A camera solved to fill the frame converges on a face-on view and throws the perspective away

Fact

Auto-framing is usually written as an optimisation: search the camera's angle and distance, keep whatever puts the most of the subject on screen. For a subject that is broadly planar — a board, a playfield, a floor plan, a page — that objective has one maximum, and it is looking straight down the plane's normal. Projected area is largest face-on, and every degree of obliqueness costs it.

So the solver reliably returns the most orthographic view available to it. The framing is optimal and the picture is flat: the depth cues the 3D view existed to provide are exactly what the objective was discarding.

Why it matters

Both halves look correct in isolation. The fill number really is higher, and nothing is wrong with the projection, so the result reads as "the perspective is too weak" and gets chased in field of view or in lighting, which cannot fix it — the angle is the variable, and it was chosen deliberately.

It bites hardest on a portrait viewport, where the shortfall from an oblique angle is largest, so the flattening appears on phones and in narrow panes and not on the developer's wide window.

How to apply

• Constrain the search to the range of angles that still read as the view you want, and optimise fill only inside it. A clamp is the fix, not a workaround: the objective is genuinely incomplete.
• Prefer solving distance and lens shift, which do not change the read, over solving orientation, which does.
• Recentre by shifting the frustum (m[8], m[9] in a column-major projection) rather than moving the camera; it slides the image without touching the perspective, and shifts NDC by exactly the amount set.
• Fit against every corner of the subject's box, not a bounding sphere, which is orientation-blind — see [[framing-a-box-is-a-per-corner-solve]].