LTX 2.3 Extend is a video-context Extend model. Instead of looking only at one still frame, it uses the video around the key frame as reference, so the prompt can ask for a continuation that follows the existing camera movement, timing, and scene rhythm.

Use it when the motion before the key frame matters: a camera move, a character action, a reveal, or a scene that should feel connected to the footage already on the timeline. In Kepler, the same purple Extend region supports both directions. After extends what happens next from the last frame before the region. Before generates what could have happened immediately before the first frame after the region.

This model needs much more specific prompts than a simple frame-based Extend. Describe the subject, action, camera movement, composition, lighting, and what must stay unchanged. Short prompts often produce vague or generic motion because the model has to infer too much from the context video.

Kepler renders up to 20 seconds of timeline video context from the same reference clip shown in the Extend preview. The model then uses the frame-limited part of that rendered context automatically. Extension duration is capped at 20 seconds. Advanced controls expose output resolution up to 2160p, optional manual seed, and Auto Improve Prompt. Auto Improve Prompt is off by default so users can keep tighter control over continuity.

Generated output is stored as new media and placed above the source at the purple region. The original source video is never overwritten.