Evaluate
Do captions cover the killfeed? How to check
Check whether AI captions or a 9:16 crop hide your killfeed: map HUD regions, work out what a crop keeps and measure overlap frame by frame. A method only.
To check whether AI captions or crops cover your killfeed, freeze a frame from the finished clip and mark where the captions sit against the game’s HUD regions. No independent measurement of any tool’s overlap was found, so this page gives a repeatable method with the crop and overlap arithmetic shown.
Sections on this page (9)
Why this is a gaming-specific problem
Talking-head footage has nothing important at the edges of the frame, but a game HUD does. Killfeeds, minimaps, health bars and scoreboards sit in fixed places, and a default caption position designed for podcasts can land directly on top of them. Vertical crops add a second risk by removing those regions altogether.
Which HUD elements matter depends on the game, and on whether the viewer needs them to understand the clip. A killfeed is what proves a multi-kill, and a health bar is what makes a narrow escape tense. As reasoning, not a finding: expect a crop to hide more critical information than captions do, and expect a tool’s default caption position to ignore the HUD. You can check the first by seeing whether crop loss accounts for most hidden HUD frames, and the second by seeing whether placement stays the same across games with different HUDs.
Work out what the crop can keep before checking anything
Crop arithmetic tells you what a layout can possibly show, and you can do it from one screenshot. Illustration only, using invented HUD coordinates on a 1920 × 1080 source. Your game’s positions will differ, so measure your own.
| HUD element | Source x range | Source y range |
|---|---|---|
| Killfeed | 1500 to 1900 | 60 to 260 |
| Health bar | 700 to 1220 | 980 to 1060 |
| Minimap | 40 to 340 | 760 to 1060 |
Full-height centre crop to 9:16. The kept width is 1080 × 9/16 = 607.5 pixels, or 31.6% of the frame, running from about x = 656 to x = 1264. The health bar survives. The killfeed and minimap are gone before any caption appears.
Stacked layout with gameplay on the lower half (1080 × 960). Cropping the full source height to that shape keeps 1215 pixels of width, from x = 352.5 to x = 1567.5, scaled by 1080 / 1215 = 0.889. Only 67.5 of the killfeed’s 400 pixels of width survive, or about 17%. The minimap ends at x = 340 and is lost.
So a killfeed in the top-right corner of this example needs a layout that deliberately moves it. No caption setting can bring it back.
Measure overlap, frame by frame
Overlap is a simple rectangle sum, repeated at fixed intervals.
- Choose a game and map its HUD regions at the source resolution, using the coordinates above as a template.
- Run the tool with default caption settings, then note any setting that changes position.
- Export the clip and take one frame per second. With ffmpeg:
ffmpeg -i clip.mp4 -vf fps=1 frames/%03d.png - Convert each HUD region into the output frame using the scale and offset from the crop calculation.
- Mark the caption rectangle on each frame, and record whether each region is covered or cropped out.
- Report how often each HUD region is covered or cropped, as a share of frames.
Check default settings first, because most people never change them. A second pass can use the best position the tool allows.
Worked overlap (illustration only, invented inputs). Using the stacked layout above, the health bar maps to output x 309 to 771 and y 1831 to 1902. Suppose a caption box sits at x 140 to 940 and y 1780 to 1880.
- Overlap width: 771 - 309 = 462. Overlap height: 1880 - 1831 = 49.
- Overlap area: 462 × 49 = 22,638 square pixels.
- Health bar area: 462 × 71 = 32,802 square pixels.
- Share covered: 22,638 / 32,802 = 69%.
A reasonable rule of thumb is to count a region as covered when 25% or more of its area is under a caption, since text elements become unreadable well before they are fully hidden. If that happens in 12 of a clip’s 30 sampled frames, the covered share is 12 / 30 = 40%.
Where platform interface adds to it
A clip can be clear of captions and still be hidden by the app. The third-party Upload-Post safe zone checker, read on 8 November 2026, lists a bottom margin of 672 pixels for YouTube Shorts and for Instagram Reels on a 1080 × 1920 frame, and 660 for TikTok. On Shorts that puts anything below y = 1248 at risk. The checker says these figures come from the platforms’ ad guidance (Google Ads Help, Meta’s Ads Guide and TikTok Ads Manager Help), relayed by the checker, and that it found no official numbers for organic posts. As reasoning, organic interfaces may differ from ad layouts, so these are a starting guide and not a measured overlap for any tool. The checker itself says “Safe zones are a margin, not a pixel-exact map”, and adds that the interface varies by screen shape, app version, ad versus normal post and caption length.
In the example above, the health bar at y 1831 would sit inside that margin. The safe zone overlay page covers these regions for each platform.
Occlusion scoring sheet
The proof window is the few seconds in which a HUD element provides the evidence for the moment, such as the killfeed line for a multi-kill or the players-left count for a final circle. These thresholds are proposals. Fix them before you score.
| Check | Good | Acceptable | Fail |
|---|---|---|---|
| Proof element during the proof window | Fully visible | Partly covered but readable | Covered or cropped out |
| Frames with any critical HUD region covered | Under 10% | 10% to 30% | Over 30% |
| Crop loss | Every proof element kept or deliberately moved | Only decorative HUD lost | A proof element removed |
| Platform interface | Proof element and captions inside the margins | One close to a margin | Either inside a band |
| Caption position control | Set freely | Presets only | Fixed |
A failure in the first row fails the clip, however clean the other rows are.
Failure modes to log
Name the type, since each has a different fix.
- caption block over the killfeed or scoreboard
- word-by-word caption animations that pop over the centre of the action
- HUD elements cropped out by the 9:16 frame
- emoji or keyword pop-ups added over gameplay
- a caption that grows to three lines on a long sentence and reaches a HUD element that one line clears
- platform interface overlap, where the Shorts or TikTok buttons cover the same area
What can bias the measurement
HUD position is the main variable, and it changes the answer.
- HUD scale and custom layouts. Many games let players move or resize the HUD, so a result from one configuration does not apply to another. Record it.
- Clips with more speech. Talkative clips have more caption frames, so compare covered frames against frames with captions showing as well as against total frames.
- Streamer overlays. Alerts and borders can sit on top of the HUD and confuse the mapping.
- Output resolution. Use percentages if you compare tools that export at different sizes.
Sample and ties
Frames in a clip are not independent, so the clip is the unit that counts. Thirty frames from one clip are closer to one observation than to thirty. As a rule of thumb, use at least 20 clips per game for a meaningful share, and more if the proof window is rare. Treat two tools as tied if their covered shares differ by fewer than 10 percentage points, also a rule of thumb. A tie on share can still be split by the first row of the sheet, because hiding the proof once matters more than brushing a health bar often.
Which layer to fix
The fix depends on which layer is responsible.
- Captions cover the HUD, and position can be changed: set your position once and rerun, then keep the tool if a saved style holds.
- Captions cover the HUD, and position is fixed: rule it out for HUD-heavy games. The caption comparison shows which tools document position controls.
- The crop removes the proof element: captions are not your problem. Look for a layout that lets you place a HUD region, or crop by hand.
- Platform interface covers it: raise captions above the bottom band and preview in the app before posting.
For a quick check, export one clip from a free tier, pause on the busiest moment and ask whether a viewer can still see what proved the play. You can log the results in the clip evaluation log template, and other checks are on the guides hub.