AI First & Last Frame Video

Pin both ends, let the model fill the middle

There's a familiar frustration in animating a single image: the motion looks great and goes somewhere you didn't want. You supplied a start, and the model decided the rest.

First/last frame inverts that: you supply both ends and the model only fills the middle. Where the picture lands is settled before you generate.

That constraint is stronger than any prompt. To take a product from closed to open, a camera from wide to tight, a scene from day to night — showing both ends beats describing the process.

The two frames need a plausible relationship

Quality mostly comes down to the ends you pick:

Two states of one scene are ideal — closed and open, wide and tight, day and night. The model reads it as one thing changing, so the bridge is smooth.

Unrelated images cause trouble — the model can only force a transition, and the middle tends to deform or jump. If you genuinely need to go from A to B with no relationship, that was always a hard cut, not an interpolation.

One frame is allowed — it then degrades to continuing forward from that image, close to image-to-video. The last frame is what you add when the landing point matters.

When to reach for a different tool

• You want the model's own idea — the video generator, from one start frame or pure text, lets it improvise.
• You want an existing clip to run longer — extend-video continues from the footage's final frames, with no end frame to prepare.
• You want a specific action transferred — motion-control maps a reference clip's movement onto your subject.
• You need mouth shapes to match speech — lip-sync.

A reasonable way to think about it

First/last frame suits the moment when you already know the answer. The start and end are in your head and all you need is the bridge — showing both ends is more precise than writing a description.

Conversely it's awkward when the direction is still open, since you'd have to produce two images first. At that stage, running a few takes in the video generator is faster.

One practical combination: make two key frames in the image generator, then bring them here to interpolate. That way both ends of the shot are pictures you chose, rather than whatever the model happened to produce.