Transit stations: empty-platform stills plus timetable overlay, not a commuter film
Architecture of the platform edge you operate. Timed overlays for headway. TTS the announcement as a file, not a talking conductor. Do not invent riders.
Architecture of the platform edge you operate. Timed overlays for headway. TTS the announcement as a file, not a talking conductor. Do not invent riders.
A station clip that informs is a facility plate: this edge, this tactile strip, this canopy, the clock operations already posted. It is not a packed platform of generated commuters, and it is not a constructed conductor walking “you should take the 7:10.” Photograph the empty-enough edge. Burn the headway. Lay the announcement as audio. Leave the crowd off the generate.
Empty-platform stills are the architecture
The honest still is a photograph of the platform you operate — same coping, same sign family, same roof line a rider already knows from the slip. Text-to-video will invent a prettier station and seat extras you did not cast. That is a failed facility plate. If the still is already wrong, every second of motion is waste.
Dawn, empty, or staff-only is a legal still. A packed platform of generated faces is not. If the only photo includes a rider’s face, crop them or shoot again. A generated commuter is a private extra. A generated conductor is a replica.
AI image to video keeps frame one as the file you sent. The prompt is for physics — a slow hold, light on the edge, no new canopy. Re-describing the station is how the model draws a second platform. Prompt the travel. Stop. “Commuters boarding, children running for the doors” is talent you did not hire. The architecture here is the coping you maintain, not a speculative concourse.
Kling O3 Standard Image to Video is the still-to-motion row: image required, reasoning-enhanced, native audio, 3s–15s, listed 56 credits. The still is the brief. The model moves it. No still, no generate. A paragraph where the platform edge should be is a wrong label.
Write only camera. “Locked edge, slow push, empty platform, no new riders.” Native audio is on. Write a quiet pad, a distant rail bed, or silence. A silent brief still gets a soundtrack you did not choose. Iterate at 3s or 5s until the coping holds. Do not ask Kling to letter the next-train time on the LED. Overlay that string. Do not invent a train that was not in the still.
Timed overlays for headway
Headway is a string operations already uses: “Trains every 8 minutes.” “Last train 12:40.” “Elevator out, use Broadway stairs.” Write that string. Do not ask a model to letter it into a destination sign.
Add timed text overlays to my video is the named agent job: you supply each line with start_sec and end_sec. Nothing is transcribed. Headway, then the next-train clock, then a service note is three authored beats on a locked plate. Font, size, colour, and position apply to the set, not per line. Flat per-video credit fee, shown before rendering.
The sibling editor surface is add timed text overlays to a video. Same job, different door. Auto-captions listen. A mute platform still has nothing to transcribe.
If you only have one persistent badge for the whole clip, that is a single overlay on the whole file. Timed overlays exist because the text changes. Picture lock first. Crop to the ratio you will ship before you overlay. Headway changes when a disruption calendar moves. The coping does not.
TTS the announcement as a file
Gemini 3.1 Flash TTS is a 4-credit text-to-speech row: thirty voices, natural-language style control, inline tags ([sigh], [laughing], [whispering]), multilingual synthesis, multi-speaker dialogue. That is a voice file. It is not platform physics. It is not a talking conductor.
Read the announcement you already post. One headway, one clock, the wording legal already cleared. Lay that file under the locked platform clip. Add voiceover to video is the finish: audio on a picture you already trust. The face stays off.
A closed-mouth TTS read under a generated conductor is still a presenter. The voice is the host. A real named operations lead is a different day and a different file. Generation can still do the edge. It cannot be the crew.
Write “no speech” on the Kling take if the mix should stay rail. Then decide whether the timetable is overlay type, a TTS bed, or both. Do not ask Gemini to invent a bulletin operations did not publish.
Do not invent riders
Photoreal commuters the model invented are a likeness problem. Minors on a generated platform are a hard stop. “Fill the station so it looks open” is not a slider. Open is a photograph of a peak you actually ran, with people who were actually there, releases in the folder — or it is the empty edge plus hours.
Do not prompt a packed platform, a child running, a conductor with a mouth in frame, or a readable ticket. Those are people. A silent empty edge is not that shape. Authored type is typesetting, not a mouth. TTS is a file, not a face.
The test is a pause. If a regular can name the station from the canopy, you locked the right still. If a parent might ask who that child is, you shipped a replica. Print the time. Speak the announcement as a file. Leave the platform as you photographed it.
FAQ
Can we add a generated crowd so the station looks open?
No. Photoreal riders and constructed faces are a likeness problem. Photograph peak service with real people and releases, or keep the still empty-enough and overlay headway.
Why TTS instead of generating a talking conductor?
Gemini 3.1 Flash TTS is a speech row: 4 credits, thirty engine voices, an announcement as a file. A talking conductor is a video generate plus a likeness. Cast a real named person, or keep the face off the line. The platform still does not need a mouth.
One overlay or timed lines for headway?
One persistent headway badge is a single overlay on the whole file. Headway, then the next-train clock, then a disruption note is timed overlays, with timestamps you set. Neither job transcribes. Neither job belongs in the Kling prompt.
Does Kling O3 Standard Image to Video work from text only?
No. Image is required. Native audio is on. Durations 3s–15s. Listed 56 credits. Photograph the empty platform edge, then animate that file. Text-to-video of a busy station will invent riders.