A Twitch overlay pack is an overlay file, not a generate

Alerts, starting-soon, and webcam frames are text overlays and PIP on a capture. /for/twitch-streamers is the person. The pack is add-text-overlay plus picture-in-picture, not a new cinematic clip.

A Twitch overlay pack is a HUD you composite onto tape you already have: a starting-soon line, a BRB badge, a follow alert, a webcam frame in a corner. It is not a Veo scene of a "cyber studio." It is not a new eight-second film every time the handle changes. Burn type. Place the cam. Stop.

The pack is a file on a capture

The capture is the product. Gameplay, a Just Chatting desk, a VOD grab — that file already exists. The pack sits on top of it. Asking a video model to "make a starting-soon screen with my branding" is a generate of a cousin set: wrong mark, melted letters, a webcam circle that is not your cam. You will still have to composite.

Treat the pack as three objects, none of which is a scene generate: a persistent headline (starting soon, BRB), timed alerts (follow, sub, raid), and a webcam frame in a corner. If any of those is still a prompt instead of a file, you do not have a pack yet. Finish the capture. Then typeset.

Overlay is one authored line, not a cinematic intro

Add a text overlay to a video is add_video_captions: one line you wrote, burned for the whole clip, top / center / bottom. Optional size, color, scrim. The tool does not listen. Silence, game audio, a music bed — none of that changes the string. Position is a three-way enum, not a free pixel. "A bit higher" is not a setting.

The agent verb is add a text overlay to my video: give it the file, the exact line, and the position. One string, whole duration. Cost is a flat per-video fee, shown before rendering. It is not transcribed subtitles. It is not a motion-graphics pack. It is typesetting on this capture.

Alerts that swap — "NEW FOLLOW" at 0s, a handle at 2s, a CTA at 4s — are timed text overlays. You write every line and every start_sec / end_sec. Font, size, colour, and position apply to the set. That is still not a generate. Running a cinematic row to "animate the alert" is how the letters become a different glyph at second three.

Lock the crop first. Top / center / bottom sit in a different safe zone after a 9:16 reframe. Overlaying a 16:9 starting-soon onto a Shorts crop is how the line lands in chrome. Offers and handles change. Copy lock can be later than picture lock. If either is open, do not burn.

Do not ask a scene model to typeset the handle. Scene models paint letter-shaped noise. Overlay is composited, so it is.

Webcam frames are picture-in-picture, not a new face

Add picture-in-picture to a video is create_ugc_video_overlay: a base video plus an overlay video, position required, size and scale optional, remove_black_background if the overlay was rendered on black. Two files in. One composite out. It does not invent a cam. It does not lipsync a JPEG. It does not generate a "facecam circle" from a paragraph.

The base is the capture. The overlay is the webcam clip you already have — a real cam, a keyed talking head, a reaction take. If you only have one video, you are in the wrong row. If you needed a talking still, that is a lipsync job on its own meter, then this composite. Overlaying a misspent generate just to "have a cam" is two identities and one messy file.

remove_black_background strips black. A green screen, a bedroom wall, a white sweep — those are a different removal job. Confirm the matte before you pay for a key that cannot run. Position is a corner enum, not a free-form pixel. For several base clips plus an overlay, that is a compose/stitch job, not this PIP.

A pack that "includes a webcam frame" is a placement, not a person. The person is footage. The frame is this composite.

Streamers /for is the person. The pack is the HUD

AI video for Twitch streamers is the adjacent playbook: hours of clippable live tape, a twenty-second vertical recut, 9:16, six posts a week, TikTok and X. That page is the person and the clip that leaves the VOD. It is not a HUD catalog. Do not rewrite it as a starting-soon generator.

The pack labels a capture so a mute-first feed can still read the line and see the cam. A cinematic "esports arena intro" with invented glyphs is a different film with a handle that is not yours. The test: mute the file. Can a scroller still read the starting-soon line and find the cam? If the answer is a pretty studio the model built, you generated a scene. If the answer is type you typed plus a cam you placed, you made a pack.

FAQ

Can I generate a full starting-soon cinematic instead of overlaying type?

You can. You should not ship it as the pack. The handle will drift. Overlay the line you wrote on the capture you are keeping. Atmosphere, if you even want it, is a later bed under type that already reads.

Does add-text-overlay listen to the stream and write the alert?

No. One authored string, whole file. It does not transcribe. Timed alerts are a different named job. Captions are a third job, and only if someone spoke on a keeper.

Can picture-in-picture invent the webcam if I only have gameplay?

No. overlay_video_url is a video you already have. PIP composites. It does not generate a face. Record the cam, or skip the cam.

Is this the same as /for/twitch-streamers?

No. That page is the person and the off-platform clip. This page is the HUD: overlay plus PIP on a capture. Do not rewrite the streamer audience as a pack, and do not replace the pack with a new cinematic generate.