In February I gave Kling a 1992 Escada scan of Gail Elliott and got fifteen seconds of video back, which felt generous at the time. FLUX 3 Max gives me twenty. It’s a version of Black Forest Labs’ FLUX 3 that fal post-trained with the lab, and fal says it sits level with the best models in Artificial Analysis’s early image-to-video rankings while making a ten-second clip in about eleven seconds. Seedance 2.5 takes 277.8 seconds for the same job.

I started with Gail again, a different scan this time: a grey jumper with yellow knitted leaves across one shoulder, yellow gloves, a gold link belt and a plain white studio wall. Knitwear is where these models usually give up. Cable stitches crawl, appliqué slides around, and a jumper turns to porridge the moment its wearer moves. Here Gail turns round and pushes a gloved hand through her hair, and the leaves stay exactly where somebody sewed them. Her face stays hers for all twenty seconds, which matters more to me than any leaderboard.

Amber Valletta was the harder test, because the original is a tight portrait with almost nothing to move except her: a red coat, a purple scarf, one pearl earring, and city lights blurred to orange behind her. I expected twenty seconds of nervous fidgeting. Instead she closes her eyes, touches her hair and lets a small smile arrive, slowly, the way people do when a photographer has been talking to them for an hour. The pearl never flickers.

It isn’t flawless. A beach photograph I tried this week came out in slow motion however I worded the prompt, and I still don’t know why. For portraits, though, this is the first video model I’d trust with a scan I actually care about.

Sources: