AI generates infinite subjects.

It repeats the same structure.

An applied AI evaluation framework for generative systems, measuring compositional bias and structural behavior across major platforms.

The Visual Thinking Lens measures what CLIP, FID, and human review pipelines miss: the spatial geometry generative models repeatedly produce.

Text-to-image systems can generate endless semantic variation, people, city streets, animals, objects, but place them within highly constrained geometric patterns under standard prompting conditions. Across evaluated outputs, semantic diversity explains less than 10% of observed spatial variance.

Composition is not prompt-driven.
It is prior model-driven.

A collage of eight images depicting various scenes of nature, urban landscapes, and abstract digital art, with red concentric circles overlayed on all images.

Current benchmarks measure if the butterfly looks like a butterfly.
This measures spatial reasoning and priors.

A side-by-side comparison showing a monarch butterfly on the left and a different butterfly on a plant on the right.

400 MidJourney prompts. 8 semantic categories. One geometric attractor.

Only 34% of horizontal space used (central attractor is only Δx = 0.005 ± 0.044)

100% of outputs within 0.15 radius of center

Semantic categories explain 6% of spatial variance

Different prompts, same pattern across engines, identical compositional bias. VTL measures the signature each engine learned from its training data and the spatial prior it applies regardless of what you ask for.

A collage of diverse images including urban scenes, nature, animals, architecture, objects, and abstract art.
Scatter plot showing the relationship between change in multiple semantic categories, with Delta x on the x-axis and r-v on the y-axis, containing various colored dots representing data points.

Introducing the Kernel

Our seven geometric primitives fingerprint any image:

  • Δx,y: Where mass sits (placement offset)

  • rᵥ: How much void surrounds it

  • ρᵣ: How compressed the marks are

  • μ: How unified the composition reads

  • xₚ: How hard the edges pull

  • θ: Orientation stability

  • ds: Structural thickness / surface depth

Run through python or a multi-engine, recursive critique field, then apply structural intelligence to prompts, compositions, and symbolic logic. It (re)builds imagery in the ways defaults cannot see. Stable, reproducible and invisible to semantic evaluation.

Diagram of an apple illustrating various concepts: void ratio, peripheral pull, placement offset, structural thickness, packing density, compositional cohesion, orientation stability, with arrows pointing to different parts of the apple and a color scale for packing density.

The Visual Thinking Lens Breaks the Pattern

A structural engine where making, breaking, and seeing are one recursive act.

These images span subjects, styles, and engines. All push against the geometric default into authorship. For before and after, see the library.

The Kernel becomes coordinates AI uses to organize space. We measure them because they're stable, reproducible, and invisible to semantic evaluation, and when applied back into AI, it gains spatial reasoning.

The image features two plots side by side. The left plot is a gradient heatmap with a density of points towards the top center, titled 'Sora Δx - rv Forbidden Heatfield.' The right plot is a box plot titled 'Sora: Compositional Metrics — Distribution Tightness (Normalized)' showing several metrics like delta_x, r_v, mho_r, mu, x_p, theta, and d_s, with a series of data points and whiskers indicating spread and distribution.

Where AI Won't Go: Evidence from 200 Sora Prompts

These aren't failures of capability. They're learned constraints. AI models have discovered that certain compositional coordinates reliably fail human evaluation, so they've learned to avoid them, even when you explicitly request them - but that doesn’t mean you have to.

Example: Extreme edge crops (Δx > 0.52) + high void = systematic refusal


Close-up of a frog with large eyes, shown in three different views: close-up top, left labeled 'Default Collapse,' and right labeled 'Generative Steering'.

Stable Territories in Compositional Space

Through systematic perturbation testing, VTL can steer toward constraint regions where AI maintains compositional integrity under stress. These aren't aesthetic styles. They're geometric regimes that resist AI's pull toward center.

Artist basins are stable territories where AI maintains compositional integrity under constraint, but often need steered to. The off-center third or the peripheral anchor or compressed mass. AI has them in latent space, we providesthe coordinates to navigate toward them.

The VTL couples a generative-physics model (how images behave as mass in a field) with a multi-engine critique OS (how different analytic voices transform or interrogate that mass) to steer. The frog has the same semantic prompt, but different geometric instruction, moving from center to steered = 0.28, basin-navigated, vs cropped.


Sequence of four photos of a beige ceramic vase against a light beige background, with lighting changes creating shadows.

What else can the Kernel do? Soft Collapse Shows in Structure First

This is a deterministic structural regression layer that monitors compositional stability across model releases and fine-tuning updates. Model degradation appears in compositional metrics 3-4 inference steps before semantic breakdown. Within the kernel, Δx drift, void compression and peripheral dissolution all signal trouble while the image still looks fine to other metrics.

The kernel detects degradation, which matters for training evaluation, A/B testing, and quality monitoring at scale.


Flowchart illustrating a framework and user learning loop for image generation, featuring steps from user input, structural refinement, lens analysis, lens output, to score and analysis.

A Recursive Lab for visual intelligence.

VTL isn't just diagnostic. It converts generative output into measurable structure:

  • Fingerprinting: stable and repeatable metrics

  • Steering toward specific compositional territories

  • Detecting: change under perturbation and pre-failure degradation

  • Controllability measurement: architectural differences through geometric signatures

  • Drift monitoring: Identifies “snap-back” or collapse behavior

  • Consistency: compare models or releases against a baseline

Consumer tools chase style. Research metrics chase numbers. The Lens chases spatial reasoning.

This is not about beauty or style

  1. This is not a prompt framework

  2. This is not subjective taste scoring

  3. This is structural diagnostics for generative systems


Comparison of three drawings of women with corresponding technical analysis and guides, including aligned photographs and sketches with markings for centroids, frame centers, and bounding boxes.

The Lens is portable, reproducible and easy.

VTL runs in top-tier conversational AI (Claude, GPT, Gemini) for measurement and steering. Python code offers drift free bench marks.

The logic is portable. The output can be research or discovery.

What works everywhere:

  • Cross-model fingerprinting (Sora, MidJourney, GPT, SDXL, Firefly, OpenArt)

  • Deterministic geometric measurement, no aesthetic judgment, no black-box scoring

  • Reproducible analysis via Jupyter notebooks or conversational AI

Core capabilities:

  • Image Fingerprinting - Compare engines by compositional signatures

  • Predictive Steering - Treat prompts as forces in geometric space, estimate drift and snap-back, try to push past it

  • Cross-Domain Analysis - Map visual geometry to rhetorical stance and narrative tension

  • Training Archaeology - Reverse-engineer learned priors from attractor behavior)

Models arrange space before they arrange meaning. VTL exposes the geometry priors before a model interprets meaning.


Comparison of four groups of images showing stability, collapse, and compositional safety concepts, featuring a white sphere, a potted succulent plant, a cardboard box, and a portrait of a woman in different lighting and background conditions.

It’s a system artists, engineers, and models can all step into.

Researchers: A deterministic post-generation instrumentation layer that flags structural instability and behavioral regression in generative systems.

Product Teams: Quality monitoring at scale. A lightweight structural fingerprinting system for monitoring behavioral consistency in generative production pipelines.

Creative tools: A structural controllability diagnostic that measures whether user edits actually reshape composition or collapse back to default priors.

New to VTL? Begin here:

  1. Mass, Not Subject - Foundational concept (15 min)

  2. Kernel Primitives - Core measurements (10 min)

  3. Monoculture in MidJourney - Empirical evidence (20 min)

Want practical application?

  1. Deformation Operator Playbook - Hands-on techniques

  2. The Off-Center Prior- Basin navigation

  3. Foreshortening Recipe Book - Constraint architecture

Researcher or engineer?

  1. Generative Field Framework - Complete technical spec

  2. VCLI-G Documentation - Measurement methodology

  3. GitHub link - Reproducible implementations

If you still believe prompts control composition, complete research package available.