Known-Architecture Latent Resonance: Non-Invasive Generative Image Attribution via VAE Reconstruction Harmonics

dataset - Known-Architecture Latent Resonance: Non-InvasiveGenerative Image Attribution via VAEReconstruction Harmonics This benchmark dataset provides the complete empirical evaluation records, per-image predictions, and execution pipelines for Known-Architecture Generative Image Attribution under matched latent autoencoder surrogates (stabilityai/sd-vae-ft-mse). Evaluated on N=1,000 authentic images (500 MS-COCO Flickr photographs vs 500 DiffusionDB Stable Diffusion 1.x text-to-image samples), it records:- Deterministic VAE reconstruction PSNR (26.64 dB real vs 29.11 dB SD 1.x, gap +2.47 dB, Cohen's d = 0.519, AUROC 0.6331).- Azimuthal 2D-FFT harmonic spike ratios at the Nyquist lattice frequency f=64 (1.013 real vs 1.110 SD 1.x, AUROC 0.7258).- Joint calibrated resonance ranking (AUROC 0.7475). Includes full per-image prediction CSVs, the executable GPU Google Colab benchmark notebook, distribution figures, and comprehensive documentation. All evaluations follow a strictly non-circular protocol excluding synthetic 1/f noise simulations or unconditioned random-latent decodes.