Score the frame, do not trust your eye
The problem with judging your own composition is that you know the answer. You look straight at the character because you put the character there. A viewer arriving cold looks at whatever is loudest, and those are usually different places.
The figure is the subject and it barely registers, because it sits at nearly the same luminance as its background and has no color of its own. The small sign in the corner is the loudest thing in the frame by a wide margin. Nobody staged it that way; it is just what the numbers say.
What the map is made of
Two terms, which are the two that carry most of the weight in the published models:
- Center-surround luminance contrast. How far this point is from the average of a ring around it. This is the core of the Itti-Koch model and it is what makes edges and isolated bright things score.
- Color uniqueness. How far this point's saturation is from the frame's baseline. One saturated object in a desaturated frame is an enormous attractor, which is why the sign wins.
What is missing matters as much. There is no orientation channel, no face detection, and above all no motion. Motion outranks everything on this list, so a still saliency map is the floor of the analysis rather than the whole of it. A slowly drifting background element can beat a stationary subject that scores twice as high here.
Using it
The value is not the picture, it is the ranking. Three questions, in order:
- Is the subject in the top three hot regions? If not, the frame is not staged, whatever the layout says.
- What is beating it? Usually one specific thing: a highlight, a saturated prop, an interface element, a bright sky through a window.
- Can that thing lose without losing the scene? Desaturating a background sign by a third usually costs nothing and moves the ranking.
The fixes are almost always subtractive. Adding brightness to the subject raises the overall level and rarely changes the order; taking it away from the competition changes the order immediately.
The levers, ranked by how much they move the score
- Saturation of everything that is not the subject. Cheapest and largest.
- Local contrast against the immediate surround, which is not the same as the subject's absolute brightness. A mid-gray figure on a dark ground beats a bright figure on a bright ground.
- Isolation. Empty space around the subject raises its center-surround score without touching its own values at all.
- Edge density. A busy texture behind the subject destroys it. Blur or simplify what is behind, not what is in front.
- Motion, which is not in this map and which beats all of the above. If something is moving, that is the subject.
Where this earns its keep
On anything a viewer sees once and briefly, where there is no second chance to find the subject: a store page capsule, a thumbnail, a trailer frame, a cutscene beat that holds for a second. On a screen the player will sit with for ten minutes it matters much less, because they will find everything eventually.
Rules of thumb
- You cannot judge your own composition, because you already know where the subject is.
- Two terms get you most of a useful map: center-surround luminance contrast, and color uniqueness.
- Read the ranking, not the picture. The subject should be in the top three regions.
- Fix by subtracting from the competition, not by adding to the subject.
- Local contrast beats absolute brightness. Isolation is free score.
- Motion outranks everything and a still map cannot see it. Check stills and motion separately.
One email when something new goes up. No newsletter, no schedule, nothing else.
Double opt-in, so watch for a confirmation email. Unsubscribe any time.