Thermostats: wall still + overlay schedule, not a talking house
Overlay.
Thermostats: wall still plus overlay schedule, not a talking house. Overlay. The product is a rectangle on a wall. The claim is a schedule you can edit. A talking house is a different catalog row and a worse ad.
Brands keep prompting a sentient hallway that recites "68° at 10pm." The model will invent a face, a voice, and a setpoint that is not the setpoint in the app. Viewers who own a thermostat do not need a mascot. They need the device they recognize and the hours they would actually program.
The wall is the still
Photograph the unit on a real wall, square-on, screen off or showing a locked UI you own. That file is the plate. Animate a slow push or a light shift. Do not generate a new thermostat.
The image-to-video row on Seedance 2.5 is the hold: required still, first frame plus optional last frame, native audio, 4s–30s, 480p / 720p, listed 47 credits, billed resolution-based per second. One frame into continuous motion, no stitch of shorter clips. If the shot must land on the faceplate, use that as the last frame.
720p is the ceiling. Enough for a wall. Not a reason to skip the still. Image-to-video keeps the upload as frame one so the bezel, the mark, and the screw holes survive.
A talking house is a presenter job. If you actually need a face, that is a different row after the plate exists. Most thermostat ads do not need a face.
The schedule is typed
Wake, leave, return, sleep. Those four lines will change in legal review, in a regional SKU, and in the next firmware screenshot. They do not belong in a video prompt.
Add timed text overlays is the tool: you write every line, you set start_sec / end_sec, Versely burns them in order. Font, size, color, and position are one look for the set. A persistent "Schedules, not slogans" badge is the simpler text overlay — one fixed line for the whole clip.
This is not captions-from-speech. Nothing is transcribed. If a VO later says the hours, add captions can burn what was spoken. The schedule on screen should still be the typed one, because ASR will not protect a degree symbol.
On-screen text in generated video is the failure mode you are avoiding: a generated "10pm" that becomes "10am" for six frames.
Do not generate the display
A thermostat screen is a UI. Generated UIs invent menus. Prompt a dark or blank faceplate. Composite the real screenshot if you must show the app. Overlay the schedule next to the unit, not as fake pixels on the glass.
Native audio on Seedance can be the HVAC bed — a faint compressor, a click of a relay — if you write it into the prompt. It is not a house that talks. Diegetic plant noise is a mix note. Dialogue is a presenter.
Keep Movie Mode for a brand film about comfort. This unit is a wall, a still, a few timed lines.
FAQ
Why not let the model write "68°" on the glass?
Because 68 is a claim. Overlay text is editable without regenerating the wall. In-frame type is a new generate every time the setpoint changes.
Is Seedance required, or will any I2V do?
Any image-to-video row can move a locked still. Seedance's I2V row is the one that lists first/last frame and a 4s–30s single pass at up to 720p. Use it when the bezel has to hold. Use a shorter 1080p close only if the still is already the ad.
Can I overlay a full seven-day grid?
You can type it. Timed overlay is built for a sequence of short lines, not a dense spreadsheet. A seven-day grid is a still (or a screenshot) composited, then a few timed callouts. Do not ask one caption render to be a programming manual.
When is a talking house acceptable?
When the brief is entertainment, not the SKU. The thermostat in the listing photo still has to match. Generate the mascot on a presenter row; keep the product plate separate.