Compositional Attractor Basins
in Generative Image Models:
Using Spatial Instruction Compliance
as a Geometric Probe

What the kernel measures:
Mass centroid, not object location.

A reader looking at an image from this corpus might locate the stone visually and conclude: the stone is in the upper-left, the prompt asked for upper-left, therefore the model complied. The VTL kernel may disagree. The kernel does not detect objects. It measures the distribution of compositional energy across the image field — the aggregate of luminance gradients, edge density, tonal contrast, and spatial weight. The output is a mass centroid: a coordinate representing where the bulk of compositional energy lives. This is a fundamentally different quantity from object location.

Object location answers: where is the named subject? Mass centroid answers: where does the composition’s gravitational center reside? In a simple image with one subject on a plain background, these coincide. In a complex scene with competing elements — a horizon, atmospheric recession, tonal gradients, textural fields — they can diverge substantially.

Artists have prioritized mass placement over object placement for centuries, and the reason is structural. A painting is not a diagram with a labeled subject. It is a field of visual energy that the eye traverses according to forces the artist controls. Object placement tells the viewer what to look at. Mass placement tells the composition where to rest.

Strong compositions are built from decisions about weight, balance, and hierarchy that operate independently of semantic content. Which mass is largest, darkest, or most isolated determines the lead — the visual subject the eye returns to. These decisions happen at the level of tonal distribution and spatial weight, not at the level of named objects.

When a generative model receives a prompt asking for a stone in the upper-left corner, it is being asked to do two things simultaneously: place a named object and distribute compositional mass. A model that places the object correctly but distributes mass toward the center has technically obeyed the instruction while compositionally contradicting it. The eye lands on the stone briefly, then drifts to where the weight actually is. The composition is not about the stone.

The more specific failure mode: in several outputs in this corpus, particularly Recraft’s landscape scenes, the compositional lead has been transferred to something else entirely. The wave ripples in the wet sand carry more tonal variation than the stone. The horizon gradient creates a horizontal mass band the eye reads as primary. When the stone is placed in a corner, the void at center becomes the largest undivided field in the frame — and a large undivided field reads as compositional weight regardless of what the artist intended. The center void becomes the dominant compositional mass field.

The argument is not that any engine is wrong. It is that the tension is currently invisible. Without measurement, neither the model nor the person prompting it can know whether the composition is supporting the intended subject or quietly overriding it. The kernel makes that invisible process legible.

MidJourney’s flat overhead scene model produces images where object location and mass centroid are nearly identical. One subject on a uniform dark surface. Nothing else competes for compositional energy. When the stone moves, the centroid moves with it. The kernel and the eye agree because the scene is simple enough that they are measuring the same thing.

Recraft’s landscape scene model produces images where object location and mass centroid routinely diverge. The horizon introduces a luminance band. The atmospheric gradient distributes tonal weight vertically. A small stone correctly placed in the upper-left contributes less to the centroid than the combined luminance of sky, horizon, and ground plane beneath it.

This tradeoff has an unexpected implication. MidJourney’s reduced scene ontology becomes an advantage for certain positional compliance tasks. By collapsing the world into a flat surface, it removes the competing compositional forces that constrain Recraft. More world simulation produces more realistic outputs but reduces instruction flexibility. Less world simulation produces flatter outputs but more literal positional obedience. Neither is simply better. They are differently constrained.

Subject, prompts, and measurement.

A smooth river stone on wet dark sand was selected as the evaluation subject: no face, no obvious intrinsic orientation, no culturally encoded placement conventions, sufficient contrast for reliable mass detection. Eight prompts targeting four diagonal corners (P1–P4) and four axis-edge positions (P5–P8). Square format (1:1) specified for all generations. No prompt engineering, no iteration, no style modifiers.

Each output was passed through the VTL kernel extractor. Compliance was assessed by comparing the sign of mean displacement vectors against the expected direction. For corner prompts, both delta_x and delta_y signs must match. For edge prompts, only the relevant axis is scored.

Green labels = MidJourney    Gold labels = Recraft — Same prompts, different scene ontologies.

Scene interpretation precedes placement.

Before reporting displacement data, a qualitative observation is necessary because it conditions all quantitative findings. MidJourney interpreted all eight prompts as a top-down texture study. Recraft interpreted all eight prompts as a landscape scene with a visible horizon line, atmospheric depth, and tonal recession. This is not a stylistic difference. It is a different compositional physics.

Metric	MidJourney	Recraft	Delta	n
Mean radial distance	0.099	0.151	+53%	MJ=32, RC=16
Mean mu (mass concentration)	0.141	0.203	+44%	MJ=32, RC=16
Mean theta (angular deviation)	0.006	0.066	+1000%	MJ=32, RC=16
Mean r_v (void ratio)	0.837	0.936	+12%	MJ=32, RC=16
Centrality pressure	0.473	0.502	+6%	MJ=32, RC=16
8-prompt compliance	4/8	6/8	+25pp	8 prompts each

MidJourney theta values are consistently near zero (range 0.0003 to 0.050, mean 0.006) — the signature of a flat overhead view with no scene geometry to introduce diagonal structure. Recraft theta values are substantially higher and more variable (range 0.008 to 0.166, mean 0.066), reflecting the diagonal gradient introduced by the landscape horizon.

Outlined points show expected positions at normalized radius 0.20. Filled points show actual mean displacement. Y-axis inverted: negative delta_y = upper frame.

MJ P6: stone at right edge, high mu = tight isolation on flat surface. RC P6/P8: correct direction, lower mu = scene complexity distributes mass.

Two measurement problems, not one.

The central finding is that Recraft and MidJourney are not solving the same compositional problem. MidJourney interprets the prompt as a 2D subject placement task. Recraft interprets it as a scene-building task. Comparing spatial instruction compliance between them without accounting for this prior is like comparing two navigators who received the same destination but are driving different road networks.

Standard generative image benchmarks assess whether the output contains the requested subject. This study shows that containment is a necessary but insufficient metric. An engine can correctly generate a stone in every output while systematically failing to place it where instructed. The failure is invisible to semantic evaluation and visible only to structural measurement.

A deeper implication follows: the geometric outputs are not independent of semantic priors. Semantic interpretation reshapes the compositional field itself. The kernels in this study do not detect whether the output contains a stone. They detect the resulting field organization after the engine has decided what kind of image it is making. That decision precedes placement and substantially constrains it.

n=2 Recraft outputs per prompt (n=16 total). Mean positions computed from two outputs are statistically fragile. Conclusions about Recraft directional behavior should be treated as preliminary. At n=8 outputs per prompt per engine, confidence intervals narrow sufficiently to distinguish genuine compliance from noise at p<0.05 using a one-sample sign test. The stone-on-sand subject introduced a scene interpretation confound not anticipated at study design. Future experiments should test multiple subjects across multiple scene types.