A young boy with brown hair wearing an orange jacket kneels in a shallow creek, examining an old, torn map laid on the water. Behind him, large, twisted tree roots extend into the water, forming a dark, natural archway.

A Theory Stack

Failure does not equal collapse. The center is the delta.

This document suite presents an evolving visual framework developed for interpreting and pressuring AI-generated imagery. At its core is the belief that visual intelligence is not measured by polish or aesthetic fidelity defaults, but by how an image holds tension, reveals structure, and survives recursive critique. These systems were developed through hands-on usage of the Visual Thinking Lens framework, image generation, comparative analysis, and prompt-based testing, less as formal proof, more as a toolkit of a refusal engine built for probing what images are made of, where they fail, how meaning breaks down under recursion and how they break the confines of any given intent.

Together, these documents outline a hybrid visual practice built on logic, constructs and code that doesn’t pretend that an image “knows” what it’s doing. It doesn’t. This work seeks structure, recursion or pressuring alternatives. Then offering a scaffold path to rebuild and replicate. See Github for Notebooks and more.

A digital painting of a muscular man kneeling on the ground, holding a large globe on his shoulders, with cracks on the globe. The right side shows a grayscale photo of a muscular man lifting a globe with arrows indicating the forces involved, labeled with 'Weight Pushes Downward,' 'Torque Bears the Shift,' and 'Strength Lifts Upward.'

System Explainers

A Refusal Engine: Documents that help explain what the Visual Thinking Lens is and what it does.

Documentation that explains what this refusal and recursive system is, how it interrogates images and then rebuilds through visual language structure. The system is made up of individual engines and when brought together form a multi-agent refusal and reasoning engine. It reveals how generative models fail, correct, and can then restructure visual intent into structured, compositionally authored images. This framework is a prompt-native scaffold that diagnoses collapse, not decorating output.

Spatial Reasoning Failures & Diagnostics

Failure Morphologies

A diagnostic taxonomy for identifying and classifying spatial reasoning breakdowns in generative images. Seven morphologies of Ghosted Overlap, VP Drift, Tonal Inversion, Columnar Cheat, Compression Stall, Contact Dissolution, and Sequence Reversal each map distinct cognitive fault lines in how models handle depth, occlusion, and geometric coherence. Includes metric signatures, detection protocols, and cross-platform benchmark framework.

Two side-by-side pencil drawings of a young woman with dark hair tied back. The woman on the left has her eyes closed, wearing a short-sleeved top and a skirt, with an illustration of a twisted, snake-like shape emerging from her shoulder. The woman on the right has her eyes closed, wearing a short-sleeved top and a skirt, with her left arm bent at the elbow and her hand hanging down.

Stability, Drift, and Collapse

Formalizing drift, collapse, and constraint basins as reproducible fields.

Generative models tend to collapse toward the statistical mean: subjects re-center, voids fill, and fracture smooths away, producing polished but predictable results. These are tools that control deformations, identify data basins of “artist” images, offer edge-case failure modes and provide ways to work within the drift. Learn to navigate around the gravity wells of latent space in constructive and exploratory ways that produce authored images to outright collapse.

A black and white photo of four simple, smooth ceramic vases and bowls on a flat surface, with a plain background. The right side shows a stylized, high-contrast version of the same arrangement.

LSI: Structural Diagnostics & Image Assessment

Image quality assessment/art evaluation tool showing where images hold or break.

LSI is a profiled composition “stress test” for AI images for artists and researchers as an instrument to study compositional stability in AI-generated images. Unlike metrics that measure similarity to training data (FID) or text-image alignment (CLIP), this focuses on structural principles that observe how compositions hold up under iteration or modification. Includes case studies and code links to GitHub.

The Visual Cognitive Load Index

The VCLI-G is a way to measure how much visual effort an image asks from a viewer. It looks at structure, balance, voids, and tension, but not at beauty or subject matter. In simple terms, it tells you whether a picture’s complexity is “earned” (coherent, intentional) or just “busy.” By combining geometric cues like curvature, layering, and void control, it turns what artists sense intuitively into a number you can track or compare. It’s like having a visible dial for visual tension and compositional focus.

A diagram of token pathways under recursive pressure showing a flowchart with steps: start with image, then token prediction, token, recursive refinement, and end. The flow includes loops for drift, decay, and drop, with an arrow leading to recursive refinement. On the right are four paintings of still life objects such as vases, cups, and fruits, labeled with scores 6.5, 7.2, 8.1, and 8.6, depicting progress in artistic style from abstract to realistic.

Interpretability and Research Probes

Bridging Lens logic with AI interpretability and research tool use

Large models simulate visual coherence, but miss structural intent. Artists detect pressure and failure where machines default to surface gloss. How do we meet in the middle? These documents showcase logic constructs for AI, tools for tear downs, the Len’s native drift friendly scoring system, how recursive symbolic and spatial pressure can be applied and a recursive analysis of AI-generated imagery.

Painting with two sections. The left side shows a man looking up with a worried expression, and the right side shows a man with multiple faces, each holding his head with a distressed look. Text at the top says, '750 WORDS ISN'T ENOUGH' on both sides.

Artistic Extensions

Pushing into symbolic recursion, refusal, and design philosophy

Explore tension, refusal, and symbolic ambiguity as design tools. Test whether an image can hold together under symbolic and spatial pressure. This is documentation that identifies where form collapses and where tension sustains volume, offers new approaches to prompting that prioritizes structural tension over aesthetic description or shows how to track prompts that bend token behavior under structural pressure.

Two portraits of the same man, on the left a realistic painting and on the right a abstract cubist style version, both featuring a man wearing a hat, against a background with houses.

Proofs and Walkthroughs

Largely taken directly from threads, lightly edited to help with flow and to take out transitional default closing prompts.

Take walkthroughs with a single image across recursive iteration, learn about critical strain, recursive refusal, or refusal as structure. This suite of papers are the side steps and exploration of how to step out of the aesthetic center defaults.

—> Author’s Note: The frameworks, metrics, and methodologies presented across these documents operate at multiple levels: from computational measurement (RCA-2, VCLI-G, LSI-lite) to dialectical prompting strategies to conceptual models of how optimization shapes generative systems.
What this work provides:

Measurement Infrastructure: Tools like the Radial Compliance Analyzer (RCA-2), Visual Cognitive Load Index (VCLI-G), and Latent Structure Inventory (LSI-lite) quantify compositional properties that were previously discussed only qualitatively. These produce reproducible numerical results: coefficient of variation, interquartile ranges, geometric distributions.
Operational Frameworks: Systems like the Visual Topology Language (VTL) and dialectical prompting protocols translate compositional intent into actionable prompt structures. These aren't metaphors, they're tested methods for navigating optimization landscapes and forcing models into less-traveled regions of their output space.
Mechanistic Hypotheses: Proposals about preference optimization, denoising dynamics, and computational efficiency as drivers of compositional collapse. These are testable claims grounded in measurable phenomena, not speculative analogies.
Conceptual Models: Terms like "compositional basin," "geometric scaffold," "repair mechanisms," and "forbidden zones" function as both descriptive language and analytical constructs. They organize observations into coherent explanations of system behavior.

What this work is not: This is not reverse-engineering of model internals. I don't have access to training data, model weights, or optimization trajectories. The claims about *why* compositional collapse occurs (preference tuning, denoising stability, attention efficiency) are informed hypotheses based on measured outputs and known architectural principles and not definitive statements about what's happening inside the black box. The frameworks use structured language and visual metaphors to make complex system behaviors legible and manipulable. When I describe mass "organizing in rings" or prompts "pushing against priors," I'm translating geometric measurements and observed prompt responses into accessible explanations. The measurements are literal; the mechanistic explanations are evidence-based inference.

In practice:

RCA-2 coefficient of variation of 4.09% is a measured fact, not metaphor
"Radial prior as path of least resistance" is an explanatory framework for that measurement
VTL protocol forcing Δx > 0.4 is an operational method with documented results
“Repair mechanisms" names the observable phenomenon where models resist off-center requests

This work lives at the intersection of measurement, method, and interpretation, where quantitative analysis meets practical application meets mechanistic understanding. It's practitioner research: built from production needs, validated through systematic testing, explained through frameworks that bridge technical precision and conceptual clarity. If you're using the notebooks: trust the numbers. If you're reading the papers: engage critically with the hypotheses. If you're applying the methods: expect them to work, at times with some trial and error, then measure whether they do.