Begin with coverage
A report with three maps does not establish attention across every scene. List the video’s scene boundaries, then mark where an analyzed frame is available. If a brief product reveal falls between samples, write “not analyzed” rather than borrowing a neighboring map.
This is a sampling diagram, not a fabricated model output. A real overlay must come from the matching analyzed image. TranSalNet’s paper describes image-saliency prediction; selected frame predictions do not record continuous gaze through a video.
Fill the scene worksheet
Download the scene review CSV. Review the original export at normal speed before filling it.
| Field | What to enter |
|---|---|
| Creative version | Exact filename or report identifier |
| Scene interval | Start and end time in the source video |
| Analyzed timestamp | Actual frame time, or “not analyzed” |
| Intended focus | Product, brand, benefit, proof or call to action |
| Map observation | What competes within this frame |
| Other evidence | Visible text, audio context or observed playback issue |
| Proposed edit | Specific change and what to preserve |
| Verification | What the revised export must visibly do |
The worksheet is an editorial tool. It does not imply that the app automatically generates a map for every row.
A filled example: observation, gap, decision
This is a fictional teaching example. The bottle, intervals and observations below are authored to demonstrate the worksheet. No model was run, no heatmap is supplied, and no viewer or campaign outcome was measured.
Imagine a six-second refillable-bottle ad: a pack shot, a refill demonstration and an end card. The review question is whether the refill action and brand are legible. A completed row should preserve uncertainty rather than invent a prediction.
Opening: 0–2 seconds
- Source observation: the fictional pack is visible, but its name is covered by a caption.
- Analyzed timestamp / map observation: not analyzed; no model output available.
- Decision: change the caption position while preserving the pack shot.
- Verification: replay the revised export at its intended display size and check whether the pack name can be read. This checks the file, not brand recall.
Demonstration: 2–4 seconds
- Source observation: the fictional refill action is shown, but the cut occurs before the action finishes.
- Analyzed timestamp / map observation: not analyzed; do not borrow an opening-frame prediction.
- Decision: test a cut that lets the action finish. Record any resulting duration change.
- Verification: inspect the completed motion in playback. If the decision depends on understanding, gather responses from intended viewers using a question that does not supply the answer.
End card: 4–6 seconds
- Source observation: the fictional brand name and next action appear together.
- Analyzed timestamp / map observation: not analyzed.
- Decision: keep provisionally; check the actual export for legibility, cropping and consistency with the offer.
- Verification: confirm that the linked destination matches the offer. No click or sales lift is inferred.
Download these three illustrative rows or start from the blank scene worksheet. When using a real report, replace “not analyzed” only with a matching source-frame timestamp and actual model observation. Keep the original asset and report reference beside the worksheet.
Inspect the original before the colors
Locate the product, brand and message on the original image. Then examine their relationship to the strongest predicted regions. If a face dominates the map, ask whether that person directs the story toward the product or competes with it. Attention is not automatically wasted when it lands away from the logo.
Do not treat red as a purchase signal. A salient element may be confusing, irrelevant or useful. The overlay supplies a spatial hypothesis, not comprehension, emotion or buying intent.
Read time carefully
Within a frame, compare regions. Across frames, check whether scale and normalization are comparable before comparing intensity. A small bright area in one image does not establish more total attention than a larger muted area in another.
Watch the sequence for the time available to read the claim. A still can look clear while disappearing too quickly in playback. Also check whether audio changes the meaning of the scene; a static map alone cannot measure sound-driven gaze.
Turn the observation into a limited edit
A proposed change might move the claim closer to the product, simplify competing text or give a key frame more readable time. Preserve the visual element that makes the ad distinctive. Do not replace the entire concept solely because one sampled frame looks weak.
The completion check should be visible: the condition is legible at normal playback, or the product remains identifiable through the transition. A higher diagnostic score is a separate result, not the definition of success.
End with the decision
Use the worksheet to rank edits by relevance to the objective, evidence and production effort. Where no map exists, use ordinary creative review and keep the evidence gap explicit. For interpretation basics, read attention heatmaps explained. Watch the demo to see how visual evidence fits into a wider review.
Evidence, authorship and next steps
This workflow is published by SaliencyLab, which sells creative review services and software. Oussama Nakhil’s background and the methodology and evidence limits explain who is responsible for the guidance and how to interpret diagnostics. This article is not an independent validation study.
For an inspectable research artifact, the opening-frame pilot includes 18 AI-assisted observations, source hashes and reproduction limits. Those observations do not validate the fictional rows above or prove a campaign outcome. TranSalNet’s cited paper concerns saliency prediction; it does not validate this worksheet as a measure of purchase intent.
Continue with the first-two-seconds review when the opening is unclear, or the brand cue placement guide when the issue is advertiser visibility. Keep one specific edit and one verification question in the final brief.