Why do the first frames of an AI-extended video clip look glitchy or broken?

The first frames of an extended clip glitch because the model regenerates overlapping frames around the join — Seedance 2.0's extend produces two overlapping frames on either side of the cut, and the first two frames of the new clip often contain error frames. Usable, reliable motion begins on frame three, so trim to there.

Treat the opening frames of an extended clip as a deliberate buffer zone, not finished footage. When you extend a clip, the model doesn't hand you a clean one-for-one start-and-end frame — it generates overlapping frames on either side of the join (two per side) so editors have alignment room. Those overlap frames are where the model is still 'kicking in': it is re-synthesizing motion, color, and subject position from a single seed frame, and the first two frames frequently come back duplicated, smeared, or slightly off-model. The reliable pick-up point is the third frame — cut your new clip in from frame three and the artifact disappears from the timeline entirely.

The deeper cause is context loss at the boundary. Frame-based extension conditions only on the last frame of your previous clip — it carries no memory of your character references, location, or camera movement beyond that single image, so the model briefly misreads motion direction and identity before settling. Color drifts for the same reason: Seedance 2.0's extend honors the lighting and color integrity of the source clip about 99% — and that remaining ~1% variance is visible when the frame is at rest, so a slight RGB curve lift and a small green-hue shift at the join is a standard correction, not an optional one.

Three practical mitigations. First, align on the overlap: use the third frame as your pick-up point and let the two overlap frames on each side act as your trim buffer. Second, hide the join with motion — a moving camera or a strong foreground action at the cut point stops the viewer scanning for artifacts; a still frame between two motion segments is the worst case for exposing the seam. Third, when you're chaining many extensions, switch from frame-based extension to Seedance 2.0 reference-to-video: it accepts character and location references alongside the full prior clip, so the model carries real context across the boundary instead of guessing from one frame. Inside invideo, the invideo agent handles this chaining — you clip the end of a segment, re-upload it, and the invideo agent attaches it to Seedance 2.0 reference-to-video with your locked references; one production built a seamless-looking 1-minute-30-second continuous shot this way, with every cut hidden.

Extending your video doesn't give you a one for one start and end frame. Very, very cleverly, it gives you two overlapping frames on either side, which is, I mean, that's genius, right?

— invideo's creative team