Real outputs, unretouched
Every clip below came out of the pipeline exactly as you see it. No cherry-picked frames, no hand-fixing, and the failure cases are on this page too.
Aerial beach
The shoreline, the shadow and the swimmer's beach all continue outward. Hardest case here — a real person in frame.
Mountain lake
Strong horizon and a mirrored reflection — the two things outpainting usually breaks. The raft extends toward the camera.
Palm trees
Trunks and sky invented above and below the original frame.
Blossom branch
Watch the top centre: the model invented a lens flare that wasn't in the original.
Aerial clouds
Soft, low-detail subjects are the easiest case — cloud and sky extend cleanly.
Pink flowers
Dense repetitive texture. Convincing at a glance, looser under scrutiny.
What this does not do well
- People, faces and hands. If a person is near the edge of your frame, the AI has to invent the rest of them, and it usually gets it wrong. Scenery is where this tool earns its money.
- Interiors and busy backgrounds. Rooms, furniture and crowds can produce objects that were never there.
- A ~6% trim. The model rescales your clip slightly, so roughly 6% of the frame edge is lost in the conversion.
- 720p output. Converting a tall clip to widescreen fits your footage into a smaller region of the new frame, so the original detail is reduced. This is not a 4K upscaler.
Making clips longer is experimental
We also test adding seconds to the end of a clip. It holds up on simple continuous motion — waves, clouds, driving — but it cannot reliably judge scale, so asking it to add a subject tends to produce something the wrong size. We are not selling that as a finished feature, and we will say so before you spend a credit on it.
Scenic footage is what this is good at today. If that is not your footage, wait for the next model.