AI in automotive marketing, the real version.

Only if you make it. Brand paint has to read true across every lighting condition, so the color rides in as a locked reference on every generation, never a per-shot judgment call. Badges, caliper text, and wheel patterns drift when their reference is small in frame, so any part under scrutiny gets its own big, clean, single-subject reference bound on every shot where it is in focus.

The lesson that generalizes: the fix for a drifting detail is a better reference of exactly that detail, not more adjectives in the prompt.

Because every shot is re-imagining the environment unless you stop it. On one luxury launch film, the single biggest client note spanned eight timecodes: two lanes with a center line in one cut, an unmarked single lane in the next, dashed white in a third, with the horizon drifting behind it.

The fix is one locked environment reference applied identically across the entire sequence, plus a QC pass specifically hunting for lane markings creeping back in. Video models have a strong prior for painted lines. You have to hold the road still on purpose.

It does if you skip the grounding step. A vehicle composited into a scene has to inherit that scene's actual light: the direction, the panel highlights, the shadow softness, the ambient color in the paint. We read those off the base image and write them into the generation explicitly.

Same discipline on motion: physics rules like smooth planted movement and consistent light states are stated at the start, middle, and end of every prompt, because a rule stated once does not reliably hold.

A shoot day is a shot list plus setups. Keep the shot list, replace the setups. From a small reference set of the real vehicle, we generate the same coverage a director would block: the low angle through the window, the through-the-windshield hero, the over-the-shoulder wide, the mirror-reflection detail with real focus falloff, the hands-on-wheel macro.

And because the shot list is locked, variants are nearly free: re-fire the same list with a different trim, wardrobe, or color reference and get the matching set back.

Against traditional CGI: no 3D model of the vehicle has to be built, rigged, and lit before the first frame. Coverage comes from photographs and direction. Drafts run at review resolution and only the client's final picks get the full-resolution render, so revision cycles stay fast.

And yes, the review process looks exactly like a shoot review: a marked-up deck with circled timecodes. The difference is the fix. A flagged beat means a targeted regeneration of that beat, not a new shoot day. One more honest note: a technically correct cut can still read flat for a car spot. Pacing has to be directed on purpose, the same way light is.