Moving a close-up subject clear of a dialog by moving the camera sinks the camera into the ground

Fact

To show a subject a share f of the half-height above the centre of the screen by translating the camera (or lowering its aim point, the same thing), the camera must move down by f · d · tan(fov/2): in proportion to its distance. In a close-up that is about the subject's own height. On an upright phone (80° vertical field, d = 3.4) clearing a sheet over the bottom 60 % takes 1.7 units, and a camera at head height ends at the subject's feet.

A view offset (an off-axis frustum, what a shift lens does) slides the picture by exactly f and leaves the camera where it was: the subject keeps its shape and nothing new comes between them. Tilting the camera down instead puts the subject off-axis, where a wide field stretches and keystones it.

Why it matters

Lowering the aim works at follow distance on a wide screen, so it gets reused for close-ups on phones and fails there in ways that look unrelated: a step in the floor hides the subject, a collision ray cast from the lowered aim point starts inside the ground and snaps the camera in, and the shot turns worm's-eye.

How to apply

• three.js: camera.setViewOffset(camera.aspect, 1, 0, share, camera.aspect, 1) lifts the picture by share of the screen height; clearViewOffset() undoes it. setViewOffset sets aspect = fullWidth / fullHeight, so normalised arguments such as (1, 1, …) squash the picture.
• Picking through unproject and frustum culling stay right, since they read the same projection matrix; updateProjectionMatrix() keeps the offset.
• Measure the room from the dialog's layout box (offsetTop; a pop-in transform skews getBoundingClientRect()), and again on every resize.
• When the room is shorter than the subject, stand back as well: it fills height / (2 · d · tan(fov/2)) of the screen. See also [[auto-framing-collapses-to-face-on]].