Generative Criticality is Not Observed in Pixel-Space Measurements of Latent Diffusion Models under Linear Projection Constraints

Latent Diffusion Models (LDMs) enforce topological closure, biasing image synthesis toward discrete objects. This paper investigates whether this bias can be overcome to achieve "generative criticality"—a regime of distributed structure without objecthood—using an inference-time method called Orthogonal Gradient Projection (OGP). Our findings show a consistent negative result: instead of producing scale-free distributed structures, the system undergoes an abrupt collapse from object-centric organization into periodic and low-variance regimes. Quantitative analysis using FFT spectral metrics, Total Variation, Gradient Variance, and connected-component analysis reveals a sharp, non-linear transition under minimal constraint. Rather than an engineering limitation, this collapse is interpreted as a conceptual insight into the structural rigidity of latent diffusion systems. The results suggest that object-centric organization is a deeply embedded property of current generative architectures, and that attempts to suppress it reveal discrete aesthetic boundaries rather than continuous transformations.