How does first and last frame AI video work?

What keyframe interpolation means

In animation, a keyframe is a drawing that fixes an important moment, and the frames between keyframes are called in-betweens. First-and-last-frame video applies the same idea: you draw or generate the two keyframes, and the model makes the in-betweens.

Classic frame interpolation, such as the FILM method from Google Research, fills frames between two photos taken a moment apart. It estimates how pixels moved and blends along that motion. It is good for smoothing slow motion, but it cannot invent new content, so large changes look like a morph.

Generative interpolation goes further. A video diffusion model is conditioned on both images and generates a full clip whose first frame matches one and whose last frame matches the other. The 2024 paper Generative Inbetweening adapted an image-to-video model for this: it runs one pass forward from the first frame and one backward from the last, and merges them. Products such as Veo, Kling, Runway and Seedance offer the same idea as a start frame and end frame option, and Google's Veo documentation describes it as specifying the first and last frames plus a prompt.

When to use it

• You need the clip to end on an exact pose, product angle or composition, for example a logo reveal or a product turning to face the camera.
• You are joining shots: the last frame of shot A becomes the first frame of shot B, so the cut is invisible.
• You want a controlled transformation, such as day to night, closed flower to open flower, empty plate to served dish.
• You want a camera move with a fixed destination, such as a wide shot that ends on a close-up of a door.

If you only care how the clip starts, plain image-to-video is simpler and gives the model more freedom.

How to design the two frames

The model has to find a believable path from frame A to frame B in a few seconds. Your job is to make that path short and obvious.

1. Make both frames from the same source. Generate frame A, then make frame B by editing frame A (change the pose, expression or light) rather than generating a new image from scratch. Same character, same clothes, same background, same lens.
2. Change one main thing. One pose change, or one camera change, or one lighting change. Two changes at once often produce a blend that is neither.
3. Keep the size of the change reasonable for the length. A head turn fits in a few seconds; walking across a room and sitting down needs a longer clip or two shots.
4. Keep the same aspect ratio and resolution for both images. Google's documentation, for example, uses 16:9 or 9:16 for both.
5. Check that the path is physically possible. If the left hand is raised in frame A and in a pocket in frame B, the hand must travel somewhere the camera can see.
6. Write a prompt that describes the change. Google's example pairs a first frame of cookies with a last frame of cookies and milk, and the prompt describes the action that connects them.

First frame: a woman at a desk, looking down at a laptop, office at dusk.
Last frame: the same woman, same desk, now looking up at the camera, smiling.

Prompt:
She slowly lifts her head from the laptop, notices the camera, and smiles.
Camera static, medium shot. Soft warm desk lamp light. Audio: quiet keyboard
clicks, then silence.

Transitions that work and transitions that fail

When you want a deliberate morph, such as a seed turning into a tree or one style turning into another, a failed path can become the effect. Say so in the prompt: 'smooth magical transformation' tells the model a morph is intended, and it will usually look cleaner than an accidental one.

Chaining clips with shared frames

First-and-last-frame tools are also a continuity tool for longer sequences. Plan a list of keyframes, then generate each clip between two neighbors.

Because clip 2 starts on the exact frame that clip 1 ends on, the joins are seamless. Colors and sharpness can still shift slightly at each join, so color-match the clips in the edit, and keep the motion direction consistent so the cuts feel natural.

Limits and fixes

• Ease-in and stall. Models sometimes hold the first frame, then rush to the last. Describe the pace: 'moves steadily over the whole clip'.
• Hidden cross-fade. If the frames are too different, the model may dissolve between them. Bring the two frames closer together, or add a middle keyframe.
• Identity drift in the middle. The ends match, but the face changes halfway. Use closer framing, a shorter clip, and frames made from the same source image.
• Last frame not exact. Some tools treat the last frame as a strong hint rather than a hard rule. Trim or hold the final frame in the edit if needed.
• Text and logos in motion. Keep them still, or add them in the edit.