AI disclosure: Drafted with AI assistance for Apiary. The café, media, and screenshots described below are hypothetical. The included geometry is synthetic; it is not proof of the way a particular browser rendered a live page.
A café commissions a wide shot of somebody finishing a cup of espresso. The steam rises on one side of the frame and the shot glass pours from the other. On a desktop page, both elements make the scene. On a phone, a tall hero shows a patch of countertop between them. Nothing is wrong with the source video. The original composition and the phone's visible window just don't agree.
The quick fix might be to move the focal point. That is a good solution when both important details fit inside the phone's window and are merely offset. If the details are farther apart than the entire visible width, no object-position setting can recover them together. Do five minutes of geometry before asking a designer to find a magical alignment.
Define the story the first view needs to tell
A lot of media QA checks only whether faces are cut off. The critical detail might be a product label, the contact point between two pieces, the first frame that explains a gesture, or the distance between a face and a tool. If a video is an ambient surface and the content on top already explains everything, its crop may tolerate more loss. If the footage demonstrates a motion that the viewer needs to understand, pretending it is just decorative creates a different access and usability problem.
The fictional café has three categories of anchor:
- Required: the spout when the stream begins, the stream while the pour happens, and the rim where the stream lands.
- Helpful: the person's hand, especially if it makes the direction of movement more legible.
- Optional: the plant, menu board, and reflection that give the desktop shot atmosphere.
Mark each anchor as a box in source coordinates. Mark the time window when it matters. A single still at the midpoint won't expose the spout leaving the right edge early or the finished cup sliding to the left as the clip resets. Pick one starting frame, one during the active step, one near the end, and one at the loop point if the hero loops. Track more frames if the movement or camera changes sharply. The anchor plan is an editorial note with measurements, not a claim to have solved automated object tracking.
Account for overlays too. If the title sits over the required anchor on portrait, the anchor is effectively unavailable even if its coordinates fall inside the video frame. Set a safe area that includes the vertical space the title, primary link, captions if applicable, and playback controls need. Check it with actual type and dynamic browser height, not an illustration stripped of copy.
Turn the shot description into a contract somebody can falsify
Write acceptance statements in terms of what a visitor can understand, not where the editor expected a subject to end up. “Cup near bottom of hero” says little about the requirement. It could pass when the rim is hidden, when the stream is out of frame, or when somebody can only see the bottom third of the cup. “From the moment the stream begins until the cup is set down, the frame shows where liquid leaves the spout and reaches the opening” has a scope, a beginning, and two pieces of information that can actually be checked.
Next decide whether those objects need to coexist in the same image or whether their sequence is enough. In a quick pour shot that relies on obvious continuity, simultaneous visibility can be helpful. In a montage made of authorized source footage, adjacent shots can support the story if the edit makes the action clear and does not imply continuity across two unrelated demonstrations. An editor can often preserve meaning with a simpler cut, but a higher-resolution export will not supply a missing relationship. State whether the decision accepts an alternate sequence before an animator or developer starts moving layers around to solve impossible geometry.
Narrow the test by asking which failures matter to someone new to the page. The café owner knows that the bright metal shape at the edge is a spout even when the shape is mostly outside the frame. A visitor does not share that mental model. They may interpret a tan streak as glare or think the beverage is already sitting on the counter rather than being made. Write the review prompt as a question for a fresh viewer. “Where did the stream come from?” gives more signal about whether the intended story survived than “does this look good on a phone?” It also separates photographic style from the information the asset was chosen to communicate.
For an implementation handoff, assign one identifier to each anchor and explain whether it is judged as a whole box or a meaningful sub-feature. An entire person's elbow being in frame is not necessary if the small motion of their fingers provides the relevant information. A product demonstration may need every letter legible, which is a stricter requirement than keeping the product reasonably centered. List decorative elements separately. This gives the reviewer permission to lose something beautiful when the alternative is clipping something necessary. A portrait crop is an editorial decision even if the ultimate patch is one line of code.
Work out the visible source window
CSS Images defines object-fit. With cover, the content keeps its aspect ratio and is sized large enough to cover the box. If the rendered media ends up wider or taller than the box, part of it is cropped. The object-position definition controls the alignment of the object in that box. A percentage is about aligning positions in the object and the box, not about placing an arbitrary source pixel directly at the visible center.
Consider a 1,920 by 1,080 source in a 390 by 560 CSS-pixel hero. To cover the 560-pixel height, the uniform scale is approximately 0.519. The visible source width is then only about 752 source pixels. The entire 1,080-pixel source height remains, and more than half of the source width is outside the viewport. At 50% horizontal position, the crop occupies approximately source x=584 through x=1,336. CSS pixel density may affect download choices, but multiplying both sides of the destination by the same device pixel ratio doesn't change the composition.
For the café, suppose the required spout sits at x=160–240 and the cup rim sits at x=1,470–1,640. Their combined extent is 1,480 pixels, almost twice the 752-pixel window. No centering or side alignment fits them together at this ratio. A desktop shot with both objects can still be excellent; it just needs a different portrait edit or a shallower portrait container. The test should make that clear before someone spends an afternoon dragging a slider.
Here is a fixture for normalized percentage alignment and ordinary cover geometry. It ignores transforms, borders, captions, and any source rotation processing. Measure the actual content box and use the effective source dimensions when adapting it.
from math import isclose
def cover_window(source, box, position=(0.5, 0.5)):
sw, sh = source
bw, bh = box
if min(sw, sh, bw, bh) <= 0:
raise ValueError("dimensions must be positive")
x, y = position
if not 0 <= x <= 1 or not 0 <= y <= 1:
raise ValueError("fixture only supports 0–100 percent alignment")
scale = max(bw / sw, bh / sh)
vw, vh = min(sw, bw / scale), min(sh, bh / scale)
left = (sw - vw) * x
top = (sh - vh) * y
return left, top, left + vw, top + vh
left, top, right, bottom = cover_window((1920, 1080), (390, 560))
assert isclose(right - left, 752.1428571428571)
assert isclose(top, 0.0) and isclose(bottom, 1080.0)
assert (1640 - 160) > (right - left) # cannot fit both anchors
# New portrait master was composed vertically at the source.
crop = cover_window((1080, 1920), (390, 560))
nearby_spout = (180, 390, 310, 610)
nearby_rim = (610, 1030, 830, 1270)
def contains(window, rectangle):
return (window[0] <= rectangle[0] <= rectangle[2] <= window[2]
and window[1] <= rectangle[1] <= rectangle[3] <= window[3])
assert contains(crop, nearby_spout) and contains(crop, nearby_rim)
The second test covers one sample frame. It says nothing about the first frame, the next second, text legibility, a narrower box, or whether the HTML actually chooses that master. Those require separate checks. The value of the geometry is to establish what is possible and to give the reviewer concrete regions to inspect.
Check the feasible interval, not just its most flattering frame
A single window is good for spotting an impossible case, but you can be more precise about a possible one. At a given time, let the horizontal extents of all required anchors run from A to B. Let the window be W wide in source pixels. Its left edge must be no farther right than A and no farther left than B minus W. Clip that interval to the actual permitted range from zero to source width minus W. A fixed position works across several checked frames only where all their permitted intervals intersect. If the intersection is empty, adding a decimal to a single setting won't help.
In the 1,920-pixel-wide, 752-pixel-window example, assume one frame's anchors span x=900 to 1,450. Its left edge could be anywhere from roughly 698 to 900. A later frame's anchors could run from x=1,100 to 1,690, permitting an edge from about 938 to 1,100. Both frames fit individually, but no fixed left edge satisfies both. That's a distinct failure mode from a single frame whose relevant region is too wide. Label it correctly: the existing crop can frame each moment but has no shared static alignment. A recut, a different camera composition, or an alternate layout that widens the visible region might solve it more elegantly than jumpy object positioning.
If that later extent instead ended at x=1,610, the two frames provide a common edge range of about 858 through 900. The remaining amount is not an automatic pass. If you require even a modest side buffer on each frame, apply the buffer symmetrically to the anchor bounds before computing the intersection. A buffer is not about multiplying every coordinate by a number someone saw in a design rubric. Its purpose is to reserve real space for actual movement, browser geometry, or the way a subject looks when its edges have no room to breathe. Record the difference between a mathematical overlap and one a reviewer is willing to use.
Perform the same check on the vertical axis if the object overflows there. When it does not overflow, treat vertical alignment as having no crop to move under the simple model; the object has no unused vertical region to reveal. A zero overflow axis is not a reason to divide by zero in a debug calculation. When an axis is cropped, a suggested left or top pixel offset can be converted back to normalized position using the corresponding overflow distance. Prove the browser actually applies that position in the layout, particularly if the media wrapper uses an additional crop or the implementation has independent transforms.
Work in the right reference frame
Several coordinate systems coexist in a media test. A frame annotated in a media editor may be measured in encoded pixels. The browser may present an effective intrinsic orientation. The surrounding DOM sizes are usually CSS-pixel dimensions. A review screenshot may capture the physical screen at another density. Mixing these numbers leads to perfectly calculated results for the wrong box. Label the source and review units explicitly and decide which conversion is real. A 780-pixel screenshot of a CSS 390-pixel width does not make the destination composition wider.
Check offsets from the correct origin. An outline drawn from the entire viewport is not necessarily anchored at the upper-left of the content box. A sticky banner, margin, or inset can move the visual origin. Use the box where the media is sized and the actual effective display rectangle. Be equally careful with screenshot measurements: a source asset x=900 is not a printed x=900 on the screenshot unless the mapping really makes it so. A reliable annotation can say “x=612 in source, maps to 87 CSS pixels from the media box left in the recorded state.” Whoever checks that claim knows which step to reproduce.
A quick diagnostic is to use a harmless preview asset with a few recognizable labeled markers. It can confirm whether the tested transform and cropping boundaries do what you expected. That preview is not proof that the real content has the same composition or selection outcome. Include both a marker fixture test and actual editorial evidence when the calculation is important. Once the cause is understood, remove the diagnostic asset from the intended release configuration and rerun the relevant real state. Nobody should accidentally ship the measurement grid as the only image that passed.
Pick the smallest intervention that really fits
If all required boxes fit inside the source window with useful breathing room, a per-breakpoint object-position can be adequate. For example, if both anchors occupy x=930–1,530 in our source, a right-shifted crop may be enough. Verify the bounds across the whole relevant segment and leave room for modest container size changes. “Fit” with one pixel to spare turns into clipping as the address bar changes the visible viewport or copy wraps to a second line.
If those anchors fit only by shrinking the height of the displayed box, balance that change against what the hero should emphasize. A shallower rectangular media area preserves more of a wide source on a narrow device, but it uses a different composition from a full-height hero. That's fine if the product wants the scene to sit beside or above the primary text instead of behind it. Don't fill the unused area with an enlarged blur and call the subject recovered; blur can be an intentional treatment, but lost detail remains lost.
If required anchors cannot fit in a suitable container, make another crop or a vertical edit from suitable source material. The café's filmmaker can pick an alternate angle with the spout above the cup and compose it as a vertical stack. Another approach is to change the sequence: show the hand initiating the flow, then cut to the rim receiving it. Done well, the meaning survives without requiring both anchors in one shot. Done poorly, the cut can imply a product action occurred when the frame shows only unrelated motions. Verify continuity, timing, and labels against the actual recording.
When neither the media nor the layout can change, show the entire asset with contain or put it in a player with a natural ratio. This may produce unused area in the layout, but it avoids hiding critical information. Sometimes the highest quality result is a simpler presentation instead of forcing a promotional frame shape onto evidence footage.
A second case: preserve a handoff when the frame itself has enough room
Imagine an independent page for a neighborhood print room. Its hero shows a volunteer placing a small preprinted card into a freestanding desktop rack. Unlike the wide café shot, this source is already close to the desired portrait aspect. A particular state keeps the card and rack opening inside the source window on every checked frame. The review initially passes because the team inspects the asset in isolation. On the page, a pale translucent callout enters as the card crosses the opening. It lands in exactly the region that connects the motion. The frame passed its crop test and failed its meaning test.
A useful follow-up does not have to redesign every asset. Store the critical box for the handoff during its short time window, then collect a second observation with the component's real copy and callout shown. If there is room above the intended visual action, move the callout there and retest readability, interaction and the next section. If there isn't, put the callout on a separate surface below the active scene. Altering the media position may work on one geometry and shove the same action into a bottom control area on another. Fix the relationship between the editorial safe area and actual page furniture rather than turning every lost detail into an arbitrary focal position.
Add representative content differences to the replay. A two-line callout might remain above the action at the default user setting, while a translated or longer authorized variant wraps into three lines. Those aren't always equivalent states. Test the content that exists and name configurations not checked. If the layout supports real changes in text scaling or controls, include appropriately selected cases in the same matrix. The solution worth shipping is the one that keeps the moment intelligible without making the controls or title worse. Any other view should remain labeled as unverified, not rolled into one passing screenshot.
What can you legitimately conclude after the fix? “The insert and opening remain visible in the named playback range, with normal and long-copy variants reviewed on the listed phone-sized states.” That's a specific earned claim. You cannot conclude that every copy, browser, video position, or language was reviewed. A scoped result is not a weakness: it tells the next editor exactly which test to repeat when an approved copy change alters the overlay dimensions.
Treat posters as their own crop decisions
The first thing a phone visitor sees could be a poster, not a decoded moment of video. The poster should work before autoplay, when autoplay is blocked, and when motion is intentionally off. It does not have to be an arbitrary frame from the same recording, but it should be an honest preview of the depicted subject rather than a shot implying missing footage.
For still imagery, the HTML art direction model describes how picture and media-dependent source elements can deliver alternate compositions. Use srcset density or size variants for resolution choices within the same composition. An image with more pixels won't bring back the missing spout. A cropped desktop composition and a deliberately composed portrait alternative are different sources even when both originate in the same session.
<picture>
<source media="(max-width: 600px)"
srcset="/media/cafe-pour-portrait.jpg">
<img src="/media/cafe-pour-wide.jpg"
alt="Espresso pouring into a clear glass">
</picture>
This is an illustrative still, not a video element. The alternate compositions convey the same subject, so the shared text alternative is consistent. W3C's image decision tree helps decide what the alternative should convey and when an image is decorative. If desktop and mobile variants tell meaningfully different stories, address that at the editorial layer instead of forcing one description to cover inconsistent content. On a hero where text provides all relevant information and the image is truly decorative, a different alternative-text decision may be appropriate.
A responsive still mechanism does not automatically select the matching video. If you separately load portrait and landscape clips, decide when selection happens, how orientation changes work, and what loading behavior you expect. A switch from portrait to landscape that downloads two large files without a play request could undo performance work. Measure network requests on a cold load and after one intentional rotation. Keep controls usable during transitions; a framing fix shouldn't restart a user-paused movie or cover navigation.
Give the no-motion state its own success criteria
For the viewer who hasn't played the clip, the goal may be to know what the depicted activity is, whether play is an option, and which primary page action is available. A poster can carry that framing without claiming to demonstrate the motion. If the application shows a poster that emphasizes the final cup, someone who needs to understand how the pour is performed still needs the actual video or a suitable text description. Review the two states separately instead of letting a passing poster certify an unseen video.
An intentionally paused video may show the exact moment somebody stopped it rather than snapping to a poster. A page may also have a design that swaps the decorative background for a still in a reduced-motion configuration. These paths are not interchangeable. List the visible asset, whether media plays, and which controls remain operable in every reviewed row. An observed state can be as plain as “page opened paused with still shown and play control present.” It shouldn't claim to verify all reduced-motion settings if the team only clicked a stop button in one browser. Conversely, observing no network request for the large source on one cold load is not evidence of the same behavior after play or orientation changes.
Check visual identity between the poster and first decoded frame. Differences can be intentional, but a transition should not confuse the user about which clip they chose. On a hero that has two editorial variants, inspect whether portrait and landscape posters are paired with the corresponding source selection and useful content. If the still's crop obscures the name printed on an illustrative image, serving a cover frame from the video might not preserve that particular detail either. Ask what content the state needs before searching for the nicest first frame.
The moderate response to a bad fallback state is often a robust still and clear text, not more complex runtime conditions. If a page opens with no decoded media, a simple rendered still can tell the visitor what the hero is about and allow them into the main content. If the media is essential instruction, include an appropriate accessible equivalent rather than treating autoplay as the only dependable explanation. If the media is decorative, the fallback can be still more modest. Account for behavior without treating aesthetic preference as a verdict about access.
Build a crop receipt the next person can repeat
A picture in a design file is useful as intent. It is weak evidence of what shipped if it lacks the actual box, artwork revision, and relevant frame time. A small receipt for the café might record:
| Field | Example value |
|---|---|
| Asset revision | portrait-pour-r4 |
| CSS content box | 390 × 560 |
| State | motion active, text visible |
| Reviewed frame times | 0.0, 1.3, 3.2, loop seam |
| Critical anchor | spout, stream, rim |
| Result | both visible except rim overlaps caption at 3.2 |
| Follow-up | shift caption area or change frame composition; rerun portrait |
This is not screenshot theater. The failing case needs to make sense when it appears on someone else's device. Include whether the browser toolbar was expanded, whether the reviewer zoomed, and whether the site displayed a banner or caption that changed the visible safe area. If the site supports translated copy, consider a representative longer label. If the layout must adapt with magnification, test that case rather than using the phone manufacturer name as shorthand for accessibility.
Check the first visit with no cached media. Then check the configured reduced-motion or manual-pause state. Finally, check natural scrolling from the hero into the section immediately after it, because some implementations trap a visitor in an oversized media stage or obscure the next interactive element on the way out. Record as separate observations. One good screen grab at the peak of the pour doesn't show that the first frame and next section behave.
Review useful edge conditions, not a pile of phone names
A table of 52 passing screenshots is hard to interpret if 49 of them are effectively the same layout. Choose representative geometry from conditions that change the calculation or content. Put one review just below and just above the breakpoint that changes art direction. Put another at the narrow end of the supported content widths, and one where a short viewport makes content and controls compete. Explicitly include the variant that changes overlay or script behavior when it is part of the design. Two tests with different device labels but the same box, asset and copy are narrow corroboration, not evidence for two entirely different interpolated ranges.
When a browser window is resized, observe which combinations are actually rendered, not what a responsive image query would have done under an assumption. The visible media and control selection can change through the player's own logic, not just CSS. Keep the route, state, current asset identifiers and source frame of the failure together. A private receipt should omit irrelevant session details and personal information. Preserve the minimum geometry and evidence needed to reproduce the issue on the next compilation.
Watch for interactions introduced by testing. A developer toolbar might shrink the desktop viewport or force an override. A screen capture starts after the moment the user actually needs to see. A controlled check from a signed-out or fresh state can detect initialization differences, but make a separate note about what distinguishes it from the everyday return path. If layout shifts when a font finishes loading or copy arrives, record the first useful display and settled state rather than choosing one flattering capture. That distinction reveals whether the user saw an understandable short sequence or two disconnected arrangements of the same beautiful image.
Find out what moved when a regression appears
A portrait crop can regress without anybody touching the media file. A marketing editor adds a line under the headline, the grid grows taller, and the larger hero box crops more of the same wide source. Or a CSS change applies a new orientation rule and chooses landscape against the intended breakpoint. If all you know is that the asset ID stayed the same, both of those errors can look impossible.
Capture the effective source dimensions and content-box size in preview. Compare the latest recording with the previous one at the same route and browser geometry. A changed source with an unchanged box points your first investigation toward asset selection or encoding. An unchanged source with a different box points toward layout. If both stay the same but content has moved, inspect position, transforms, and any additional layers cropping the result. A screenshot is more useful when it records which of those facts it was showing.
An editor can use the same technique before a template is copied to a second location. The new location might have a longer title or a logo that extends farther across the hero. Keep the source well framed, but rerun the anchor test rather than equating “same template” with “same visible window.” If both copy and framing need room, consider separating the type from the moving surface instead of making one always fight the other.
Don't penalize a short device for being short. A full-height section that fits on a tall review phone may occupy the entire usable view in a mobile browser with chrome and a consent banner present. If the visitor must see the primary action without scrolling, measure that requirement separately from whether the media itself crops nicely. These are different acceptance tests, and a well-composed still can pass the second while the page fails the first.
Make a cropping regression actionable after the review meeting
Create an issue with a class and one primary cause hypothesis, clearly marked as provisional. If the source window cannot contain the required rectangle, assign it to editorial framing or layout before chasing asset encoding. If the geometry passes but the subject disappears behind a control, give the owner the actual obscuring element and state. If the wrong variant loaded, record which decision path chose it. These are separate tasks, even though all might appear in the same symptom report: “I couldn't see where it poured.”
A reviewer shouldn't be expected to find the cause with a single screenshot. Establish a reliable reproduction, then compare two supportable facts before making the patch. For instance, confirm that the same source frame is visible and that the content box changed after the headline became longer. That does not alone prove the copied text caused the regression, but it identifies the appropriate difference to isolate in preview. Retest with a minimal reversible change and one control geometry whose behavior should remain unchanged. Record both outcomes. A fix that restores the phone hero and cuts the desktop scene should not be silently accepted as a passing release.
Make the proof small enough to update on future projects. Catalog the anchors only for the scenes that tell a necessary story. A hint like “ambient hand movement; no required relationship” can be enough for several decorative shots. For critical relationships, preserve a handful of checked frames, measurements, and the named overlay states. This makes QA part of editorial production rather than paperwork that grows with every photo in the folder. The amount of proof should rise with ambiguity, not with the storage size of the movie.
Decide what to retain after the material changes
When a new asset is staged, avoid carrying old coordinates forward as if a media filename preserved composition. Two revisions might be exported at the same resolution and display length while the colorist, editor, or asset tool chooses a slightly different crop or performs image stabilization. Treat a newly encoded revision as a new evidence object, even when its design ID remains unchanged. Transfer the labels that describe what needs to survive, then measure their positions again where the new revision calls for it. This preserves the editorial work instead of merely preserving some old numbers.
Some checkpoints are independent of the exact asset. The CTA being reachable without covering the primary title is a page acceptance condition that can remain valid across a new shoot. A specific rim reaching pixel y=1,270 at a named frame applies only to one revision and effective geometry. Keeping these scopes separate makes the next release cheaper: reviewers rerun the checks whose inputs changed and do not throw away all prior reasoning. It also prevents an inherited screenshot from standing in for a comparison it never made.
If a crop is supposed to illustrate continuity across a cut, save neighboring evidence as a pair. The viewer may see one shot establish the toy, place or object, and the next zoom in on a movement. A single frame after the cut might not show why the change was clear. Conversely a review cut timed to music may make a transition feel obvious on a large screen while it looks discontinuous on a smaller one without audio. Include an ordinary sound-off review if the interaction expects that use. Label that as a narrative test, not a measurement of lip sync, audio access or playback quality.
One cheap check is to ask a colleague unfamiliar with the open review to watch once and state what the clip shows. Don't prime them with the desired answer or allow the developer tools to obscure the controls. Ask what moment was hardest to interpret. Combine that observation with the most relevant geometry. If the person describes the intended action even though a decorative object fell out of frame, consider preserving the simpler version rather than reframing to satisfy a checklist nobody actually asked for.
Know what the eye can catch that the coordinate test cannot
All anchor boxes can fit and still make a physically unconvincing shot. In the café clip, the stream could jump sideways after a cut, or two differently lit frames could give the impression that the cup changes shape. Watch the shot at ordinary speed, then scrub the seam where the motion repeats. A still-frame crop proof can show that each object is in view. It cannot prove movement continuity, realistic contact, or smooth pacing. Do not call a zoom on a single frame proof of character or object animation; report the asset as the still or camera move it is.
If multiple people work on the same scene, give each review a scoped task. One reviewer checks visual continuity. One checks phone layout and input. One checks alt text or video accessibility. A person can fill multiple roles, but a design pass shouldn't accidentally stand in for accessibility and behavior verification. When you spot a fail, save the smallest witness that will help the next person reproduce it: box, timecode, source revision, and an annotated rectangle are often enough.
The final decision is editorial, not mathematical. A shot needs to keep the part the user came to see. Measure the window and prove what fits. If moving the focal point retains the story, make that simple adjustment. If the story doesn't fit the shape, recut the scene instead of polishing the lost middle. That is how a high-definition image becomes an understandable first screen rather than an expensive background.