Universal AI Image Detector
State-of-the-art universal detector for synthetic AI imagery. Powered by frozen CLIP ViT-L/14 feature representations, 2D FFT frequency lattice analysis, PRNU sensor noise forensics, and C2PA provenance triage.
Analyzing Image with UnivFD...
Forensic Inspection Completed
UnivFD detected clear synthetic neural feature signatures consistent with modern generative models. Ambiguous feature distribution. The image may be heavily compressed, stylized, or an artistic composite. The image aligns with natural real-world camera optics and Poisson-Gaussian sensor noise patterns.
UnivFD (Universal Fake Detector, CVPR 2023) maps images into CLIP ViT-L/14 high-dimensional latent space. Generative models (GANs, Diffusion, Flow Matching) inhabit distinct sub-manifolds from authentic camera captures.
Spatial Patch Breakdown
Evaluation across 5 strategic image regions to detect localized editing or face replacement:
2D Fast Fourier Transform (FFT) decomposes pixel luminance into spatial frequency components. Generative upsampling (e.g. transposed convolution) creates unnatural periodic grid spikes along cardinal axes.
Radial frequency energy exhibits high-frequency distribution with an axial spike ratio of .
Interpreting the Map: High-frequency lattice dots andεε-axis rays reveal artificial convolutional upsampling (e.g. Stable Diffusion UNet layers).
Physical camera sensors exhibit Poisson-Gaussian shot noise and Photo-Response Non-Uniformity (PRNU). AI-generated imagery exhibits synthetic smoothing, missing grain, or uniform mathematical noise.
High-pass Laplacian residual variance measured at .
Interpreting the Map: Physical optical camera sensors exhibit fine, stochastic grain (PRNU). Pure synthetic images exhibit smooth patches or unnatural uniform noise.
Cryptographic C2PA manifests, EXIF camera hardware parameters, and PNG generation parameters (prompts, samplers, model hashes).
Detected Generator:
Physical Hardware Camera:
No embedded generation tags or EXIF hardware serials discovered. Assessment is derived purely from neural & spectral feature extraction.
How UnivFD Universal AI Detection Works
Most AI image detectors are trained on a single generator and fail when tested on new diffusion architectures. UnivFD solves this fundamental generalization problem.
1. Why CLIP ViT-L/14 Generalizes
UnivFD leverages OpenAI CLIP ViT-L/14 backbone, pre-trained on 400M diverse image-text pairs. Frozen features preserve rich semantic and structural invariants, while a trained linear probe separates authentic photos from synthetic distributions.
2. 2D Spectral Frequency Signatures
Neural generators produce subtle checkerboard frequency artifacts due to deconvolutional upsampling. 2D FFT spectral analysis unmasks these invisible mathematical fingerprints regardless of post-processing filters.
3. Multi-Crop Inpainting & Face Swap Detection
By segmenting the image into spatial crops, UnivFD pinpoints whether an image is 100% synthetic or an authentic photo with localized AI-generated modifications (e.g. generative fill or deepfake face swaps).