0.35.0 — 2026-08-12

The frame you approve now has your actor in it

Stills are drawn from your cast photographs, text cards stop being filmed, and a generation that fails gives your tokens back.

Attach up to four reference images when generating an image, not one. The picture follows all of them — a subject, a style, a scene. Video generation has taken four for a while; images were limited by us, not by the model.
Seedance 2.5 is available in the video model picker. Takes up to 30 seconds in a single shot — twice anything else we offer — and generates sound alongside the picture rather than as a separate step.

The still you approve is now drawn from your cast photographs. It never was — the frame was generated from the words alone, so your actor's grey tee came back navy and a product's "neon green logo" wandered onto a shoe. Words steer a picture when the photographs are missing, and you were approving frames your actor had never been in. Redrawing a single frame carries the photographs too.
A text card is no longer filmed, and no longer charged for. A 90-second film that is mostly white text on black was quoted 190 tokens — most of that to have a video model produce black rectangles it is not even given the words for. Text belongs on the timeline, where it stays sharp and editable. Cards are now marked as cards, priced at nothing, and if we spot shots that look like cards we ask before changing anything.
A generation that fails gives your tokens back. Nine failures in ten days were charged and never refunded: whichever part of the app noticed the failure first settled it, and only one of those parts could pay you back. Tokens already lost have been returned.
Images generated from a Chinese prompt can be downloaded again. The filename is built from your prompt, and a single Chinese character was enough to make the download answer with an error instead of a file. It worked in English, which is why it lasted.