Seven Signals That an Image Came From a Diffusion Model

The advice you read in 2023 — count the fingers, look for garbled text — is mostly obsolete. Current diffusion models render hands correctly and produce legible signage. Detection has had to move to subtler ground.

1. Noise that is too well behaved

A real camera sensor produces noise that varies independently across colour channels and increases in shadow regions. Diffusion output tends to produce noise that is statistically smooth and correlated across channels, because it emerges from a denoising process rather than from photons hitting silicon. This is the single most durable signal we measure.

2. Impossible depth of field

Generated images often show a depth-of-field falloff that no physical lens produces: the blur gradient does not correspond consistently to distance. Look for two objects at the same apparent distance where one is sharp and one is not, with no optical reason.

3. Texture that repeats without repeating

Fabric weave, foliage, gravel, crowd faces. Generated texture often has the right local statistics but the wrong global structure — it looks correct at a glance and becomes strange under magnification, without ever tiling exactly.

4. Lighting that does not resolve

Check every shadow against every light source. Generators are good at making individual shadows plausible and bad at making all of them consistent with one lighting setup. Multiple soft shadows pointing in incompatible directions is a strong signal.

5. Reflections that disagree

Mirrors, eyes, windows, water, chrome. A reflection is a second rendering of the scene, and generators frequently do not keep the two consistent. Zoom into catchlights in eyes: real ones show the actual light source shape, generated ones often show a generic blob.

6. Frequency-domain uniformity

Photographs have a characteristic frequency spectrum shaped by lens, sensor and compression. Generated images frequently show either a suspiciously clean spectrum or periodic artefacts left by upsampling layers. This is what our compression layer indirectly probes.

7. Metadata that is absent or synthetic

Real camera files carry EXIF: make, model, exposure, lens, GPS sometimes. Generated files usually carry nothing, or carry software tags. Absence proves little on its own — social platforms strip EXIF aggressively — but combined with the signals above it strengthens the case.

What still defeats all of this

Laundering. Screenshot a generated image, or run it through a platform that re-compresses uploads, and you destroy most of the artefacts listed above. You end up analysing the screenshot rather than the generation. Always seek the original file, and treat a laundered image as unverifiable rather than authentic.