What is first and last frame in AI video?

First and last frame works like image to video with a second anchor. The model places your first picture at the start of the clip and your last picture at the end, marks both as frames to keep, and generates the frames between them from noise until they join the two up in a way the prompt describes.

In text to video the model invents everything; with one picture, everything after the opening. With two it has a destination, so its job is finding a believable route: the mode suits a planned change and struggles when no such route exists.

Under the hood: two frames held, the middle generated

The report on the open Wan model describes the usual design. The picture is compressed and placed as the first frame, with empty frames after it, and a binary mask marks which frames are kept and which are generated. An image encoder reads the picture’s content as well, so its look carries into frames it never showed. The report says the same design extends to first-last frame transformation and to video continuation.

The name FLF2V comes from that family: the Wan team released Wan2.1 FLF2V, first-last-frame-to-video, on April 17, 2025, at 720p only.

Research calls the task keyframe interpolation or inbetweening, after the animator’s job of drawing the frames between two key poses. One ICLR 2025 paper takes an image-to-video model trained to move forward in time from one picture, tunes a copy to move backward, and runs the two together so each step blends a guess from each end.

How to write a first-and-last-frame prompt

Describe the change, not the pictures: both are already fixed. Say what moves, in what order, and how the camera travels from the first framing to the last. Kling’s guide for its O1 model gives the pattern: take the first image as the start frame, the second as the end frame, then describe the changes between them.

• Name one action that turns the first picture into the last: “she turns from the window to face the camera”.
• Name the camera move that gets from one framing to the other: “slow push in from wide to close”.
• Say how the clip arrives: “settles”, “comes to rest”, “the light finishes rising”, so the last second isn’t a snap.
• Leave out what both pictures already show: the face, the clothes, the room.

A picture can steer a video in several ways. Which you choose decides how much of the shot is yours and how much the model makes.

First and last frame is the only one of these that fixes where a shot finishes; a longer sequence is several such shots, cut together.

Most failures start before the prompt, with two pictures that can’t be joined in the time. Kling’s start and end frames guide asks for two similar images on the same theme, and warns that large differences may trigger a shot switch: a cut instead of a move.

• Give both pictures the same shape and size as the clip, so nothing is cropped or stretched at either end.
• Keep the subject, the light, the lens and the style the same; change one or two things, such as a pose, a camera position or the time of day.
• Make the change fit the length: a head turn fits in 5 seconds, a sunrise over a city does not.
• Make the last frame from the first: edit a copy with the AI image editor, or grow it to a wider framing with the AI image expander, so both share every detail you didn’t change.
• Keep faces, hands and text large and clear in both: the middle frames are invented, and small details are redrawn first.

Each maker sets its own limits. What their pages said when we read them on October 29, 2026:

• Transitions: the same street by day and by night, a room empty and furnished, summer to snow.
• Before and after: a makeover, a renovation, a dish raw and cooked, both ends exactly as photographed or designed.
• Reveals that land on a planned final image: a product, a title card or a logo.
• A product turning from its front to its back, with both views exact.
• A character moving between two poses drawn for a storyboard.
• Loops, with the same picture at both ends.

On DreamBench, Seedance 2.5, Seedance 2.0, Veo 3.1 with its Fast and Lite versions, Kling 4.0, Kling 3.0, Wan 3.0, Wan 2.7, MiniMax H3 and LTX-2.3 can be told where to end. HappyHorse 1.1, Grok Imagine 1.5 and Vidu Q2 start from a first frame only. The table is read from the same settings the DreamBench composer uses:

“First frame only” means the model can open on your picture but can’t be told where to finish: the ending is left to the prompt.

Why doesn’t my video end on my last frame? The first frame is placed exactly, but the last is a target: the model generates the final frames to meet it, and a path that runs late, or a change too big for the seconds, lands near it or snaps into it in the last moments. The closer the two pictures, the closer the landing.

• Middle frames can still flicker or drift: a failure of temporal consistency, worse the further apart the ends.
• Two ends fix one shot. A person who must look the same across several shots needs the same pictures in each, which is character consistency.
• Some models refuse some pictures: ByteDance’s tutorial says Seedance 2.5 takes no images showing real human faces.
• Only use pictures you have the right to use, and ask before you animate a photo of someone else.

The logo animation tool is DreamBench’s first-and-last-frame run. You upload a logo and pick a style, and the model is given your logo, as you uploaded it, on its own backdrop, as the frame to end on. The first frame depends on the style: the logo itself for Light sweep, the empty backdrop for a reveal, or a frame drawn from the logo, broken up for Glitch, a dim outline for Neon, edge-on for 3D spin.

Each run is a five-second clip on Kling 3.0: 720p and silent, 1080p, or 1080p with sound. Some runs don’t land cleanly on the logo frame; if the last second looks wrong, run it again.

The image-to-video tool is the one-picture version: a photo becomes the first frame of a five-second clip. And when you describe a film, DreamBench writes it as shots, draws an opening frame for each from your reference pictures, and the video model generates the shot from it; where it goes is written in its prompt.