LTX 2 Retake regenerates a selected moment inside an existing video shot while keeping the surrounding source video as context.
Use Retake to change what happens inside a shot: an action, reaction, movement, spoken line, or sound. It is designed to redirect the scene while preserving its visual context. It is not a reliable object-replacement tool; use a dedicated editing workflow when the main goal is to replace one visible object with another.
In Kepler, place the purple region inside one video clip and choose whether Retake should replace Video, Audio, or Video + Audio. The selected source span must resolve to 2-20 seconds after timeline speed is applied.
Write the prompt as a precise direction for the changed moment. State what should change and how it should happen, then name any important details that must stay unchanged and the intended camera behavior. Short prompts can be interpreted too conservatively.
Example:
Same portrait shot of the same cat in the same room and lighting. During the selected segment, the cat suddenly jumps upward with all four paws leaving the surface, then lands back in the same place. Keep the camera static. Preserve the cat's appearance and the background. Natural-speed motion.
For predictable Retake input, conform source video to one of the published LTX 2 model frame rates: 25 or 50 fps. The provider does not publish a separate input-FPS whitelist for Retake. Until the provider documents broader input support, conform other source rates, including approximately 29.97, 30, or 60 fps, before Retake. The LTX API documents MP4, MOV, and MKV video with H.264 or H.265, a 4K maximum input resolution, and at least 73 input frames. Kepler sends the original source video with the source-relative Retake span, stores the returned video as new media, and places an aligned variant above the original clip. The original source is never overwritten.