Beyond Seed Integers: Spectral Features of the Initial Noise Tensor as Predictors of Structural Fidelity in Diffusion-Based Image Generation

A prior study (Ticudean, 2026b) established that 32-bit seed integer encoding is effectively useless for structural fidelity prediction in Stable Diffusion 1.5; seed bit models achieve AUC-ROC of 0.456-0.478, worse than random. The root cause is the Mersenne Twister PRNG: designed for maximal decorrelation, it destroys any recoverable relationship between the seed integer and the noise tensor geometry it generates. I propose bypassing the integer entirely and extracting spectral features directly from the 4×64×64 noise tensor each seed produces, operating at the true causal level of the diffusion process. Applying this to the existing 50,000-image labeled dataset (no new generation required), spectral feature models achieve AUC-ROC of 0.916-0.932, with logistic regression reaching 0.932, a 0.45-point gain over seed bits and above the 0.859 best from the prior study. Individual spectral features contribute two to three times more importance than seed bits, confirming that noise tensor geometry carries genuine predictive signal. Output type remains the dominant predictor (type_line_chart, 0.430), consistent with prior findings.