Upload a video, scrub to the moments you want to change, and capture each as a keyframe. Write a short prompt for each keyframe and the AI generates a still image that previews the edit. When the references look right, click Render — the AI applies those changes to the matching moments in the video and produces a new clip.

Editing a single frame is faster and cheaper than rendering a whole video. The per-keyframe reference lets you confirm the visual direction — colour grade, object placement, costume change, lighting — before spending the larger render budget on the final clip.

You can upload longer videos, but only the first 15 seconds are editable and rendered in this workflow. Anything past the 15s mark is ignored at render time.

Per-keyframe image previews use Seedream 5.0 Lite Edit. The final video render uses ByteDance Seedance 2.0 Edit (higher fidelity) or Seedance 2.0 Fast Edit (cheaper, quicker). You can switch models from the settings panel before rendering.

Localised changes — outfit swaps, object replacement, colour grading, relighting, background replacement, weather changes — all work well because they can be anchored to a moment and a reference image. Very broad rewrites (e.g. changing the entire genre of footage) are less reliable.

Yes. Every render lands in the versions strip alongside the original source. Click any version to make it the working video, tweak prompts or keyframes, then render again to produce a new version on top of that one.