Evaluate
How to check facecam and gameplay layouts in AI clips
Check how an AI clipper handles your facecam and gameplay layout: a scene test card, pass and fail anchors, a worked pass-rate sum and sample-size limits.
To check an AI clipper’s facecam and gameplay layout, run it on a handful of your own scenes and count how many vertical clips have both the face and the game correctly framed. Vendors describe this feature but the sources read on 8 November 2026 give no failure rates, so the count is yours to make. This page gives a scene test card, pass and fail anchors and the arithmetic for a small sample.
Sections on this page (8)
What vendors and platforms say
Several sources describe automatic layouts, and none of the ones read gives an accuracy figure as of 8 November 2026. Twitch’s clip help says portrait clips are “automatically cropped to keep your cam and gameplay in frame”, with a Stacked layout of two boxes and a Full layout of one box. It also describes the fallback: “If your cam or gameplay can’t be confidently identified, you’ll see a centered crop as a starting point.” Vizard’s Twitch page says its AI “ensures that your gameplay or facecam remains at the heart of the action” for YouTube Shorts. Both are product descriptions, not measurements. Twitch’s fallback is still useful to know, because as reasoning, a centred crop on a corner facecam will miss the camera.
Why layouts break on streams
As reasoning, not a finding, layout accuracy probably depends more on the type of scene than on a tool’s overall ability, so a single percentage hides where it breaks. You would see this if a tool’s pass rate differs sharply between scene types. Streams are difficult for reasons that can be stated without a measurement:
- cams change shape and position between scenes
- animated frames and overlays sit around the camera
- some games have portraits that look like faces
- a streamer may switch to full-screen gameplay or full-screen camera partway through
Build a layout test card from your own scenes
A test card is a short list of scene types that together cover how you actually stream. Pull eight to ten suggested clips for each scene you use:
| Scene | What it checks |
|---|---|
| Cam in your usual corner, gameplay behind | The easy case and a baseline |
| Cam in a different corner or a different size | Whether the tool searches or assumes |
| Full-screen gameplay with no cam | Whether it invents a face region |
| Full-screen cam, such as a talk segment | Whether gameplay is forced in anyway |
| A scene change inside the clip | Whether framing follows the change |
If your game has character portraits in the HUD, include a stretch with them on screen, and note it. As reasoning, that is the likeliest route to a wrong region.
To run it, submit a stretch holding those scenes to each tool with default settings and export the vertical clips before making any layout edits. Mark every clip against the anchors below. Time how long it takes to fix one failed clip per category, because the fix cost often decides whether a tool is usable.
Pass and failure anchors
Each clip gets one primary result: pass or the most serious failure category. The thresholds are proposals. Fix them before you score.
| Category | Pass | Light fix | Failure |
|---|---|---|---|
| Missed cam | Cam region found | Found late, within the first second | No cam region although one is visible |
| Wrong region | Cam crop contains the face and nothing else of note | Includes a little of the overlay border | An alert, overlay or in-game portrait used as the cam |
| Face cut off | Forehead to chin inside the frame throughout | Slight cut at the hairline for under a second | Eyes or mouth outside the frame for over a second |
| Unstable framing | Crop holds still unless the scene changes | One visible correction | Repeated jumps, or drift of more than about 5% of frame width with no scene change |
| Gameplay crop error | Gameplay crop holds the action | Action near the edge but readable | Crop centres on empty space or a menu |
Check one thing the table cannot show: whether clips from the same scene use the same crop. A set that all share a layout looks intentional, while one that varies clip to clip means extra fixing even if every clip passes alone.
What each failure looks like in practice
Each category points to a different cause, so log examples as well as counts.
- Missed cam: the vertical export shows gameplay only while the streamer is plainly on screen.
- Wrong region: the top box shows a sub alert, or a character portrait from the HUD, where the face should be.
- Face cut off: a tight crop loses the streamer’s chin the moment they lean back.
- Unstable framing: the gameplay half slides left and right as the tool chases motion.
- Gameplay crop error: the gameplay box sits on the sky during a firefight at the bottom of the screen.
Worked calculation: rate by scene and by your stream
The headline rate means little until it is weighted by how you stream. Illustration only, with invented inputs for an unnamed tool:
| Scene | Clips | Passes | Pass rate | Share of your stream |
|---|---|---|---|---|
| Usual corner cam | 15 | 13 | 87% | 70% |
| Different corner | 10 | 7 | 70% | 10% |
| Full-screen game | 10 | 8 | 80% | 15% |
| Mid-clip change | 5 | 1 | 20% | 5% |
- Raw pass rate: 29 / 40 = 72.5%
- Weighted to this stream: (0.70 × 0.87) + (0.10 × 0.70) + (0.15 × 0.80) + (0.05 × 0.20) = 0.81, or about 81%
- Fix time: 11 failures × 1.5 minutes each = 16.5 minutes
For this invented streamer, mid-clip changes barely matter. For a streamer who switches scenes constantly, the same table gives a much lower number, so work out your own shares before comparing tools.
Bias and sample limits
Your own overlay is the biggest source of bias, because a result on one layout says little about another. As reasoning, a simple rectangular cam on a plain background is an easier case for detection than an animated frame, so describe yours when you share a result. Resolution, camera aspect ratio and overlay transparency can matter too.
Small samples mislead most when a tool looks perfect. If a scene passes 30 times out of 30, a 95% upper bound on its real failure rate is still 1 - 0.05^(1/30), about 9.5%. With 10 clean passes the same bound is about 26%. This is arithmetic from the binomial distribution: it is the failure rate at which a clean run of that length would still have a 5% chance of happening. So use eight to ten clips per scene for a first read, and treat a one- or two-clip difference between tools as a tie. Aim for more clips and at least one mid-stream layout change if you are choosing between close tools.
Deciding from a layout result
Match the outcome to how often the failing scene occurs on your channel.
- Passes your usual scenes, fails rare ones: fine, and fix the rare ones by hand.
- Fails your usual scene: check whether the tool lets you set or save the cam region. If it does, one correction may carry across clips. If it does not, move on.
- Passes on every clip, but crops vary: look for a saved layout or template option.
- Fixing takes longer than building it yourself: follow the facecam and gameplay vertical layout guide and use the tool only for finding moments.
Counts per tool and scene can go in the clip evaluation log template. Other measures are on the guides hub.