LSI: Structural Diagnostics & Case Studies

An image quality assessment and AI art evaluation tool showing where images hold or break.

LSI-lite: A Composition Analysis Tool 

LSI-lite is a profiled composition “stress test” for AI images that measures Δx (off-center gravity), rᵥ (void ratio), and ρᵣ (rupture/mark energy), combines them into LSI_lite_100 (0–100). Essentially it is MVP compositional image quality assessment tool that studies structural stability/failure on single images or across iterations rather than aesthetics. It’s practical for education, batch analysis, and AI guardrails and can run solo or complement industry tools such as FID/CLIP/SSIM. GitHub Link (v2)

LSI: lite Case Studies 

LSI-lite case studies apply a profiled composition stress test across varied subjects (landscape, portrait, interior, still life, abstract, figure, animal), reading Δx (off-center gravity), rᵥ (void ratio), and ρᵣ (rupture/mark energy). These showcase LSI-lite as a research 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. The goal is not to replace human artistic judgment, but to provide a systematic way to discuss why some compositions feel more stable than others.

LSI: lite Color Telemetry - Addendum

This addition to the LSI-lite adds Color Telemetry, which lets LSI look in color alongside grayscale. Purely for diagnostics, not for changing the score or acceptance. It computes a color-based subject/background mask and a luminance-only balance read, and exports those plus simple “difference” values. It provides a one-line Color audit when those color reads meaningfully disagree with the gray read, flagging where color may be skewing the composition. This is a first step in adding color to the compositional “stress test” and go beyond just B&W.

LSI: Structural Composition for Images

LSI measures composition with three pre-attentive primitives—Δx (gravity), rᵥ (void), and ρᵣ (rupture)—and makes a per-frame decision via the Compositional Triage Gate (CTG_100): ≥55 and no RED bands. Color/Tonal telemetry (rᵥ/Δx deltas, tonal span/separability/bright-mass) runs in parallel, shares the Δx ROI, and never gates—providing concise, one-line audits that explain why a frame looks structurally off. Built on that base, optional θ and dₛ plus sequence diagnostics (S/K/R and a barycentric A/B/V map) read behavior across batches and iterations, positioning LSI as a complement to fidelity or prompt-alignment metrics, not a replacement. Contact if interested.

Lens Structural Index (LSI) Full Proposal

LSI is a research diagnostic that treats images as dynamical systems, measuring five primitives: Δx (centroid offset), rᵥ (void ratio), ρᵣ (rupture proxy), θ (tilt), and dₛ (skeleton depth) and tracking their trajectories to derive Stability (S), Consequence (K), Recursion (R), and a composite LSI100 with an “in-basin” validity gate. It complements FID/CLIP by asking whether structure holds under pressure (not just resembles data), offers recursive, perturbation, and single-frame modes, and is explicitly non-universal/bias-revealing for artists and engineers to debug compositional resilience. Contact if interested.

Lens Structural Index (LSI) Case Studies 

LSI is presented as a structural diagnostic (not an aesthetic judge) that stress-tests images under iteration using five primitives: centroid offset, void distribution, edge/rupture integrity, axis alignment, and depth, with an intentional volumetric bias to expose AI’s compositional assumptions. The case studies map distinct regimes and highlights productive friction between machine reads and artist intuition. The paper positions LSI as a research instrument (anti-universal by design), and suggests adjunct flags (e.g., Refusal Mode, Frame Placement Index, Axis Integrity, Directional Coherence, Domain Presets, Field-Noise Split) to contextualize scores and reveal blind spots. Contact if interested

Quiet Color, Loud Structure: Why Gray-Only Evaluators Miss Late Cézanne, and How to Fix It

Most image quality evaluators assume compositional structure lives in luminance (light-dark edges), which causes them to undercount or reject images where color relationships carry spatial separation. This small study leverages LSI-lite and introduces a lightweight telemetry (TEL) layer that runs alongside a stable grayscale gate, measuring when color-derived metrics diverge from grayscale-derived metrics without changing acceptance thresholds. Applied to nine Cézanne Mont Sainte-Victoire paintings (1878–1906), the TEL signal flips from negative (early works: grayscale-dominant) to positive (late works: color-dominant), quantifying Cézanne's known shift toward using color as structural architecture rather than decoration—a pattern gray-only evaluators systematically miss.

Scroll Structure, Not Style: 1700-21st Century Traditional Scrolls vs. 2025 AI-generated Scrolls

This study compares old scroll paintings to AI-generated scrolls. Traditional artists leave more blank paper (89% vs 84%) and spread their ink across a much wider area of the page (76% vs 54% of the width). AI basically makes a narrow column of content with margins around it, while real scroll artists treat the whole page as compositional space. The takeaway: AI doesn't understand scroll format, it just centers stuff, so to fix it you need to explicitly push for more emptiness and wider distribution of marks across the page.

Beyond the Default: How a Visual Lens Shifts AI Out of Center-Fill

The analysis shows that AI tends to create images that are safe and centered, like putting a subject directly in the middle of a page. Using the Visual Thinking Lens and LSI-lite, this study tracks structural metrics like void ratio, lane width, and mark density to quantify whether images make deliberate compositional choices rather than statistical ones. Through a 16-iteration recursive process, the study shows measurable divergence from AI norms with more empty space, broader structural lanes, sparser marks, and livelier rhythm, demonstrating that “failure” can signify creative refusal, not collapse. The result is a method to measure intentional composition and reclaim authorship from algorithmic equilibrium.

Artist-AI Exploration as Recursive Process: Using Geometric Metrics to Guide Latent Space Navigation

This study argues that common AI “quality” scores encode centered, aesthetic defaults that mask geometric information and suppress exploration. In a 68-iteration, artist-led study, trajectories traverse measurably distinct basins (stable operating regions) and include frames that reward models would flag as “failed,” yet remain geometrically coherent and artistically useful. The result is a practical separation of measurement (what an image is, geometrically) from preference (what a brand or curator wants), enabling reproducible exploration of generative systems without encoded aesthetic bias.