Chapter 25. Action Continuity and the Handoff Frame#

25.1 A continuous action is a cross-shot contract#

When an action begins in a medium shot and completes in a close-up, both shots must share one handoff state. If the previous shot ends with the document ten centimeters above the table, the next cannot begin at chest height — unless the cut deliberately elides and the direction of movement agrees.

A handoff contract stores action progress, hands, props, body orientation, speed, eyeline and emotion. It is far more reliable than "continue the previous action."

25.2 Action phases#

Break an action into preparation, initiation, passing, completion and aftermath. Cut points usually sit in the movement after initiation, or after completion and before the reaction.

Figure 25-1 Three phases of passing a document. Preparation establishes control and direction. The handoff carries the document across the midline of the space. Completion confirms the change of ownership. The orange path is the action; the vertical marks are the available editing handoff window.

What matters most in the diagram is not how far the hand travelled but who owns the object at each stage. During preparation the document is still under Lin Yun's control. During the handoff, whether the other party touches it decides whether the action can complete. At completion, the document's position, which hands are occupied, and the relationship have all changed. If the next shot skips one of those steps, the edit must supply a comprehensible elision — the document cannot change owner without cause.

action_handoff:
  action: ACT_RAISE_AUTHORIZATION
  shot_out: E001_S08
  phase_out: passing
  state_out:
    right_hand: gripping the top-right corner of the document
    prop_height: 10cm above the table
    motion_vector: upward, slightly toward frame right
    gaze: the document
  shot_in: E001_S09
  phase_in: passing
  state_in: same
  phase_end: completed_at_chest

25.3 A handoff frame is not automatically usable#

The previous shot's last frame can serve as a compositional reference for the next, and it also carries compression, motion blur and any small facial error forward. Select a handoff candidate frame from the stable end section of the previous shot, and clean it into a bridging image where necessary.

The next shot references three things at once: the golden identity image, the location and wardrobe assets, and the handoff frame. The handoff frame governs action state; the golden assets govern re-anchoring. Their responsibilities are not interchangeable.

25.4 Where to cut on a matched action#

Cutting inside an action lets the audience overlook differences in background and angle. The outgoing shot shows the intent; the incoming shot shows the critical detail. Direction, speed and phase must be compatible either side of the cut.

Scissors approaching the badge can begin in a medium shot, cut to a close-up of the hand completing the cut, then cut to fragments landing on the table. An impact sound can carry across the cut. Do not have the badge cut in the outgoing shot and cut again in the incoming one, unless the repetition is a deliberate emphasis.

25.5 Direction and speed of motion#

A character moving frame left to right should continue in that direction in the next shot. A hand raising cannot suddenly descend. Speed may change at a cut with motivation — the action being blocked, for instance.

Record motion_vector and a speed level. A generative model does not need exact physical values, but "continue upward at the same speed and stop on completion" is far more specific than "she raises the document."

25.6 Hands and prop contact#

Contact is a high-risk zone. Confirm the hand is empty before the action, that finger positions are plausible during contact, and that prop ownership updates afterwards. When a model cannot stably show a grasp, use a three-shot illusion: the hand approaching; sound with partial occlusion; the next shot with the object already held.

Audiences accept elision. They do not accept a prop appearing without cause. Sound, eyeline and reaction can complete a contact that was never shown.

25.7 Head, eyeline and expression handoffs#

When a head turn crosses a cut, record the starting angle, direction and final target. If the previous shot already ends looking at Lin Wei, the next cannot start from frontal and turn again. Expression inherits too: a peak of surprise can settle back into restraint, with a breath or a regulating action.

Characters re-emerging after closing their eyes, covering their faces or turning away are prime sites for identity drift. Re-inject the golden reference for the matching angle in the next shot.

25.8 Edit handles#

Generate roughly half a second to a second of stable handle before and after each shot, adjusted to the action. Models frequently wobble or deform as they start and finish; the final edit uses the stable middle.

Handles are not meaningless stillness. The leading handle keeps the preparation, the trailing handle keeps the completed state and an optional reaction — which gives the edit several possible cut points.

25.9 First- and last-frame constraints#

Generation that supports first and last frames works for simple displacement, provided the two states are physically reachable and no complex contact is required in between. A hand at the table edge to a document at the chest is reasonable. Standing alone to already embracing another person will fuse the bodies in between.

The last frame is not a wish list. Design the action path first, then decide whether one video can complete it.

25.10 Strategy for splitting a continuous action#

Split complex action by informational value: establish the position, show the intent, the key contact, the result, the reaction. Not every step has to be shown.

For the badge in E001: a medium of security raising the scissors; an extreme close-up of the blades closing; the fragments landing on the agreement; Lin Yun looking at them; Lin Wei ordering her removed. The physical contact is brief, the proof object is clear, and the emotional consequence is complete.

25.11 Options after a failure#

When action continuity fails, ask in order: can the middle be deleted; can an insert or reaction cover it; can the action be simplified; can it be composited in layers; is it worth regenerating.

Do not burn budget repeatedly out of fidelity to the original shot list. What the story needs is the badge identity being stripped away — not necessarily every millimeter of the scissors.

25.12 Action continuity QC#

Review frame by frame for hands, props and body first. Then watch at normal speed for the feel of the movement. Then listen to the sound alone to judge impacts and rhythm. Frame-accurate correctness does not guarantee naturalness at speed, and smooth motion at speed can conceal a prop jump.

Record one of four outcomes: handoff passes, repairable in the edit, needs regeneration, needs a shot design change.

25.13 SOP for action handoffs#

First, define phases for cross-shot actions. Second, choose the cut point and the shared state. Third, record hands, props, direction, speed, eyeline and emotion. Fourth, generate leading and trailing handles. Fifth, select a handoff candidate from the stable end section. Sixth, have the next shot reference identity, assets and handoff frame together. Seventh, check motion direction and whether the action repeats. Eighth, review frame by frame, at speed, and by sound alone. Ninth, on failure, work through deleting the process, covering, simplifying, compositing, and only then regenerating.

25.14 Fault tree#

Symptom: the action happens twice across two shots. Phases were not recorded and the outgoing shot already completed it. Move the cut point or delete the duplicated section.

Symptom: the prop jumps to a different height instantly. The handoff states disagree. Redo the incoming shot's opening frame or cover the elision with an insert.

Symptom: the action connects and the face has drifted. Only the previous frame was referenced, with no identity re-anchor. Add the angle-matched golden reference.

Symptom: the direction of movement reverses. Screen direction is wrong, or footage was mirrored. Regenerate, and check carefully whether flipping is safe at all.

Symptom: repeated generation still fails on the grab. You treated a high-risk contact as mandatory to show. Use the three-shot illusion of intent, occlusion, sound and result.

25.15 Checklist, exercises and deliverables#

Check that actions are split into phases; that handoffs share state; that hands and props agree; that direction and speed match; that edit handles exist; that the handoff frame is clean; that the next shot carries the golden assets back; and that failures are simplified before being regenerated.

Exercise one: design the raising of the document as two shots and write the handoff contract. Exercise two: design an alternative badge-cutting sequence that never shows contact. Exercise three: review one action three ways — frame by frame, at speed, and by sound only. Exercise four: simulate five shots of drift referencing only the previous frame, then add a re-anchor.

Deliverables for this chapter: the action handoff, the phase table, handoff candidate frames, the reference policy, the handle plan, action continuity QC, and the repair decision record.

25.16 An action contract constrains both shots, not just the second#

A handoff is not the incoming shot unilaterally catching up. An action contract states the actor, the object, the world direction, the starting phase, the cut phase, the ending phase, the change of prop ownership, and the features visible in both shots. If the outgoing shot has already passed the contract's cut phase, even a correct incoming shot repeats the action. If the incoming shot starts before the cut phase, the action runs backwards.

action_contract:
  id: AC_E001_AUTH_HANDOFF
  action: Lin Yun pushes the authorization toward Zhou Lan
  world_vector: door_to_table_head
  shot_a:
    phase_in: file_under_right_hand
    cut_window: file_crosses_table_seam_40_to_55_percent
  shot_b:
    phase_in: file_crosses_table_seam_45_percent
    phase_out: file_stops_before_zhoulan_left_hand
  invariant:
    - document face up
    - pushed with Lin Yun's right hand
    - Zhou Lan has not touched it yet
  ownership_event: none_until_next_shot

A range in the contract is more useful than a single frame. Hitting an exact frame in generated motion is difficult; supplying a 40–55 percent cut window is not. The editor then looks inside both windows for the position where speed, shape and direction agree most closely, which preserves choice.

25.17 Sound is the second continuity track#

Visual contact can be elided; sound events cannot drift arbitrarily. Footsteps, paper friction, a lock, an impact and breath all help the audience read two generated clips as one action. Add sound events to the action contract: attack, peak, tail and spatial position. The cut may land before or after the sound peak, but an impact cannot precede the contact, and a document landing on a table cannot have no weight.

A sound bridge also covers elision in picture. The outgoing shot shows scissors approaching the badge; cut to Lin Yun's reaction with a crisp snap on the track; then cut to fragments landing. The audience completes the physical process that was never shown. Here sound is not a patch but part of the design: it reduces the generation risk of hand contact while increasing subjective impact.

Play any action review at least three times: muted, to check physical direction; sound only, to judge rhythm; and together, to check causality. Smooth picture with repeated sound makes the action seem to happen twice. Correct sound with a jump in ownership means the elision was never properly established.

25.18 Substitute grammars for frequently difficult actions#

Handing over an object can run as approach, occlusion, already held. An embrace can run as one person entering close distance, a shoulder filling frame, a reaction close-up. Getting into a car can run as the door opening, a hand on the frame, already seated. A fall can run as the loss of balance, an object landing with its sound, the state on the floor. Removing a coat can run as beginning to unfasten, action off screen, the coat already over the chair back. Substitute grammars keep the causality the audience needs without insisting on the intermediate physics models handle worst.

Confirm the story value before substituting. If a chip covertly exchanged during a handshake is the core evidence, the contact cannot be fully elided — use layered compositing, real photographed hands, or a separate prop shot. If the handshake merely signals cooperation, reaction and sound suffice. Shot difficulty is set by narrative necessity, not by technical display.

Teams can build an interaction pattern library, storing for each pattern the applicable conditions, the recommended shots, the common failures and the acceptance criteria. It is not a fixed template but a store of production experience. Each time a new kind of action is solved, distil the reusable structure so later projects need less trial and error.

25.19 The regression scope after repairing an action#

Redoing one handoff shot affects at minimum the tail of the previous shot, the head of the next, the prop state, the sound sync and the character's emotion. After a repair, do not review only the new shot. The system generates a minimum regression pack from the action contract: the last two seconds of the previous shot, the new shot, the first two seconds of the next, the corresponding sound effects, and the state differences.

If the repair changes prop ownership or the moment of completion, the impact can propagate down the sequence. If the document originally passed to Zhou Lan at S10 and now passes at S12, Zhou Lan leafing through it in S11 becomes impossible. The dependency graph should mark every shot reading that state. Only replacements that change no state and correct local deformation can be regressed within a single shot.

Action continuity is not complete when the cut feels smooth. It is complete when physics, state, sound and emotion all hold. Any track that connects only through audience misreading must be explicitly accepted or repaired before release.

A note on sources#

Short-form generation suits single brief actions, and traditional matched-action and insert cutting can organize several stable fragments into one continuous event. This chapter moves that editorial logic upstream of generation, which reduces wasted retries.