Chapter 78. Action and Ensemble Lab: Breaking an Uncontrollable Shot into a State Relay#

The test scene has Lin Xia handing the red folder to Gu Zhou while Lin Wei tries to stop it from the background. It contains two principal faces, four hands, a critical prop, a shift of eyeline, dialogue and a third party's reaction — the classic high-risk combination for a generation system.

78.1 Action facts, three-phase boards, and diagnosing failure#

Write action facts, not shot wishes.

There are only five action facts: the folder starts in Lin Xia's left hand; Gu Zhou's right hand starts on the table; both hold it briefly; Lin Xia releases first; at the end Gu Zhou holds it and Lin Wei has seen it. A push-in, a focal length, slow motion and dramatic expressions are not action facts.

action_invariants:
  object: prop_red_contract_folder_01
  start_holder: char_lin_xia.left_hand
  receiving_hand: char_gu_zhou.right_hand
  dual_grip_min_frames: 6
  release_order: lin_then_gu_secures
  end_holder: char_gu_zhou.right_hand
  observer: char_lin_wei

If those five cannot be determined from the text, the shot does not enter generation.

The three-phase action board.

Preparation: Lin Xia moves the folder from beside her body to the table edge while Gu Zhou's hand stays still. Handoff: his right hand contacts the spine, both hold it, eyelines maintained. Completion: Lin Xia releases, Gu Zhou draws it into his own area, Lin Wei reacts.

Produce one state frame per phase and accept each on hands, prop orientation and character slots separately. The video model's job is connecting states that already hold — not inventing the whole process at once.

Why the single-shot version failed.

The first version asked for 4.5 seconds in which Lin Xia speaks, pushes the document, Gu Zhou turns and receives it, Lin Wei stands to intervene, and the camera orbits the table. The results duplicated the folder, fused fingers, swapped Gu Zhou's hands and crossed the axis. The root cause is too many concurrent state changes, not a prompt lacking cinematic flavor.

Complexity can be read roughly as the number of simultaneously changing objects, multiplied by contact relationships, multiplied by identities. Shortening the duration does not necessarily reduce it; compressing five actions into two seconds is often less stable.

78.2 Cuttable plans, ensemble attention, and screen direction#

Four cuttable plans.

Plan A is a two-person medium carrying the complete handoff. B is three inserts: Lin Xia's hand, the slide across the table, Gu Zhou's hand. C pushes the document, cuts to Lin Wei's reaction, then cuts to Gu Zhou already holding it. D has the folder placed on the table and Gu Zhou picking it up in the next shot. A is the most continuous and the riskiest. B is physically clearest. C is the most dramatic. D is the most stable and weakens the direct transaction between the two.

The team generated B first as insurance, then attempted A. A succeeded on the fourth attempt and entered the main cut, with B retained as a fallback version. Insurance coverage is approved at shot design — it is not improvised after failure.

Controlling attention in an ensemble.

With three people in frame, only one primary action is permitted. During the handoff peak, Lin Wei stays still in the background with only an eyeline move; her standing goes into the next shot. If a background character moves substantially at the same time, the model can transfer action ownership and the audience misses the folder's state.

The cast layout fixes Lin Xia foreground left, Gu Zhou mid-ground right, and Lin Wei background centre-right. Expression intensities are 2, 1 and 3 respectively, while only Lin Xia and Gu Zhou need high identity detail. Resource allocation follows the shot's responsibility.

Eyelines, axis, and screen direction.

Lin Xia passes the document toward frame right and Gu Zhou receives toward frame left. Hand inserts generated separately must inherit the same motion direction and table texture. One candidate had beautiful hands with the folder sliding left; cut in, that returns the prop to Lin Xia, so it was eliminated.

Lin Wei's reaction shot keeps her looking down-left toward the handoff area. Eyeline targets are written as world objects rather than as an abstract instruction to look left, because frame left stops being reliable once the camera moves.

78.3 Contact acceptance, fault injection, and the experiment report#

Accepting a contact action.

Check frame by frame for penetration before contact, at least six frames of dual grip, stable fingers, no duplicated folder, the release order, the end position and the shot's handles. A candidate that looks approximately right with a wrong end state does not go to the editor.

Once the action passes, review the performance: Lin Xia does not look down at her hand as she passes it, which reads as trusting the action; Gu Zhou takes half a beat of judgment before receiving; Lin Wei's reaction occurs after the dual-grip peak, so she does not appear to know in advance.

Fault injection and repair routing.

Deliberately mirror Lin Xia's reference and lint should block at the handedness layer. Delete the handoff peak and the state check should find the holder teleporting from Lin Xia to Gu Zhou. Move Lin Wei's reaction earlier and the knowledge sequencing check should alarm. Swap Gu Zhou's camera position and the axis rule should demand re-establishing.

Of those errors, only local finger deformation might route to local repair. Holder, reaction timing and axis errors belong to the design or selection layer and cannot be painted over in post.

SOP, checklist and deliverables.

Extract the action facts. Write start, peak and end states. Build the three-phase board. Compute concurrency complexity. Design the primary plan and insurance coverage. Fix slots and action ownership. Generate a low-resolution previsualization. Accept frame by frame. Submit action state events. Add failures to the regression set.

• One primary state change per shot.
• Contact actions include a dual-grip peak.
• Inserts inherit motion direction and material.
• Background reactions never precede their visible trigger.
• Insurance coverage is approved before the budget is spent.

Exercise: design a six-shot state chain for handing over a glass, receiving it and the glass shattering, delivering action_contract.yaml, phase_frames/, coverage_matrix.csv, continuity_events.jsonl and frame_qc_report.md.

The action experiment report.

Compare plans across five columns: are the action facts preserved, does the audience understand, is the physics credible, is it usable in the edit, and is the cost bearable. Even after succeeding on the fourth attempt, single-shot A is not automatically better than three-shot B. If A lacks handles either side and cannot adapt to a different rhythm, B may carry more production value.

Action knowledge must state its preconditions. "Document slide inserts are stable" depends on a flat table, a single prop, low-reflectance material and a fixed direction of movement — it does not generalize to two people struggling over a glass. Failed candidates keep thumbnails and structured reasons and are barred from being selected as references again. When building a recipe library for frequent actions, store the prohibited conditions and the insurance coverage alongside, or the recipe will only propagate hidden errors.

Before promoting a recipe, re-test it across different characters, sleeve cuffs and table materials. Only when state correctness holds across two scenes does it graduate from a lesson about one shot into a project-level action template.

A note on sources#

This lab reconstructs an action decomposition method. What transfers is writing invariants before shot wishes, budgeting concurrency, and approving insurance coverage as part of design rather than as a reaction to failure.