A hidden face is the documented failure case for motion transfer

The shot

Motion-transfer tools map a recorded performer's movement onto a different character, a feature relevant to eyeline and screen-direction continuity whenever the mapped performance has to cut against footage shot the ordinary way. Runway documents this feature as Act-Two, which transfers a driving performance video's movement, speech, and expression onto a supplied character image or video.

What the documents show

Runway's own best-practice list for Act-Two states specific input constraints as documented requirements, not suggestions: the performance video must feature "a single subject," must "ensure the subject's face remains visible throughout the video," and should frame the subject "at furthest, from the waist up"; "no cuts that interrupt the shot" are allowed, and certain expressions "are not supported." These are the vendor's own stated conditions for reliable output, meaning occlusion of the face or a camera angle that loses it falls outside documented, supported use. A separate, independent academic study, "Can Pose Transfer Models Generate Realistic Human Motion?", tested three pose-transfer methods -- AnimateAnyone, MagicAnimate, and ExAvatar, not Runway's own model -- on actions and identities outside each model's training data, and found that participants in a controlled study correctly identified the intended action only 42.92 percent of the time, and judged the generated motion consistent with the source video only 36.46 percent of the time. This is an independent test with a stated method and named models, distinct from any vendor's own claim.

The workflow

Per Runway's documentation, an operator records or selects a driving performance that keeps the subject's face, and for gesture control the subject's hands, in frame throughout, without mid-shot cuts, then applies it to a character image or video and adjusts a facial-expressiveness setting that the documentation states trades expressiveness against character consistency and visual artifacts at its extremes. Matching the driving performance's implied camera angle to the character reference's angle is left to the operator; Runway's documentation does not describe an automatic correction for a mismatch.

What the tool does not change

Neither the vendor documentation nor the independent study claims motion transfer removes the need for a human check on whether the transferred performance still reads correctly for eyeline and screen direction once cut against surrounding footage; the cited study's low action-identification rate on out-of-distribution material is a documented reason to check rather than assume.

• Does the driving performance keep the subject's face unoccluded for its full duration, as Runway's documentation requires?
• Is the requested action within the range the cited independent study actually tested, or well outside it?
• Has the transferred shot been checked for eyeline match against the shots it will cut next to?

Reading the vendor's input requirements next to an independent test of a different set of models shows the same shape of risk from two directions: documented constraints on one side, and a measured failure rate on unfamiliar action outside them on the other.