AI image generation has reached the point where human visual inspection alone cannot reliably distinguish synthetic images from authentic photographs. Modern diffusion models, the technology behind Midjourney, DALL-E, and Stable Diffusion, generate images with photorealistic detail, natural lighting, and convincing facial features. The result is that visual inspection alone catches fewer than 30% of AI-generated images. Reliable detection requires analyzing the forensic signals embedded in (or conspicuously absent from) image files.

Check EXIF metadata first

Every photograph taken by a camera or smartphone embeds EXIF metadata, camera make and model, GPS coordinates, shooting parameters, and software tags. AI generators do not embed this data. A JPEG with no EXIF is strongly suspicious. A PNG with no EXIF is expected (PNG rarely carries EXIF), but PNG is itself the default output format for most AI generators. If a JPEG has EXIF but no camera make or model, it may have been processed by editing software after generation. Any image whose software tag references Midjourney, Stable Diffusion, DALL-E, or a known AI platform is definitively synthetic.

Analyze sensor noise patterns

Real camera sensors introduce small amounts of random noise into every pixel, even in uniformly colored areas like clear sky or smooth walls. This noise is an inherent physical property of semiconductor image sensors, it cannot be eliminated without also destroying image detail. AI-generated images do not have this noise. Diffusion model outputs are mathematically smooth in flat regions, with near-zero pixel-to-pixel variation where a camera would show 2 to 5 luminance units of noise. Measuring noise in flat regions of an image is one of the most reliable forensic signals for AI detection.

Check image dimensions

Diffusion models require image dimensions to be multiples of 64 (or 8 for some architectures) due to how the U-Net backbone processes images. Common AI output sizes include 512x512, 768x768, 1024x1024, 896x1152, 1344x768, and 1216x832. If an image is one of these exact sizes, it was almost certainly generated by an AI model. Camera sensors do not produce images in these dimensions, a smartphone typically outputs 4032x3024 or similar, and professional cameras produce even larger non-square outputs.

Use a free AI image detector

The ScanErase free deepfake detector analyzes all four forensic signals automatically, EXIF metadata, sensor noise, dimension patterns, and color statistics, and returns a probability score. The analysis runs entirely in your browser using the Canvas API, so your image is never uploaded to any server. You can check any JPEG, PNG, or WebP image in seconds with no account required.

Frequently asked questions