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?
Not necessarily. Record what is uncertain so later production does not rely on it as fact. Use Reverse Prompt for a draft visual brief, then verify occluded and fast moments yourself before using the description in a film.