How the test was runSame photo, same prompt, first clip back
Seven photos, one plain prompt each, written the way a person types into a box. Every model ran once per photo through fal.ai on 18 September 2026, at its 720p tier and five seconds long, with sound switched off where the model has a switch. Nothing was retried or cherry-picked. The clip you see is the first one each model handed back.
On this page every clip is re-encoded at one setting, so file size and sharpness do not favour any model. Open a clip to watch it at 720 pixels.
Our score is a mark from 1 to 5 per clip. A 5 moved the way the prompt asked and still looks like the photo. A 1 changed the subject or broke the scene. Readers vote for one clip per photo, one vote per person, and can change it.
Photos by Nguyễn Phát, Edyta Stawiarska, Ilona Frey, Fajar Arroisi and Анатолий Стафичук on Pixabay, and StockSnap. The cabinet card is in the public domain, via rawpixel.