How do you frame chain AI shots from first frame to last frame?

Frame chaining AI shots is a four-step loop: plan the key frames, generate a clip between two of them, extract the real last frame, and start the next clip on that frame. Every seam then shares one real frame, so the sequence plays as one continuous move with no cut.

Plan the sequence as a set of key frames first: a still for the start of the sequence, and a still for each beat it should reach. The worked example below planned 5 key frames as one composited photo-wall image, so the "next" frame already existed inside the same still, cropped differently.

Generate a video clip between two key frames: the first key frame as the video model's first-frame input, the next key frame as its last-frame input. The model fills in continuous camera motion between them.

Extract the true last frame of the rendered clip. This is the frame that seeds the next clip, and it is not always identical to the key frame you designed. In this sequence, one shot's real hand-off frame drifted enough from the planned still that the next shot was written to match the real frame instead of the plan.

Feed that real last frame into the next generation as its first-frame input, and repeat down the sequence. Check each seam by eye, at full speed and frame by frame, before moving on.

A frames-first pipeline builds each AI video shot from its start and end frames: the stills are designed first, and the video model fills in the motion between them. Designing both ends of the move gives control over shot size and composition, for example a wide establishing frame that ends on a close-up. Chained shot to shot, the last frame of one clip becomes the first frame of the next, so a multi-shot scene plays with no visible cut.

The worked example on this page runs that pipeline on a real production: the step-by-step loop, three shot-size transitions, the retake when an anchor is ignored, and what the sequence cost.

To change shot size inside one AI video move, give the video model a first frame at one shot size and a last frame of the same scene at another, such as a wide and a medium. The model authors the dolly or push-in between the two frames. Three worked examples follow: wide to medium, medium to close-up, and a big jump chained through a medium.

Wide to medium. First frame: a wide of a lighthouse keeper on the gallery rail at dusk, the whole tower in view. Last frame: a medium of the same keeper, waist up, rail in the foreground. Prompt: "Slow dolly-in from the wide to a waist-up medium of the keeper, dusk light unchanged, one continuous move, no cut."

Medium to close-up. First frame: that medium. Last frame: a close-up of her face as the beam sweeps past. Prompt: "Gentle push-in from the medium to a close-up of her eyes, the beam crossing her face at the end, no cut." Name the ending framing in words as well as in the still, because the words are what pulls a drifting clip back onto its target.

Push-in across a big jump. Wide to extreme close-up in one clip usually breaks: the model invents a cut or lands on the wrong detail. Chain it through a medium instead, two clips that share one frame, so each move only covers one step of shot size.

For every pair, keep the light, wardrobe and lens feel identical in both stills, and keep the subject near the same point in the frame. The shot size is the only thing that should change.

When a video model ignores the last-frame anchor, keep the same two frames and retake the shot with a prompt that describes the last frame in words as well as in the image. The words pull a drifting clip back toward its target. Budget one retake for every few chained shots, because the anchor does not always hold.

The failure mode worth planning for is not a bad clip, it is a clip that moves well but lands on the wrong frame. Twice in the sequence below, the video model was given a specific last-frame image and produced a continuous, good-looking move that nonetheless ended on a different composition than the one it was told to end on.

The fix both times was the same: keep the same two anchor frames, but make the prompt spell out exactly what the last frame contains, in words, in addition to the image. One shot then matched on the retake. The other matched the same subject but in a different framing, a documented, honest seam rather than a hidden one.

Budget for this. A clean production run assumes every anchor holds; a real one holds a retake in reserve for at least one seam per few shots.

A frame-chained AI sequence in FXI Studio costs about 25 credits per 5-second chained shot with the fast transition model, plus a retake in reserve for every few shots. A four-shot worked example finished for 90 credits, both retakes included.

The four chained video generations kept in this sequence's final cut ran 90 credits total, at 25 credits for a 5-second clip with the fast transition model used here, prorated for two shorter clips. That total includes both retakes.

The key frames cost nothing extra: the stills were generated on FXI's in-house fast still model, which runs at no credit cost on paid plans. The only line item is the chained video itself.