Latent Resonance: Zero-Shot Autoencoder Inversion and Azimuthal Spectral Forensics for Diffusion Image Attribution

Abstract

Modern latent diffusion models (LDMs) synthesizephotographic imagery with fidelity that evades human perceptionand frustrates conventional convolutional neural network classi-fiers. Existing forensic detectors rely predominantly on supervisedbinary classifiers that overfit to specific semantic domainsor depend on closed API parameters. This paper introducesLatent Resonance, a zero-shot, white-box-free forensic frameworkgrounded in the physical and architectural asymmetries betweenoptical camera acquisition and latent autoencoder synthesis. Bymapping candidate images through a deterministic autoencoderprojection ˆx = D(μ(E(x))) without noise injection (σ = 0),we isolate two distinct physical phenomena: spatial manifoldresonance and transposed-convolution harmonic lattice spikes.Synthetic diffusion images originate directly from the decodermanifold D(z), yielding near-zero spatial reconstruction error(MSE = 0.000809 ± 0.000013, PSNR = 36.94 ± 0.07 dB).Conversely, authentic optical photographs suffer irreversible lossof physical sensor shot noise and Photo-Response Non-Uniformity(PRNU) across the 8× latent spatial bottleneck, producingsignificantly higher residual error (MSE = 0.001955 ± 0.000062,PSNR = 33.11 ± 0.14 dB, ∆ = +3.83 dB). In the frequencydomain, azimuthally averaged 2D Fast Fourier Transform (2D-FFT) analysis reveals that diffusion residuals exhibit sharpperiodic harmonic spikes (2.268× baseline) induced by transposedconvolution upsampling strides, whereas authentic photos followa smooth, continuous 1/f α power-law decay (1.141×). On cleannative images, Latent Resonance achieves a verified 100.00%AUROC with zero false accusations. Under rigorous adversarialstress testing (lossy JPEG Q ∈ {95, 85, 75}, bicubic resampling384 → 512, and Gaussian blur σ = 0.8), we demonstrate thatcombining spatial reconstruction margins with azimuthal spectralharmonics provides robust forensic discrimination where single-domain detectors collapse. All models, benchmark pipelines, andinteractive inspection interfaces are released as a fully open-sourcereproducible artifact.Index Terms—Image Forensics, Latent Diffusion Models, Vari-ational Autoencoders, 2D-FFT Spectral Analysis, Sensor Noise(PRNU), Transposed Convolution Harmonics, Open Science.