Describe the visible state before and after an occlusion, then mark the hidden action as uncertain. A person passing behind a pillar does not prove which hand they used or what happened out of view. A useful reverse prompt distinguishes recorded evidence from the new action you choose to invent for a recreation.
Locate the missing evidence
Pause before the subject disappears and again when they return. Record which objects are visible, who holds them and the direction of movement. Then note the interval where the view is blocked. A brief silhouette or reflection may provide partial evidence, but it is not permission to describe a fully observed sequence.
Occlusion is different from a cut. A pillar can block the subject while the same camera view continues. A cut replaces the whole image with another viewpoint. It is also different from a sampled-video gap: an analysis system may not inspect every frame even when the original file contains them.
Keep observation and reconstruction separate
Observed: the person enters behind the stone pillar holding a closed box in the left hand. After the blocked interval, the person emerges with the box open. The opening movement and the right hand are not visible. Recreation choice: use a separate close-up showing the same box opening once, then return to the wider view with the box already open.
This example supplies an editorial bridge without pretending to recover the hidden original prompt. If fidelity to the visible clip is the priority, keep the action hidden in the recreation too. If the goal is a clearer tutorial or story, add an explicitly new shot rather than smuggling speculation into the source description.
Check fast intervals against the original
Google's video-understanding documentation describes sampled video analysis and configurable frame handling. It is a useful reminder to verify brief actions in the source; it does not establish the sampling configuration of Imagild or guarantee that every contact was analyzed.
For a critical moment, review the original at normal speed and advance through the relevant frames. If you still cannot see contact, keep the uncertainty. Do not resolve it with a more confident adjective. A prompt can be executable while remaining honest about what the reference reveals.
Choose the right level of detail
| Evidence | Useful wording | Wording to avoid |
|---|---|---|
| Hand visible before blockage | The left hand holds the box | The right hand secretly turns a latch |
| Object changed afterward | The box is open on reappearance | The source definitely shows an opening action |
| Subject exits and returns | Movement resumes from the right edge | Exact unseen walking path |
For a recreation, decide whether the missing action needs a separate shot or a deliberate time omission. Either is better than a dense instruction that asks the model to keep the view blocked while also showing hidden details clearly.
Can AI infer the most likely action?
It can propose one, but probability is not observation. Label the proposal as a reconstruction choice and check that it connects the visible states plausibly.
Should uncertain details be deleted entirely?
Keep the uncertainty that affects the recreation: “Box opens during an interval blocked by the pillar; mechanism and hand are not visible.” Omit unnecessary guesses, but preserve the missing-evidence note so another writer cannot mistake a proposed bridge for an observed event.
Filled occlusion record: no source view can show the latch
In this constructed source, the camera remains continuous while a stone pillar blocks the box between 2 and 3 seconds. Manual frame review of that interval still shows stone, not the hands. This distinguishes a physical source occlusion from a summary that simply failed to include visible frames.
| Source interval | Hypothetical visible evidence | Reconstruction decision |
|---|---|---|
| 1.8 seconds | Closed box in the left hand | Retain the observed state |
| 2.0–3.0 seconds | Pillar blocks both hands and box in every reviewed frame | Mark opening mechanism and hand uncertain; denser extraction cannot reveal it |
| 3.2 seconds | Box is open on reappearance | Retain changed state without claiming the opening was observed |
For a fidelity-focused recreation, keep the pillar and the same deliberate omission. For an explanatory tutorial, propose a new close-up of a latch opening and label it new staging. Do not describe the new shot as footage recovered from behind the pillar. This choice is different from the sampling-gap guide, where the original video actually contains usable contact frames that the selected sample set omitted.
Workflow boundary and sources
This is documentation-based guidance with constructed examples, not a reported generation or application test. Use Imagild Film Script for a written plan and inspect the actual supported input roles in Film Prompt before reference-video recreation. External software controls belong to their named products. No exact reconstruction, native analyzer, existing-clip repair or ranking result is promised.
- Google Gemini Video Understanding — Sampled visual analysis can miss rapid events; Google API processing configuration is not an Imagild setting.
- Runway Image to Video Prompting Guide — Motion planning and source/temporal review principle
Editorial revision: November 13, 2026.