Every video on this page was generated by TrainWindowAI from the input photo shown beside it, with the same fixed scene customers get. Nothing was trimmed, colour-graded or re-voiced. Problems are labelled, not hidden.
What does a real TrainWindowAI generation look like?
Model: ByteDance Seedance 2.0 Mini via APIMart. Format: 15 s · 720 × 1280 · 9:16 · MP4 with generated sound. Of these 8 videos, 3 kept the car and camera consistent, 2 had minor issues and 3 had visible problems. Model render time ranged from 3 min 0 s to 3 min 47 s; safety checks before and after rendering add time, and no delivery time is guaranteed.
In the October 29, 2026 test batch we submitted 8 photos: 7 generated, and 1 was rejected by the provider at submission with an invalid-image-link error, so no video exists for it. Each photo was submitted once, without retries. This is a small sample; it does not establish a long-term success rate.
Paint, rear spoiler, tail lights and wheels stay recognisable from miniature to full size. Same rear three-quarter angle throughout and one tunnel exit.
The side-view photo is turned into a consistent rear three-quarter view. Body shape and rear spoiler match; the model had to invent the rear, which the photo does not show.
Black mid-size SUV
Consistent
INPUT · Photo by Duc Van on Unsplash · Number plate and a rear-window sticker blurred before upload.
Body style
SUV
Render time
3 min 0 s (model only)
Scene version
Current fixed scene
What we saw
Dark paint, roof rails, light bar and wheels stay consistent. The input photo crops the front of the car; the rear-only camera angle hides this.
Soft top, side intakes and wheels stay consistent. The car leaves the tunnel earlier than planned, and the model adds a rear number plate with unreadable characters.
Colour, chrome grille and body lines match well. The car faces the camera for the whole clip instead of the planned rear view, following the front-facing input photo.
Paint and grille stay consistent, but the truck faces the camera for most of the clip and is suddenly seen from behind in the final outdoor shot, an obvious orientation jump.
Colour and stripes carry through, but the miniature is shown from behind and then turns to face the camera during the transition, so the direction of travel flips.
The home-page example. Car identity and the single tunnel exit hold, but a human hand appears at the window edge in the first seconds, and the window frame leaves the shot quickly. It used an earlier scene prompt; the current one explicitly forbids hands.
How well does AI preserve car colour, shape and wheels?
Paint colour carried through in every video above. Body shape and wheels stayed recognisable in most, but the model invents what a photo does not show: a side-view photo gets an imagined rear, and a front-facing photo can pull the camera to the front, which caused the orientation jumps in the pickup and muscle-car results.
Which photo gives the best result?
The most consistent results here came from photos taken from the side or rear three-quarter, showing the whole car in even light. See the photo tutorial before you upload.
Can I make the same train window video with my car?
Yes. Upload one photo of your car on the home page. It costs $8.99 for one video; a completed video that looks different from your car is not automatically re-generated for free. Read cost and quality and the service disclosures first.
