This tool does not use AI to decide if an image is fake. Instead, it extracts forensic views that let you judge. Just because an image doesn't show signs of AI generation, does not mean it is real.

Noise Residual — Strips away image content to reveal the underlying noise pattern. Real photos have organic, uniform noise. AI-generated images often show unnatural smoothness, grids, or repeating textures here.

FFT Power Spectrum — Shows how noise energy is distributed across frequencies. Bright spots or regular patterns in the spectrum can indicate artifacts from generative model architectures.

Error Level Analysis (ELA) — Re-compresses the image as JPEG and measures the difference. In an unedited photo, all regions should have similar error levels. Spliced, painted, or AI-generated regions often compress differently. You can adjust the JPEG quality and brightness scale to fine-tune what's visible.

Block Artifact Grid — JPEG compression operates on 8×8 pixel blocks. This view measures the gradient energy at every 8-pixel boundary and highlights blocks that deviate from the norm. Red/yellow = grid mismatch (splicing), blue = suspiciously smooth (AI/inpainting), dark green = normal.

PCA Minor Component — Runs Principal Component Analysis on the RGB channels and shows the least significant component. This decorrelates color information to surface manipulation artifacts, cloning, or inpainting invisible to the naked eye.

Local Noise Variance — Measures the standard deviation of noise in 16×16 pixel windows and renders a heatmap. Authentic photos have relatively uniform noise; edited or AI-generated regions often have noticeably different noise levels.

Luminance Gradient — Computes the Sobel gradient magnitude across the image. Pasted objects often have subtly different edge characteristics — too sharp, too smooth, or double-edged — compared to natural edges in the scene.

Statistical Moments — Shows kurtosis, skewness, entropy, and standard deviation for each color channel across a 4×4 grid of regions. Real camera noise follows predictable distributions; synthetic or manipulated regions often deviate.

Please check the reference section at the bottom of the page to compare your results with the behavior from other image types.