For game streams and reaction videos: the streamer's webcam and the game (or the video they're reacting to) laid out together in a 9:16 Short, in the layout you choose.
A game stream is one wide picture with the game in the middle, a webcam in a corner, and chat, alerts and panels around the edges. Cropping it to 9:16 the standard way follows the biggest face, and on a game stream that can be a game character, a portrait, or the person in a video the streamer is reacting to.
It is a switch of its own, off unless you turn it on, and it can't be combined with Vertical Live, Podcast or Longform. With it off, nothing about processing changes. If anything in it fails, that clip is made the standard way.
It is for streamers on a real camera. The detection is built to find people. VTubers aren't supported: an avatar is never taken for the streamer, so a VTuber stream gets the game on its own.
Eleven layouts, chosen from cards that show the video's own frame in each one.
All of them are data (gaming/layouts.json), not code.
| Layout | What it is | Boxes it uses |
|---|---|---|
| Split | Webcam band on top, the game fills the rest. The divider sets the webcam's share, 25–50% (38% to start). | Webcam, Game |
| Basecam | Split with the game on top and the webcam at the bottom. | Webcam, Game |
| Half | Webcam and game 50/50. | Webcam, Game |
| Fullscreen | The game cropped to fill the Short, no webcam. | Game |
| Blurred | The whole game in the middle, on a blurred copy of itself. | Game |
| Small facecam | The game fills the Short, a small webcam near the top; move and resize it on the preview. | Webcam, Game |
| Circle facecam | The same with a round webcam. | Webcam, Game |
| Game UI | Webcam on top, the game below, and a piece of the game's UI (a scoreboard, a map, a timer) as a layer over the game: against the webcam to start, moved and resized on the preview. | Webcam, Game UI, Game |
| Mosaic | Webcam and a game UI panel side by side on top, the game below. | Webcam, Game UI, Game |
| Dual facecam | The game fills the Short, two round webcams near the top (duo streams), always the same size. | Webcam, Webcam 2, Game |
| Duo split | Two webcams side by side on top, the game below. | Webcam, Webcam 2, Game |
On top switches the webcam and the game in Split, Half, Game UI, Mosaic and Duo split; Basecam is the same switch set the other way.
The game is either zoomed to fill its part of the Short or shown whole. Shown whole, it sits right against the webcam: never a band of blur between the streamer and the game. The webcam and the game are stacked together, and the space left over is a blurred copy of the game, above the two (just clear of the platform's top bar, so the streamer's head isn't under it) and below them, where the platform's captions and buttons go anyway. Only Blurred, the game alone, sits in the middle of the Short. There are no black bars.
With no webcam found, a layout that needs one falls back to Blurred. A layout that needs a Game UI or second webcam box that was never drawn falls back to Split or Small facecam.
TikTok, Reels and Shorts draw their own buttons, captions and top bar over the video. A webcam crop that is only centred can put the streamer's head under the top bar: measured on one Marvel Rivals stream, the head sat 2–26 px from the top of every clip, because that webcam has barely any room above the head.
So the webcam is placed from the streamer's head, not the middle of the box:
- the head is found in the webcam box of each clip (the pose model, a few frames, inside the box only; it places the crop, it never decides who);
- the head's top lands at least 6% of the region below its top and below the chosen platform's top bar, and the chin above the caption area where the region allows it;
- the chin comes first: when a tight webcam's face is too tall for its band, the hair goes under the top bar (and the preview says so) rather than the chin being cut off by the game. A bigger webcam share, or the whole game, gives it room;
- when a webcam at the top of the Short has too little room above the head, the picture moves down inside its region (by just what's missing, at most 30% of the region), with a blurred copy above it rather than a cut-off head. A webcam lower down (under the game in the split) doesn't move: there the top bar isn't over it, and the move only put a band of blur between the game and the streamer's head and pushed the chin toward the captions;
- small and round webcams go inside the platform's safe area.
The same stream re-rendered: the head top at 135–158 px with the webcam on top (TikTok's top bar ends at 140), about 200 px in Small and Circle facecam.
Safe areas, on a 1080×1920 Short. None of the apps publishes these for ordinary posts, and the guides that do disagree by tens of pixels, so they were measured (September 2026) from replicas of each app's feed on a phone taller than 9:16, where the apps fill the height (kreatli.com's safe zone checkers): each value is just past the app's own element — TikTok's tabs, the Reels and Shorts title bars, the column of buttons, and a caption block of two lines.
| Platform | Top | Bottom | Left | Right |
|---|---|---|---|---|
| TikTok | 180 | 420 | 60 | 190 |
| Instagram Reels | 245 | 400 | 60 | 205 |
| YouTube Shorts | 250 | 350 | 60 | 190 |
| All three | 250 | 420 | 60 | 205 |
The platform is chosen in the preview (TikTok to start), which draws that app's layout over the Short from the same measurements: the top bar, the buttons down the right with their counts, the name, caption and sound at the bottom (Shorts with its Subscribe button). The icons are open-licensed sets in each app's style (Material Icons, which YouTube itself uses, and Lucide), not the apps' own artwork or logos; "All three" shades the areas they share. It says whether the face is clear of them: "✓ Face clear of TikTok's UI", or what to change. Camera on top is usually the fix: with the game on top, the webcam band sits in the caption area.
Every game and every stream overlay is different, so the layout can be set up on the video's own frames before any processing starts. Tick Gaming / Reaction on a video in the Generate bar (a YouTube, Twitch or Kick link, or a file) and Choose a layout opens:
- Layouts: the eleven cards, each a live miniature of this frame;
- the frame, with a box for each thing the layout uses (Webcam, Game, Game UI, Webcam 2), already on it when it opens, as in StreamLadder: drag a box to move it, any of its eight handles to resize it. Boxes snap to the frame's middle and edges, to each other and to the stream's solid panels (a game box lands exactly on the chat bar's edge), with a guide line. While a box is dragged the grid shows (thirds and the middle) and it snaps to that too; Grid keeps it on. Alt places a box freely. The Game box stops at the webcam: while it's dragged, the webcam's edges show as dashed lines across the frame, 1% of the frame past the webcam box so its border is left out too, and the game box snaps to them and stops there instead of going over the webcam, so the streamer isn't shown in the game as well. With Alt it goes over, and a note says the webcam will show in the game. Zoomed to fill, the Game box has the shape of the game's space on the Short: change the webcam's share on the preview, the layout or which goes on top, and the box takes the new shape (its size and middle kept, clear of the webcam), so what's inside it is exactly what renders. Before, the render cut a slice of a different shape out of the middle of the box. The box as drawn is kept, so going back and forth doesn't wear it down; with Whole the game is shown whole, so the box stays as drawn. The Game box starts on the area the layout takes the game from; move it and it's yours (Let Clips Kitty pick the game area gives it back). The dashed line inside is exactly what the layout will show. Snap to the webcam's border pulls the webcam box out to the overlay's own edge. The stream's solid panels are hatched Left out;
- Moment: a slider over the whole video and five frames spread across it (nothing is downloaded for a link; each frame is read straight from the stream);
- Preview: the 9:16 result, live, with the platform overlay and the face check. Drag the line between the webcam and the game to change their shares. In the facecam layouts drag the facecam on the preview to move it, and any of its eight handles to resize it (it keeps its shape, from the opposite side; two facecams resize together); the same for the Game UI layer. While dragging it snaps, like a design editor, to the middle of the Short, its edges, the platform's safe lines and the other parts' edges and middles, and a pink line shows what it lined up with (hold Alt to place it freely). Grid shows thirds, the middle and the platform's safe box, and adds the thirds to what it snaps to. Reset puts the layers back. The Game UI in Mosaic and Game UI is cut to its space's shape, like every other part: no blur round it;
- On top: Camera or Game. Webcam: Draw it (the default: the box on the frame), Find it (by who is talking, when processing) or None. Game: Whole or Zoom to fill (and then Left, Centre or Right).
When it opens, the webcam box is moved onto the webcam when Clips Kitty finds one: a person in the same spot on at least four of the five frames, no more than a third of the picture, with a real border round at least two of its inner sides. A game character moves between frames minutes apart, chat has no person in it, and an avatar has no webcam border, so none of them are suggested. It's only a starting point, for you to check. With no suggestion the box waits in the corner for you to drag onto the webcam.
The editor has window buttons top right: fullscreen (the whole monitor, like a video player; Esc leaves it, and it opens that way next time if you left it so), and minimise to a bar in the corner, to get at the rest of the app and come back with Restore. Esc closes the editor, never the clip editor behind it.
Use this layout sends it with the video. Remember for this creator's next videos keeps it for them, so their next videos (and a watched channel's) start from it. Change layout… in the Generate bar opens it again.
The same editor is in the clip editor (Effects → Layout → Gaming / Reaction → Change layout…) to change one clip: shown in Update preview, saved on Apply. Processing the video again with Gaming / Reaction on keeps a clip you edited in its own layout. When that clip's split was found automatically or remembered for the creator, a webcam you set up for this run, or one remembered for the creator, reaches it too. A split you set up before an earlier run, a webcam you drew for that clip, or a split you turned off for it stays as it is, even with Gaming / Reaction off for the new run (a Sports job uses its own layout). Change it in the clip editor.
TalkNet decides: the streamer is the face that speaks in sync with the stream's audio. Size never decides. This is the same fix that ended the "largest face" problem in standard processing: a character can be as big as it likes, but it doesn't move its mouth with the stream's audio.
Measured on real streams, "who is speaking in this clip" isn't always "who the streamer is", so four things sit around TalkNet:
- On screen for most of the clip. TalkNet scores whatever face it is given. A driver glimpsed through a car window in GTA for 3% of a clip got a full speaking score. A webcam is on screen the whole time.
- The most confident speaker. When two faces both speak (a watch party, the streamer and the person in the video), the one TalkNet is surest of wins: 3.3 against 0.3 on the stream we measured.
- A real person. Game characters that lip-flap to voice acting (a 3D visual novel) and VTuber avatars score as speaking, but TalkNet is never confident about them. A webcam's confidence reached at least 0.3 in some clip of every stream measured. Characters and avatars stayed at -0.2 or below.
- The whole video, not one clip. In a reaction the person in the watched video can out-talk the streamer for a whole clip. So the webcam is decided once per video, from four of its clips spread through it: the face that speaks from the same spot in the most of them.
The webcam box starts as the streamer's own box, with no margin (a margin ran past a speedrunner's webcam into the chat beside it). Each side then grows out to the webcam overlay's own border where there is a clear one, and stops just inside it, so the half shows the webcam and nothing beside it. On six streams with a webcam, every side of every box landed inside the hand-marked webcam.
Streams don't keep one layout. Many open with an hour of Just Chatting before the game; a reaction streamer often goes full screen on their camera between videos; and the webcam itself moves when the streamer changes scene (measured: top left over a browser while chatting, lower down on the left for the game; bottom left, then top left, in a two-hour reaction). So the layout is decided per clip, from what is on screen in it, not per video, and not from the stream's category (YouTube and Kick don't say which part of a stream is which). Each clip gets a quick look (two frames a second), and somebody at the webcam keeps the split with no more checks. Otherwise, even with a webcam drawn in the setup or remembered for the creator:
- The streamer's camera filling the frame: framed like a talking-head clip by the standard renderer. TalkNet runs only when somebody on screen is too big to be in a webcam, and only a confident real person takes the clip: the streamer watching an old stream of their own, their face filling the video, kept the split with their real webcam.
- The webcam somewhere else: followed there. The clip is looked at the way the setup editor suggests a webcam (a person in the same spot in nearly all of six frames, overlay-sized, inside a real border), about half a second a clip, and the split uses that box for this clip. Game characters standing around in GTA were never taken for it.
- Nothing certain (a cutscene, an empty chair) keeps the webcam that was set up: a person's choice isn't overridden on absence alone.
- The automatic webcam search looks past those clips. A clip showing the camera filling the frame says nothing about where the webcam is, so it doesn't count as one of the four looks; another clip, from further across the stream, is looked at instead (eight at most). Before, an opening hour of chatting could take three of the four, leave one vote for the webcam, and every game clip rendered without the streamer.
- Such a clip is saved as what it got, so the clip editor shows its standard framing and a re-render keeps it; Gaming / Reaction in its Layout row switches it back.
- Scoring is the same across the parts. What is said counts as on any stream, and the game adds on top; what the picture shows only takes points off a quiet clip, so a Just Chatting clip with no gameplay isn't marked down for it.
The game area is never detected. Earlier attempts looked for the part of the screen with the most going on, and scrolling chat won every time. So:
- Whole game: the biggest picture beside the webcam and the solid panels (below) that leaves them out, so the streamer isn't shown twice. With no webcam, the whole stream.
- Clear of the webcam's border: both keep 1% of the frame away from the webcam box on every side. The box stops just inside the overlay's border, so a game cut right at its edge showed the border as a line down the game (seen on a reaction whose webcam moved to the top left).
- Zoom to fill: a crop at the region's shape, as tall as it can be while it stays clear of the webcam and the solid panels, on the middle of the game picture.
- A drawn game area replaces both.
Solid panels are the parts of the stream layout on their own background: a black chat bar under the game, a speedrun's splits timer. Cropped into the Short they're a useless bar, so they're found and kept out:
- four or more stacked lines of text in the same place on most of five frames spread through the video (the game's own text comes and goes; a panel stays);
- on the same background colour on every frame. Behind see-through chat the game changes, so chat drawn over the gameplay stays in the picture: it's part of the stream;
- grown out over that colour to the panel's own box, one side at a time.
It is OpenCV only, a fraction of a second, and no motion is involved. On the 13 test streams it found the speedrun's black chat bar and splits timer and nothing else: not a night-time game (its dark sky shifts between frames), not a HUD, not see-through chat.
Standard scoring judges talk: hooks, opinions, drama, quotable lines. On a game stream the moment is usually something that happened in the game (a kill streak, a boss going down, a goal) and the reaction to it, often with little said. Gaming / Reaction scores that way, and so does a live that was already vertical: tick Vertical Live and set Vertical Live content (the row under it, like Longform's output) to Gaming / reaction instead of Talking / IRL, in the Generate bar, a queued video's settings and a watched channel's, whether it's a game or reacting to videos. One Vertical Live, two kinds of content, not two pipelines. Everything below applies to both; only the framing differs (a Vertical Live keeps its own 9:16 layout, so the webcam and game area aren't needed). With neither on, scoring is exactly the standard one.
What goes into it (research: stream-highlight papers on chat, audio and facecam signals; what gaming clip tools look for; what performs as a Short):
- The game. Twitch says which game is played over which part of a stream
(a stream that goes from Just Chatting to Rust to Fortnite), Kick gives its
category, and a YouTube video's title and tags often name it.
config/gaming.yamlturns about 150 game names into a kind of game (shooter, battle royale, MOBA, sports, fighting, racing, horror, soulslike, action-adventure, sandbox, speedrun, party, strategy, reaction) with what a highlight is in it and what a streamer says when it happens. - The AI is told. Every scoring prompt says it is a gaming or reaction stream and which game (for the part of the stream it is reading), to judge what is said exactly as on any stream, and to count in-game moments too: what a highlight is in that game, that a clear in-game moment is a strong clip even when little is said, and that menus, queues, loading screens and reading out donations score low. The moments the signals found are read eight at a time: all at once, a 20-minute stretch's 45 moments ran past the model's answer budget and every one fell back to a neutral 50.
- Chat's reactions (Twitch VODs and YouTube live replays; Kick keeps no chat). A burst far above the stream's own message rate marks a moment, dated about 6 seconds earlier for chat's delay, and what chat says names it: hype (POG, NO WAY, すご), laughing (KEKW, LUL, ｗｗｗ, ㅋㅋㅋ, хаха, jajaja), surprised (！？, あ, WHAT), scared (monkaS), a fail (F, NotLikeThis, BigSad) or "clip it", including a channel's own emotes by their ending (kittyPog). Bursts of hellos don't count, and a burst of long messages is chat discussing something, not reacting.
- The streamer's voice. A sudden jump in loudness while they are talking (a shout, a laugh, a scream): the cheap stand-in for seeing their face react. It is measured against their own talking over the minute around it, twice their usual voice and up. Against the stream's overall level, which on a quiet game stream is its silences, just talking counted as a shout.
- The game's own sound. A small sound model (PANNs, trained on AudioSet's
527 everyday sounds; 24 MB, bundled) listens to every second for gunfire,
explosions, a crash or a shield shattering, a crowd cheering, a referee's
whistle, screaming and laughter. What each means depends on the game
(
config/gaming.yaml'sgenre_sounds): gunfire is the fight in a shooter and nothing in a football game; a crowd roars at a goal, and Rocket League's goals explode. A stream mostly sounds like the streamer's voice and music, so a sound is judged against its own level in that stream: a Rocket League goal's explosion scores about 0.25 out of 1 and still stands out. It runs beside transcription, about 600x realtime on a GPU and 120x on a CPU (a 3-hour VOD in about a minute and a half). - What the game writes on screen. Games announce their moments in big
letters (ELIMINATED, VICTORY ROYALE, PENTA KILL, "X A MARQUÉ", YOU DIED,
ENEMY FELLED) and fill the screen with text when nothing is happening (a
settings page, a queue). OCR (RapidOCR, bundled) reads the middle of the
frame in the game-moment windows, strongest first: every 2 seconds near the
moment, since a banner is only up for two or three, then every 4, for at
most 2 minutes a video (about 0.4-0.8 s a frame on a CPU). A banner with an
event word for that kind of game becomes an event the AI reads ("ON SCREEN:
ACE") and the clip's breakdown shows; it adds no points of its own, because
most videos never show one and a caption can look like one. A window where
at least two thirds of the frames show menu words is marked down a little
(-6, less the more is said over it: nothing for a clip full of talk) and
gets no bonus, whatever chat made of it.
config/gaming.yaml'sscreen_textlists the words in the languages games are commonly played in. It reads Latin script and kanji, not kana. - The AI looks at the best clips. Last, before the final ranking, the local AI model is shown four frames (at 896 pixels) of each of the best 12 clips, with what was said and what the signals picked up, and asked what is on screen and what happens: a clear moment raises a clip by up to 10, a menu, a loading or black screen lowers it by 10, ordinary play leaves it alone. What takes points off counts only as far as the clip is quiet: the full 10 when nothing is said, half when half of it is talk, nothing for a Just Chatting clip full of talk. The clip's breakdown says what it saw ("SEEN: scores a goal"). About 10 seconds a clip on a GPU, for at most 4 minutes a video. Only a local model that takes images does this (Gemma 3 and Gemma 4 do, gemma:7b doesn't), and it is first asked the colour of a plain red square, since a model can say it takes images and not get them. The video picture never goes to a cloud AI, so with one this step is skipped, and the job log says why.
- Talk first, the game on top. A clip with talking is scored exactly as on any stream, with the same weights: the words count as much as ever. The game adds to that the way creator context does, never taking anything away: up to +7 for how strongly the game channel (everything above) marks the clip, and +5 more when two independent witnesses agree (chat, the game's sound, the streamer shouting or laughing) or chat and a loud moment do; +12 at most. The game's sound with loudness alone doesn't count as two, because gunfire is loud. When little is said, the weight the words would have had goes to the game channel and the sound instead, so a quiet streamer's big play isn't marked down for its silence. (Weighting the game at 40% of every clip, as this did at first, left a two-hour reaction to BlizzCon with 13 clips: most of a reaction is talk.) The standard "active content" bonus (a person on screen, moving) doesn't apply: a facecam over a moving game is that all stream long.
- Each moment is a candidate of its own, from about 4 seconds before it to the reaction, 15 to 35 seconds, even when nothing was said near it.
The clip's score breakdown shows game and what marked the moment.
What it found on real VODs (September 2026, public Twitch VODs):
- A 94-minute Japanese Apex Legends VOD. Chat alone marked 4 moments: a death the streamer took (chat: the channel's RIP emote, BigSad, NotLikeThis), and also a black screen and a character-select screen. With the game's sound, 23 moments (the budget for that length); the two where chat and gunfire agreed came first, and 8 of the top 9 were fights on screen. Reading the screen marked the settings page chat had reacted to as a menu (感度, 設定), and none of the 22 others; it named no kills, since the game was in Japanese and its kill text is kana.
- 30 minutes of Rocket League, played in French: 27 moments, nearly all goals (the explosion, with its replay a few seconds later counted as the same moment). Reading the screen named 5 of them from the banner ("A MARQUÉ"), and marked the one crowd cheer that landed on the matchmaking menu after a match (COMPÉTITIF, INDISPONIBLE, MODE DE JEU).
- The AI looking (gemma3:4b, the model setup installs for a PC without a graphics card) at six of these moments got all six right: the Apex settings page and a black screen with only the webcam (-10 each), two Apex fights (+6, "firefight with multiple opponents"), a Rocket League goal (+6), and the matchmaking menu (-10). At 512 pixels a frame it had taken the fights for ordinary play. gemma4:e4b on the Ollama release installed here said it takes images but described a settings page as "a dark, abstract background": the red square catches that.
- 30 minutes of a stream filed under a football game that was really two streamers watching someone else's IRL stream: one weak moment, nothing strong enough to become a candidate on its own.
What this can't do. Reading the screen only knows the words it has been given, in the scripts the OCR reads. The AI only looks at the best 12 clips, not the whole stream, and only with a local model that takes images: with gemma:7b or a cloud AI, a menu chat reacted to is caught only if its words are on the list. Nothing here has been run on a whole quiet vertical stream yet.
Scored as gaming, the "reaction" signal (is a person on screen, being emphasised?) is left neutral, in the split and in a Vertical Live alike: on a game stream it counts game characters as people, and a top-down game as nobody at all; on a reaction stream, the people in the video being watched. A Vertical Live set to Talking / IRL keeps it.
September 2026, public Twitch VODs, three 40-second windows from each of 13 streams. Webcam positions were marked by hand from a frame of each and compared with what was found.
| Game | Stream | Result |
|---|---|---|
| Zelda: Breath of the Wild | speedrun, webcam bottom-left, splits timer and chat around it | ✓ split: webcam found, inside the hand-marked box on every side |
| Zelda: Tears of the Kingdom | VTuber | not supported: the avatar wasn't taken for the streamer, and the game shows alone |
| "Zelda" category (really a gacha RPG) | no webcam, a large anime character on screen | ✓ the game alone; the character was not taken for the streamer |
| World of Warcraft | webcam bottom-left, bags and action bars | ✓ split |
| World of Warcraft | just chatting, camera fills the frame | ✓ framed the standard way |
| Grand Theft Auto V | roleplay, no webcam | ✓ the game alone; a driver seen for a moment was not taken for the streamer |
| Grand Theft Auto V | reacting to a bodycam video, small webcam | ✗ webcam not found: the streamer mostly listened, and TalkNet was never confident about their face. Draw it in the setup. |
| League of Legends | lobby and loading screens, webcam bottom-right | ✓ split, inside the hand-marked box on every side |
| League of Legends category | a watch party: two webcams and a documentary | ✓ split with the streamer's webcam; when her camera went full screen, framed the standard way |
| Rust | webcam top-left, chat under it | ✓ split, chat left out of the webcam half |
| Dota 2 | two casters' webcams and a player cam | ✓ split with a caster's webcam |
| Persona 3 Reload | VTuber, voiced characters | ✓ the game alone: no character was taken for a webcam |
| Pixel-art game | VTuber | not supported: the game shows alone |
| Grand Theft Auto V, after an hour of Just Chatting | camera full screen at the start, then the webcam top left over a browser, then lower on the left in the game | ✓ 27 of 27 clips (9 clips, each with the webcam drawn on a chat frame, drawn on a game frame, and found): full-screen camera framed the standard way, the webcam followed to where it was in each part |
| World of Warcraft reaction to a games showcase | webcam bottom left, top left for a while | ✓ followed to the top left in that stretch; the other 5 of 6 clips kept the drawn webcam |
Time on an RTX 3060: finding the webcam looks at four 40-second pieces of the video, about 15 seconds each (6 for person tracking, 8 for TalkNet), and a few more when some show the camera filling the frame. Each clip then checks that the webcam is there, a few seconds, instead of the standard face tracking. A split set up before processing skips the search; its clips get the quicker look (two frames a second) for a camera filling the frame.
scripts/gaming_detect_bench.py repeats the measurement on any footage.
- VTubers aren't supported. The detection is for people on camera.
- One layout per clip. A clip that moves between the game and a full-screen camera, or across a webcam move, keeps one layout: the one most of the clip shows.
- The Game UI and second webcam boxes are drawn by hand. Nothing looks for a scoreboard or a second streamer.
- YouTube frames for the editor are read over IPv4. On some networks FFmpeg's IPv6 connection to YouTube hangs for minutes; the editor's frames go through a small local relay that forces IPv4.
- Separate recordings (the game and the webcam as two files, as some recorders make) aren't supported yet.
Installed, enabled SDK pipelines with category: gaming appear in the Gaming
detector dropdown. Choosing one specializes finding and scoring for that game,
with shared Gaming context, layout, captions, editing and publishing. The
external detector owns game-specific finding and scoring; the generic Gaming
finder is not also run. Its windows still go through Gaming's shared visual
review, using the existing OpenCV activity signals, available screen-text OCR,
chat/game sound evidence and configured local image model. The usual model
availability and image-capability checks apply. YOLO webcam detection/tracking
and rendering use the same Gaming path and settings. Detector defaults are used
in this flow; selecting a detector installs no additional vision model or FFmpeg.
Other plugin pipelines retain their standalone finder behavior, including when a Gaming layout is applied to them. Gaming detectors are not offered in Sports; the API rejects their rate/understand/edit steps for Sports. Game-specific recognition stays in the external SDK plugin, not in the host.