Keeping a face the same across cuts is still a manual task

The shot

Shot-reverse-shot is the coverage pattern that carries most dialogue scenes: a setup on one speaker, then a second setup showing the reverse angle, cut together so the conversation reads as continuous. A production guide's explanation of the pattern states that it "refers to when a filmmaker places a camera setup on a subject, and then uses a subsequent setup to show the reverse view," and notes the standard practice is to keep focal length, shot size and depth of field matched across both setups so the cut does not call attention to itself.

What the documents show

Holding a performer's face, wardrobe and framing steady across those two setups is exactly what reference-image conditioning is meant to do, and vendor documentation treats it as a workflow to manage rather than a guarantee. Runway's own article on using reference media lists "generating consistent characters across different scenes, lighting, and styles from a single image" as a use case, without describing an automatic check that two separate generations of the same character match. Luma's own character and object consistency guide is more direct about the limit: a single reference image "has to invent what it cannot see" from a new angle, which the guide names as "drift," and it recommends building a multi-angle reference pack, locking specific features in writing, and issuing correction prompts when a detail changes between versions.

The workflow

Applied to a shot-reverse-shot pair, that documentation implies a specific sequence: build reference images of the performer from the angles both the shot and its reverse will use, describe the features that must not change, generate both setups against that shared reference, and compare the results before cutting. Luma's guide frames this as an ongoing task through a scene rather than a one-time setup, since it recommends re-anchoring to the original reference pack rather than to whatever version was generated most recently.

What the tool does not change

Neither vendor's documentation claims its reference feature is a guarantee; Luma's own material names drift as an expected risk to correct, not an edge case. Reviewing two generated setups side by side and deciding whether a viewer would read them as the same person remains an editorial judgment, the same check a script supervisor would make on a physical set.

• Do the reference images cover the angles both halves of the reverse-shot pair actually need?
• Which specific features were locked in writing before either setup was generated?
• Was the second setup checked against the original reference, or only against the first generated shot?

A vendor's stated feature is a starting point here, not a finish line: both companies' own documentation describes consistency as something a generation workflow has to maintain across a scene, not a property a single model call delivers on request.