used here as the separation metric. Selection of the axes that define the basin runs through an explicit gauntlet: F-statistic > 50 (the axis discriminates), bootstrap coefficient-of-variation < 0.20 (it is stable), and pairwise |r| < 0.70 (it is non-redundant). Only axes that survive every confound control (§12) enter the canonical basin.

2. Data Integrity & Measurement

The authored corpus was fetched under a scope filter enforced at the source (oil + painting + European/American + end-date ≤1900) with full provenance — source, accession ID, title, artist, date, license, dimensions — retained for every image. A single manual cull step removed 30 multi-figure or non-portrait works and one low-resolution scan; the criterion was membership in the single-sitter portrait tradition, never taste. Range is the asset, not a liability, so no image was dropped for being unusual.

The Midjourney classes were generated from a prompt deck of 112 sitters spanning eight registers — Dutch Golden Age, Spanish Golden Age, 18th-century English, Flemish Baroque, Italian Renaissance, early American Colonial, 18th-century French Rococo, and Northern Renaissance — with the historical tradition embedded in each filename, which is what makes the per-tradition analysis (§8) possible. All four variants of each generation were kept; no curation toward a house look.

A confound stated up front. Oil paintings and photographic-style AI renders differ in medium as well as authorship. This study does not pretend otherwise. Two devices address it: the two-regime audit (every axis is also measured luminance-normalized, so lighting-only effects are caught) and the same-render control (§9, Midjourney-vs-Midjourney), which isolates composition from render character. The honest frame is structural persistence — what survives transformation and same-render — not structural truth.

3. Finding 1: Position Is Weak — Both Center

Every position axis is weak (|Cohen d| ≤ 0.46), and where it tips, artists center more than Midjourney does. The pre-registered hypothesis — that AI portraits would read as centered, symmetric, and bokeh-blurred against looser human composition — failed. The pipeline reported the failure rather than confirming it. The condition actually shouldn't be a surprise. While an artist might semantically place a subject to one side, through mass and distribution, an artist typically uses offsets to "balance" the composition, orienting the gravity to the center.

This is a credibility result, not an embarrassment. Formal pre-modern portraiture is mass-center by convention. Artists were largely not studying composition — "left-leaning state portrait," or any other state — and the engine found exactly that: frame-gravity is higher in oil (DGI 40.45 vs 34.89), the opposite of the prior. The hand-designed position priors were wrong in direction, the gauntlet correctly demoted them, and the search moved to where the signal actually lives: in treatment, not placement. A measurement system that only ever confirms its author's hypothesis is not measuring; this one disconfirmed its own.

4. Finding 2: The Tell Is Treatment

With position excluded, the discriminating axes resolve into two co-equal pillars: mass (how coherently visual weight is built) and value-independent color (whether color makes structural decisions independent of light). Every treatment axis below clears its permutation test at p < 0.001 (5,000 label shuffles), and the confidence intervals exclude zero by a wide margin. A note on dimensionality: the color pillar is one axis. The kernel offers two color-autonomy measures, CBI and color_independence, but they correlate r = 0.94 (color_independence is defined as half CBI), so we carry only CBI in the core, in keeping with the gauntlet's own non-redundancy rule (|r| < 0.70). The canonical basin is therefore three axes: r_v, mu, CBI.

The separation is real, not an Ω artifact. Independent classifiers trained on the same axes — LogisticRegression and Linear Discriminant Analysis — both reach 91.2% five-fold held-out, so the structure is in the data, not in the distance function. Gauntlet selection is stable across random seeds (Jaccard 1.00). The 3-axis core is the defensible scientific object; the 91% full-basin classifier is the ceiling. We report both and headline the conservative one. (Dropping the redundant color twin costs ~2 points of hold-out, inside the noise, and buys a non-redundant core.)

5. Finding 3: Mass vs Gradient — How the Quiet Is Built

The dominant axis is r_v (Cohen d = 1.25, 95% CI [1.08, 1.43]): the artist builds with coherent mass — a quiet gradient field broken by deliberate edges — while Midjourney builds with distributed gradient, a busy field of small activations. Oil mean r_v = 0.78 vs MJ 0.62. It survives luminance-normalization, so this is structure, not lighting. To anchor the effect: a random oil portrait carries more mass than a random Midjourney portrait 82% of the time (common-language effect size).

The named extremes make the axis concrete. The corpus maximum is Whistler's portrait of Arthur Jerome Eddy — a tonalist who titled his own painting as a structural arrangement ("Arrangement in Flesh Color and Brown") — and whose r_v of 0.99 is the engine agreeing with him. The minimum is a Midjourney monk whose every square inch carries fine rendered detail, so almost no region reads as quiet. Goya's "Portrait of General José Manuel Romero" holds the highest cohesion in the corpus (mu = 0.89): a single dominant mass against a resolved ground.

r_v measures amount, not arrangement, and the quiet is built two ways. Two images can share an r_v by opposite means, so the brightness of the quiet pixels is diagnostic. Oil's quiet is predominantly dark shadow-mass (chiaroscuro): dark-quiet 63.6% vs MJ 51.1%. Midjourney leans roughly 1.5× more on bright sheen (5.3% vs 3.5%), glossy highlights deployed as an artificial balancing device where the artist would use shadow. Same r_v, reached by shadow versus gloss; the steering experiment (§9) later moves MJ off sheen and toward shadow, confirming the mechanism.

Sensitivity. Removing the most extreme 5% of each class on r_v does not weaken the finding — it strengthens it, from d = 1.25 to d = 1.68. The separation is carried by the bulk of the distribution, not by outliers (full table in Appendix B).

6. Finding 4: Color Autonomy — the Mourgue Result

The second pillar is value-independent color — whether color draws boundaries the light does not — measured by CBI (chromatic boundary independence): the fraction of strong chroma edges that fall where luminance is flat. Authored color does this structural work (CBI d = 0.85; the redundant twin color_independence agrees at d = 0.95); Midjourney's color rides luminance. And the instrument, pointed at paintings, recovers art history unprompted. (CBI captures one color grammar, color decoupled from value; it does not capture relational, all-over color where hue and value move together — a separate study.)

The metric rediscovers what painters already knew. A color-autonomous image in 1,496 Mourgue has a large, smooth, near-white field of a dress — a huge reservoir of weak-luminance pixels — isolated saturated accents on a muted ground. The green bow, the gold bracelet, the orange-red drape, the boy's blue jacket: each a high-chroma island against low-chroma surroundings, so each border clears the 85th-percentile chroma-edge bar. Cool-shadow, warm-light rather than pure value. Hue turning where the light barely does is precisely color-without-luminance-edge, registered across the flesh and fabric. The engine did not know it was a Mourgue; it measured CBI = 0.271 and put it at the top. Within authored oil overall, the Flemish tradition leads color autonomy (color_independence 0.122), while the tonal Dutch and American traditions lead mass (r_v 0.82 and 0.82). A measurement system that reproduces the Venetian/Flemish colorito-versus-tonal divide from pixels alone is measuring something real about authored treatment. A random oil portrait shows more color autonomy than a random Midjourney portrait 78% of the time.

Classical technique gives the mechanism a name: form modeled by warm–cool temperature at equal value, colored shadows instead of merely dark ones, hue boundaries placed where no luminance boundary falls. That is precisely what CBI counts — strong chroma edges with weak coincident luminance edges — and precisely what Midjourney's near-zero scores show it does not do.

7. Finding 5: Operating Volume & Monoculture

The deepest difference is not where Midjourney centers — it is how much structural possibility it can produce at all. On the directional treatment axes — cross-scale mass wander, orthogonal partitioning, vertical and diagonal energy — authored portraiture spans 6.76× the operating volume of the Midjourney default (generalized volume, the √det of standardized covariance: oil 0.379 vs MJ 0.056). These are the construction choices nobody is prompting for, and it is exactly there that the engine collapses to a narrow monoculture. The effect is direction-specific, and being precise about it is what makes it strong: on the core discriminating axes (mass + value-independent color) the gap is only 1.10× — expected, since the core is where the classes are defined and where steering can reach, so MJ keeps some range. The monoculture is not in the signal we classify on; it is in the unprompted treatment.

On the most directional axes — cross-scale mass wander, orthogonal partitioning, vertical and diagonal energy — the artist distribution is 2.46×, 2.22×, 1.89×, and 1.70× wider than Midjourney's. This reframes the question as compressed versus expansive construction strategy. It may eventually outrank the classification result: the same machine could ask whether some human traditions are already "AI-like" (compressed) while others explode the space — the library-defines-the-basin thesis, measured directly.

8. Finding 6: Does the Tradition Move the Engine?

Naming a historical tradition (Dutch Golden Age, Spanish, Flemish, Rococo…) moves Midjourney's mass a little (r_v η² = 0.15) and its color autonomy essentially not at all (CBI η² = 0.08). The engine defaults regardless of the tradition it is asked for.

This is the mirror image of a content-sensitivity result. Some structural response to the tradition exists — "Dutch Golden Age" does pull the engine toward a denser, darker field, consistent with the tradition's tonal character — but it explains only 15.2% of the variance in mass and less than 8.0% in color autonomy. The deepest authored behavior, color leading structure, does not arrive by naming a colorist tradition; §10 shows it does not arrive even by describing the colorist mechanism.

9. Finding 7: Same-Render Control — Mass Is Promptable

Holding the render engine constant (Midjourney) and varying only composition (448 steered portraits, same sitters) isolates treatment from render character. Because both classes are Midjourney-native, this contrast is resolution-matched for free, and the result holds at 1536, 1200, and 1024.

Mass is promptable — r_v moves toward oil (steered − naive d = 0.45, p < 10⁻¹⁰), identical at every resolution. So the dominant axis is composition, not render character. But steering is a rotation: MJ gains mass and palette economy while losing independence, contour liveliness, and midtones. Steered Midjourney is a new construction regime — not default-plus-oil — which is why it reads as a distinct third cluster, not an interpolation.

9.1 Which steering strategy moves which axis

The 448 steered images were generated under six rotated strategies; matching each image to its strategy through the prompt deck (432 of 448 recovered) lets us see the dissociation from inside the steered set. Mass strategies move mass; no strategy moves color autonomy toward oil, and the strategy explicitly built to create color autonomy moves it the wrong way.

d = Cohen's d vs the naive default; + = toward oil. The diagonal is clean for mass and absent for color — internal corroboration of §10.

9.2 Where do steered images land — interpolation or rotation?

Train the oil-vs-default classifier (3 core axes) and apply it to the steered set. If steering were interpolation toward oil, most steered images would cross to the oil side. They do not.

Steering lifts the steered set from 20% to 28% "oil" — real movement, but a minority crossing, not a flip. The steered images gained mass while shedding other authored properties, so the classifier still reads most of them as generated. This is the quantitative signature of a rotation, not an interpolation: a new construction regime adjacent to oil on one axis, distant on others.

Steering trades color for mass — it does not approach oil. Read on the three core axes, the trade is explicit: r_v moves 0.62 → 0.68 (toward oil 0.78), mu 0.24 → 0.28 (toward oil), but CBI moves 0.097 → 0.084, away from oil (0.159). Two axes improve, one degrades, and the net basin distance is flat: the steered set closes only ~2% of the naive→oil gap in deviation-Ω. The honest statement is not that steering moves Midjourney toward authored portraiture, but that it pivots — gaining composed mass at the cost of color autonomy.

10. Finding 8: Color Autonomy Is Not Reachable by Prompt

If mass is promptable, is color autonomy? Eight register-matched colorist prompts — El Greco, Rubens, Venetian colorito — aimed at the exact mechanism (colored shadows at equal value, warm–cool modeling, hue boundaries without value boundaries) tried to make Midjourney's color lead structure.

Dissociation: hue variety is promptable; color autonomy is not. Midjourney heard "color" and poured in more hues — hue_entropy shot past the oil mean — but kept every one of them coupled to luminance. The autonomy axes did not move toward oil (both p ≈ 0.26). What this establishes experimentally is narrow and firm: color autonomy is invariant under prompt manipulation, even manipulation aimed precisely at the mechanism. The most parsimonious explanation is that the color–luminance coupling lives in the weights rather than the prompt, but that is the interpretation; the invariance is the result.

This completes the dissociation that organizes the whole report: mass is promptable (§9), hue variety is promptable (this section), and color authorship is not (this section). The prompt can dodge the basin on the first two; the deepest part of authored treatment is reachable only by the weights — the argument for open-weight models as the next sandbox.

11. Robustness

Resolution, the first external attack, is ruled out. The authored images were measured at 1536, the Midjourney images near 1200 native, raising the worry that the difference was a resampling artifact. Re-measuring both classes at matched 1024 and 1200 long-edge removes it: every treatment axis survives, and the two matched resolutions agree (1024 ≈ 1200), which a resampling artifact would not.

One residual, stated plainly. The negative-space-fraction axis was found degenerate during review (near-constant in MJ, σ = 0.001, which inflated its inverse-variance weight ~130× and skewed Ω); it was dropped, costing essentially no separation and rebalancing the per-class accuracy. The guard — watch any axis whose class variance approaches the regularizer — is now part of the protocol.

12. The Recognition Test — the Gate

A coordinate system earns its keep only if its ordering matches a trained eye. The gate is a blind forced-choice test: shown two portraits, which one "holds"? Ω's pick is scored against the judge's. This is a pilot (n = 1, a generalist judge, 60 pairs), reported as a pilot.

Ω is validated at the class/medium level, not yet at within-class ordering. The judge separated authored from generated almost perfectly (the minimum "extremes" bar), but within Midjourney she picked the Ω-default as "holds" — an inversion. Several explanations remain plausible, and this pilot cannot distinguish them: (1) the judge tracked polish, not authorship — if Ω ≈ independence and a lay eye ≈ finish, then within AI "less default" reads as "less finished," and the two anti-correlate by construction; (2) Ω's within-class ordering is genuinely wrong even though its class-level resolution is right; (3) the judge was a generalist, not the brief's trained-eye target, and a portrait professional would order differently; (4) n = 1, and a single judge's within-class picks are high-variance. The decisive next step is the same 60 pairs run by trained eyes, which separates "Ω is wrong" from "the judge tracked finish." We report the inversion and decline to resolve it here.

13. Interpretive Synthesis

Authored portraiture and the generative default both center; they part company in the treatment, and most of all in how independently their image systems are allowed to behave.

The concurrent behaviors

• Centering is shared. Position does not separate the classes; both resolve to balance, artists slightly more.
• Mass vs gradient. Authored images build coherent tonal masses (r_v d = 1.25); the engine distributes fine gradient and balances with bright sheen where the artist uses shadow.
• Color autonomy. Authored color draws boundaries the light does not (CBI d = 0.85); MJ color rides luminance. The engine reproduces the Flemish/Venetian colorito divide.
• Compressed vs expansive. On the discriminating core the volume gap is modest (~1.1×); on directional construction — the choices nobody prompts for — authored work spans ~7× the space. The collapse is in the unprompted treatment.
• Promptable vs prompt-invariant. Mass and hue variety move on command; color authorship does not, even when the prompt names the exact mechanism — pointing to a property held in the weights, not the prompt.

What the coordinate system does not say

These findings locate the two classes; they do not rank them. A low CBI is not "worse" than a high one — it is a different compositional philosophy. The engine measures; it does not evaluate. The cross-engine comparison reported here is exactly the study that single-engine measurement papers defer: both datasets run through the same kernel in the same configuration, with the medium confound named and bracketed rather than hidden.