A horizontal ad shows a folding phone stand, the hand that opens it, a $29 offer, and a final “See how it folds” prompt. Export the same edit as a centered 9:16 crop and the file can pass a basic delivery checklist—1080 × 1920, H.264, twenty seconds long—while the demonstration loses a hand, the price disappears, and the CTA begins off-screen.
That is the real resizing problem. The task is not to preserve every source pixel. It is to preserve the information that makes the ad persuasive: the product, the action that proves the claim, the complete offer, and the next step.
The most reliable workflow is shot-specific. First define what must survive each shot. Then use the smallest intervention that can preserve it: a fixed crop, a different fixed position, a tracked crop, a reflowed layout, another take, or a timing change. Finally, check the exported file in the exact placement configuration. A universal center crop is too blunt; a universal “redesign everything” rule wastes time.
This guide uses an original fictional 20-second stand ad to make those decisions visible. You can open the synchronized comparison demo, switch between 9:16 and 1:1, and compare a fixed center crop with a shot-aware version. The demo is a composition exercise, not a live campaign or performance test.
Start With the Placement, Not the Canvas Preset
“Make it vertical” is not a complete delivery brief. The editor also needs the placement, campaign or ad type, interface configuration, language direction, caption treatment, and the date on which the specification was checked. Those details affect both technical acceptance and the space the interface may cover.
Here are the three working targets used in the example, checked September 8, 2026:
| Selected use | Working export | Current qualification that matters to the edit |
|---|---|---|
| YouTube skippable in-stream, horizontal | 16:9 at 1920 × 1080 | Google lists 1920 × 1080 as the recommended horizontal Full HD size. Standard skippable in-stream accepts any length, while reservation campaigns have a separate 12-second-to-6-minute range. Viewers can skip after five seconds, so a reason to buy that appears only in the closing card may never enter the decision. See Google’s skippable in-stream specifications. |
| YouTube in-feed, square | 1:1 at 1080 × 1080 | Google recommends 1080 × 1080 for square video. The current in-feed specification accepts any length. Presentation also includes platform-controlled title, description, and thumbnail behavior, so those elements should not be confused with copy baked into the video. See Google’s in-feed video ad guidance and general video ad specifications. |
| TikTok Auction Non-Spark In-Feed, US/LTR, without Anchor | 9:16 at 1080 × 1920 | TikTok recommends vertical 9:16 at no less than 540 × 960 for this configuration; 1080 × 1920 is our production choice, not its minimum. Non-Spark video may run up to 10 minutes, must be no more than 500 MB, and must be at least 516 kbps. TikTok renders the ad caption in a uniform white style. The same June 2026 document says the safe zone depends on video dimensions, caption length, and additional formats, and it provides different files for standard versus Anchor configurations. See TikTok Auction In-Feed Ads. |
These are selected examples, not a permanent cross-platform spec sheet. Google says overlays, calls to action, and buttons can move according to format, campaign type, and screen. TikTok says its preview is not device-specific and may differ slightly from the live version. Even within YouTube, a vertical initial impression and a player compressed during interaction can expose different parts of the frame. The appropriate official reference is therefore a dated input to the edit, not a stencil you can apply forever. Google’s mobile video guidance documents that player-state qualification.
A useful delivery record fits on one line:
TikTok Auction Non-Spark In-Feed · 9:16 · US/LTR · no Anchor · final caption supplied · spec checked 2026-09-08 · preview capture pending
That line is more actionable than a file named final_vertical_v7.mp4.
Mark the Message Invariant for Every Shot
Before moving a crop window, make a shot manifest. A message invariant is the meaning that must remain understandable even when its coordinates change. It may be a product, but it can also be the relationship between two objects, the result of an action, a price plus its qualifier, or a CTA paired with the item it refers to.
The fictional stand ad uses six shots:
| Time | What must survive | What a careless crop can destroy |
|---|---|---|
| 0–3s | The stand and the claim that it folds flat | The centered product fits, but a wide headline can be clipped at both sides. |
| 3–7s | Phone, stand, and their contact point throughout the movement | A stationary portrait crop can lose the moving subject even though every individual frame is crop-able. |
| 7–10s | Both hands, the hinge, and the opening relationship at the same time | Tracking one hand preserves an object but removes the demonstration. |
| 10–12s | A hinge detail on the left side of the source | A center crop misses it; a simple fixed offset is enough. |
| 12–16s | Product, price, quantity, qualifier, and caption without collision | A crop can show the product or the offer, but not both; shrinking everything may make the qualifier unreadable. |
| 16–20s | Product, brand, and “See how it folds” | A technically present CTA is useless if it is clipped, covered, or detached from the product. |
This changes the review question. Instead of asking, “Is the subject inside the frame?” ask, “Can a viewer still understand what happened, what is offered, under what condition, and what to do next?”
The opening hero shows why separate layers need separate decisions. The centered stand fits the fixed crop, but the wide headline is clipped. Keep the product framing and reflow the headline; the product does not need a new shot or subject track. The left-detail shot shows a second low-cost answer: move the crop once for that shot. Tracking is not inherently better than a fixed position.
The remaining shots need more than one global setting. To know which kind of “more,” use the geometry before relying on taste.
Test Whether a Crop Can Work Before You Animate It
Take a 3840 × 2160 landscape source and preserve its full height while making a 9:16 version. The portrait window is:
2160 × 9 ÷ 16 = 1215 source pixels wide
At a given frame, suppose everything required lies between horizontal coordinates a and b. A crop with left edge x can contain that region only when:
b − 1215 ≤ x ≤ a
The crop must also remain inside the source, so 0 ≤ x ≤ 2625. Combining the two gives the feasible range:
L(t) = [b(t) − 1215, a(t)] ∩ [0, 2625]
You do not need to put this formula in every production brief. Its three outcomes, however, are operationally useful.
1. One fixed position works
In the 10–12 second detail shot, the required region spans source x-coordinates 450–1250. The feasible crop-left range is 35–450. Choosing x = 300 holds the detail for the entire shot. A center crop starts at 1312.5 and fails, but “center crop failed” does not mean “tracking required.” It means the crop was in the wrong place.
2. The region fits at every moment, but not in one stationary window
From 3–7 seconds, the required 700-pixel-wide phone/contact region moves from 1150–1850 to 1950–2650. At the start, feasible crop positions are 635–1150; at the end, they are 1435–1950. Those ranges do not overlap. A fixed portrait crop cannot preserve the whole motion, but a shot-local track can.
This is the situation automatic reframing tools are designed to help with. Apple says Smart Conform analyzes faces and other areas of visual interest, then allows position adjustments after reframing. Adobe’s April 2026 Auto Reframe documentation offers motion presets and warns that fast action, multiple points of interest, or rapid movement may need keyframe refinement. Those tools can provide an efficient first pass; the message invariant tells you whether they tracked the right thing. See Apple’s Smart Conform documentation and Adobe’s Auto Reframe instructions.
3. No crop position can preserve the required relationship
In the two-hand demonstration, the necessary span is 1760 source pixels. The portrait window is only 1215 pixels wide. Because 1760 > 1215, no pan, subject track, or extra keyframe can make both hands and the hinge fit at that scale.
The remedy has to change something other than x-position: use a wider or open-gate take, scale the demonstration into a panel, split the layout, change the camera angle, or replace the shot. More sophisticated tracking cannot solve a window-width problem.
Square changes the answer again. A full-height 1:1 crop from the same source is 2160 pixels wide, so the 1760-pixel two-hand relation can fit. The moving region’s complete 1500-pixel union can fit too. Recomposition is not a moral virtue; it is work you add when the target frame actually requires it.
Recompose the Offer Instead of Shrinking the Whole Ad
The offer shot exposes a different failure. In the horizontal master, the product occupies x = 520–1480 and the offer group occupies x = 2620–3600. Their combined span is 3080 pixels. Neither the 2160-pixel square window nor the 1215-pixel portrait window can contain both at the original scale.
A fixed crop therefore creates a false choice: keep the product and lose the offer, or keep the offer and lose the object it prices. The shot-aware versions preserve the meaning by changing the layout. The product moves above; the offer becomes a separately sized card below; price, quantity, and qualifier remain grouped; caption space is reserved outside that group.
Trying to “fit” the whole horizontal frame can avoid literal cropping, but it may create a readability failure. Scaling 3840 pixels to 1080 gives a factor of 0.28125. A 64-pixel-high source glyph becomes 18 pixels in the output. If that square output is displayed 360 CSS pixels wide, the glyph appears about 6 CSS pixels high. A 40-pixel source qualifier falls to 11.25 output pixels, or roughly 3.75 CSS pixels at that display size.
Those numbers are not a universal minimum-font rule. They show why “all pixels retained” and “offer retained” are different claims. Rebuild critical graphic copy at the target canvas size and inspect it at a plausible rendered size rather than admiring a 100% zoom preview on a large monitor.
This is also where source-file quality becomes a production constraint. A layered project lets you move the price, rewrap the qualifier, and reserve caption space. In a flattened MP4, baked-in copy is part of the image. You can crop, mask, cover, or replace those pixels, but you cannot reflow them as editable text. When the source is flattened, ask for a clean master, graphic files, fonts, brand rules, and unused handles—or choose a different shot. Detecting old edit points does not reconstruct the missing layers.
A practical source package for resizing should include:
- the highest useful resolution or open-gate media available;
- clean plates or footage without baked-in supers;
- editable text, offer, logo, and CTA layers;
- captions as a separate timed file or editable track;
- several seconds of handles around shots likely to be replaced;
- the exact offer wording, quantity, qualifiers, and legal copy;
- the original music and dialogue stems when edit timing may change.
Measure How Long the Complete Offer Survives
Spatial checks catch only half the problem. An offer can be fully inside the frame yet never be fully available for long enough because its components enter at different times or another caption covers them.
In the designed timeline below, the price is visible from 12.0–16.0 seconds. The condition “for one stand” appears from 12.6–15.8. An editor-added caption covers the offer area from 14.8–16.0. The unobscured interval containing both price and condition is therefore:
[12.0, 16.0) ∩ [12.6, 15.8) ∩ [12.0, 14.8) = [12.6, 14.8)
That is 2.2 seconds, or 66 frames at 30 fps—not the four seconds suggested by the price layer alone.
Moving the caption outside the offer group raises the clean overlap to 3.2 seconds. Making the condition share both the price’s start and end times—12.0–16.0 seconds—raises it to four seconds. Neither change requires a longer ad. They require treating offer timing as a system rather than checking each layer independently.
Do this calculation with the final caption state, not a silent export that will receive captions later. Check the exact words too. “$29” and “$29 for one stand” are not interchangeable messages, and a line wrap that hides “for one” can change the offer even when the price remains large.
Verify Four Different Things: Edit, Export, Preview, Served Impression
A clean editing timeline does not prove a clean file. A clean file does not prove a clean placement preview. A preview does not prove every served impression will match it. Keep the evidence layers separate:
- Editor view: Check shot framing, keyframes, graphic layers, caption tracks, and line breaks. Use overscan or an equivalent view to see what source area remains available outside the target frame.
- Exported file: Open the actual output—not only the timeline—and inspect start, end, transitions, tracking, offer timing, captions, audio sync, dimensions, frame rate, and duration. Scrub the intervals where required regions touch an edge.
- Platform preview: Load the final asset into the selected ad configuration with the final caption, CTA, language direction, and optional formats. Capture each relevant device or surface view with the date and configuration visible in the record.
- Served impression: When the campaign is live, treat a real device/surface observation as a separate piece of evidence. Do not relabel a preview screenshot as a served ad.
Regenerate your review captures after changing the asset or copy, and check the device views offered by your selected ad configuration. TikTok’s current specification explicitly warns that preview and live versions may differ slightly because the preview is not device-specific. See TikTok’s Safe Zone section.
The review record should name what was actually checked:
| Check | Record |
|---|---|
| Export | Filename, checksum or version ID, dimensions, codec, frame rate, duration, caption state |
| Composition | Shot/time range, invariant, method used, result, reviewer |
| Platform preview | Platform, placement, campaign/ad type, CTA, caption copy, LTR/RTL, optional format such as Anchor, device view, capture date |
| Served observation, when available | Date, device, app/browser version if known, surface, screenshot or screen recording, limitations |
This is deliberately more specific than “safe-zone pass.” A safe-zone reference can identify likely interface conflicts for one documented configuration. It cannot prove that the demonstration still makes sense, that the qualifier remains readable, or that a moving subject stays in frame.
A Minimal-Intervention Resizing Workflow
The complete process can be run as a sequence of decisions rather than a blanket redesign:
- Choose the exact placements. Record ratio, recommended and accepted dimensions, duration rules, interface options, caption behavior, and specification check date.
- Duplicate the source timeline for each target. Do not overwrite the horizontal master or assume one adaptive sequence will produce every deliverable correctly.
- Write the shot manifest. Mark the product, action, relationship, result, offer components, branding, and CTA that must survive—plus the time range in which each is needed.
- Generate a fixed center crop as a diagnostic baseline. It is useful because its failures are easy to see, not because it is the default final answer.
- For each failed shot, test interventions in order: shot-local fixed offset; tracked crop; scaled panel or split layout; reflowed graphics; alternative take; edit-timing change. Stop when the message is preserved cleanly.
- Rebuild critical text at target size. Keep price and qualifiers together, reserve caption/interface space, and inspect at a plausible display size.
- Export and inspect the real files. Verify dimensions, frame rate, duration, transitions, motion, captions, and the complete-offer interval.
- Preview the exact ad configuration. Use current first-party guidance, save dated captures, and regenerate previews after asset or copy changes.
- Log what remains unobserved. A platform preview is not a served impression, and neither is a conversion test. Resizing alone should not be credited with a performance change without comparable campaign evidence.
The principle is simple: preserve the reason to buy with the least intervention that works, then verify the result where it will appear. Sometimes that is a center crop. Sometimes it is one fixed offset. Sometimes the subject must be tracked. And sometimes the frame is mathematically too narrow, so the honest answer is to recompose or replace the shot.
Correct pixels are only the entry ticket. A successful resize keeps the product recognizable, the proof legible, the offer complete, and the next action available at the moment the viewer needs them.
Sources
- About video ad specs — Google Ads Help; checked September 8, 2026.
- Skippable in-stream ads — Google Ads Help; checked September 8, 2026.
- In-feed video ads — Google Ads Help; checked September 8, 2026.
- Use square and vertical video to engage mobile customers — Google Ads Help; checked September 8, 2026.
- TikTok Auction In-Feed Ads — TikTok for Business; last updated June 2026; checked September 8, 2026.
- Adjust framing with Smart Conform in Final Cut Pro for Mac — Apple Support; checked September 8, 2026.
- Add Auto Reframe effect to sequences — Adobe Help; last updated April 15, 2026; checked September 8, 2026.
