Lighting errors are subtler than anatomy errors but often more decisive, because a scene's lighting is a single physical system. Every shadow in a real photograph traces back to the same set of light sources. A generative model does not simulate that system. It paints each region of the image with locally plausible shading, and the local choices do not always add up globally.
The three checks that catch most cases:
Shadow direction. Find the brightest highlight on a face or object and ask which direction the light must be coming from. Then look at the shadows cast on the ground or on a nearby surface. If the highlight says the light is upper-left and the cast shadow falls to the right, the two cannot both be true.
Shadow presence. A strong light source produces a shadow. Look for objects that are brightly lit but cast nothing, or cast a shadow that is far too short or too soft for the light described.
Multiple light sources. Count the highlights in a pair of eyes, or the specular highlights on a shiny object. Two highlights in each eye imply two light sources. If the rest of the scene shows only one shadow direction, the lighting is inconsistent.
Reflections follow the same logic. A mirror, a window, a lake, or a pair of sunglasses must show a reflected version of the scene that is consistent with the camera's position. Common failures: a reflection that shows a different background than the one behind the subject, a reflection of a person who is not in the frame, or a reflection that is simply blurred into an unrecognizable smear.
These checks are more reliable than anatomy checks in one respect: they are harder to produce by accident. A real camera cannot easily create a shadow pointing the wrong way. But they are also harder to spot, because you have to reason about the scene rather than just count. That is why they belong in the second pass, after you have scanned the obvious features.