What Is a Tracking Shot?

What the camera is actually doing

In a tracking shot the camera moves through space. That is the whole distinction, and it is a physical one rather than a stylistic one. A pan rotates from a fixed point, so the background swings past with flat parallax. When the camera travels, near objects slide by faster than far ones, and the eye reads that difference as "I am moving with this person."

The rig matters less than the geometry. Dollies on rails, Steadicam, gimbals, car mounts, drones, and a phone taped to a skateboard all produce the same class of shot. What separates a controlled one from a wobbly one is whether the distance between lens and subject holds steady, because that constant distance is what keeps the subject the same size in frame while everything behind them changes.

The four geometries worth knowing

• Follow. Camera behind the subject, shoulder height. Withholds the face, builds anticipation, and reveals the space to the audience at the character's pace.
• Lead. Camera ahead of the subject, moving backwards. You watch the reaction instead of the destination, which suits a character walking in with a decision already made.
• Lateral. Camera parallel to the subject. This is the walk-and-talk workhorse and the version with the strongest parallax, so it reads most clearly as movement.
• Orbit. Camera circling a subject who stays roughly in place. Slow reads as intimacy, fast reads as entrapment.

What the move buys you

A travelling move spends screen time to buy continuity. Nothing is cut away, so the audience accepts that the space is real, connected, and being crossed in real time. That is why it is the default tool for arrivals, escapes, and any sequence where geography carries meaning.

It also costs something. A travelling move locks you into one performance and one path, and it removes the editor's ability to fix pace later. If the walk is boring, the shot is boring, and there is nowhere to cut to.

Prompting a tracking shot in AI video

Camera movement is the least reliable thing to get out of a video model, because the model has to separate camera motion from subject motion inside one sentence. What actually works:

• Put the camera in its own clause, first. camera tracks left alongside a man running beats a man running, tracking shot. Subject-first prompts usually animate the person and leave the camera parked.
• Use screen-relative directions. camera tracks right, camera follows from behind, camera pushes in are understood. camera moves east, camera goes around the corner, camera follows him upstairs mostly are not.
• State what must stay constant: constant distance, subject stays centered in frame. This single clause removes most drift.
• Give the near foreground some texture: railings, parked cars, columns, doorways. Parallax is what makes the move legible, and a flat background often collapses into a slow zoom that reads as nothing at all.

Three failure modes show up constantly. The camera drifts one way and then reverses partway through the clip, because nothing in the prompt commits it to a direction. The subject skates forward with no leg cycle, because the model is translating the whole figure instead of animating a walk (adding walking, full stride, feet visible helps). And the background warps as new geometry is invented at the edge of frame, which gets worse the faster and longer the move is.

Shorter and slower fixes all three. If you need a long travelling move, generate it in segments and feed the last frame of one clip in as the first frame of the next.