Guide · 9 min read · Updated September 2026

Signal two: physical consistency, or what light refuses to do

A camera is a physical instrument recording a physical world, and that world has rules. Light travels in straight lines from a finite number of sources. Reflective surfaces return what is actually in front of them. Parallel lines converge on a horizon at eye level. A lens focuses at one distance and blurs smoothly away from it. Physical consistency analysis asks whether an image obeys these rules — and unlike artefact analysis, it survives compression, because geometry is carried by the large structures in a picture rather than the fine detail.

By Ira Peoples

Shadows: direction, length and hardness

Start by finding every shadow in the frame and tracing the direction each one points. Outdoors in daylight there is one dominant source, the sun, effectively at infinite distance — so shadows run parallel and all point the same way. Indoors there may be several sources, but each object should cast one shadow per source, and every shadow of the same source should agree about where it is.

Length encodes the source's elevation: a low sun makes long shadows, a high one short. All objects on flat ground should show the same ratio of shadow length to height. Hardness encodes the source's size: a small hard source like the sun or a bare bulb gives crisp edges, a large soft one like an overcast sky or a softbox gives diffuse ones. A crisp shadow next to a diffuse one in the same light is a contradiction, and objects added to a scene very often bring the wrong kind.

The shadow that is missing

More common than a wrong shadow is no shadow at all. An object resting on a surface must show contact shadow — the dark compression right where it meets the ground — or it appears to hover. Generated scenes and composites routinely include a cast shadow while omitting contact, producing a subtle floating quality that people register as odd without identifying why.

Check the ground beneath feet, table legs, cups and wheels specifically. It is a two-second look and it catches a great deal.

Reflections: the hardest thing to fake

A reflection is a second rendering of the scene from a different viewpoint, and getting it right requires understanding the three-dimensional layout — which is exactly what a statistical image model does not have. Mirrors, windows, still water, polished floors, car paint, sunglasses and eyes all carry reflections.

Check that what is reflected is what should be reflected: the same objects, in the reversed arrangement, at the correct scale, occluded in the way geometry demands. Eyes deserve particular attention. Both eyes of a person under a single light source should show catchlights in the same relative position and of the same shape. Different catchlight shapes in the two eyes, or catchlights that do not correspond to any visible light, is one of the strongest single signals available in a portrait.

Perspective and scale

Parallel lines in the world — building edges, road markings, tiled floors, shelving — converge towards vanishing points on the horizon, and the horizon sits at the camera's eye level. Trace the strong lines in an image and see whether they agree on where those points are. A composited element often brings its own perspective from wherever it was cut out.

Scale follows from the same geometry. A person twice as far away appears half as tall, and figures standing on the same plane at the same distance should measure consistently. Synthetic crowds frequently contain a person who is subtly the wrong size for their position, and doorways, chairs and vehicles are useful rulers because their real dimensions are narrow.

Depth of field and atmosphere

A lens focuses on a plane, and blur increases smoothly with distance from it. That means blur should be a function of depth and nothing else. When a background is uniformly soft regardless of how far away its parts are, or when an object is sharp while something at the same distance beside it is blurred, the blur was applied rather than captured.

Over larger distances the air itself intervenes: haze reduces contrast and shifts distant objects towards the sky's colour. A far mountain range rendered with the same contrast and saturation as the foreground has skipped atmospheric perspective, which is a common tell in generated landscapes.

Motion, in video

Video extends physics into time. Objects have momentum, gravity accelerates falling things at a constant rate, and feet planted on the ground must stay planted while the body moves over them. Watch a synthetic walk cycle for feet that slide, and watch loose material — hair, clothing, water, smoke — for motion that is fluid but does not respond to what the body is doing.

Frame-by-frame stepping is the technique. Consistency over time is expensive for a model to maintain, so details that should persist often drift: a pattern that changes, an item of jewellery that moves, a background object that reshapes between frames.

Where physics checks mislead

Real scenes break these rules more often than you would expect. Multiple light sources produce multiple shadows in directions that look wrong until you find the second lamp. Wide-angle lenses stretch objects near the frame edge, making scale comparisons unreliable. Panoramas and phone computational photography stitch several exposures together, producing genuine discontinuities in an unmanipulated photo. Reflective surfaces are often curved, which distorts the reflection legitimately.

The discipline is to look for the mundane explanation before the exotic one. A single inconsistency you cannot explain is worth noting; two or three independent geometric contradictions in one frame is when physical consistency becomes a serious finding. Because these signals survive compression so well, they are often the only family still readable in a heavily shared image — which makes them disproportionately valuable, and worth checking by eye even after an automated verdict comes back inconclusive.

The short version

Geometry survives compression, so physics checks work on images where artefact analysis cannot.
Shadows must agree on direction, length ratio and hardness; missing contact shadow is a frequent tell.
Reflections and eye catchlights require 3D understanding, which is where synthetic images fail hardest.
Blur should be a function of depth alone, and distant objects should lose contrast to the air.
Wide lenses, multiple lamps and panoramas break these rules legitimately — look for the mundane explanation first.