LARUNDStart free
← All articles

GUIDE · 2026

How to use start and end frames in AI video

Plain text-to-video is a lottery: you describe a shot and hope. Giving the model a first frame, and ideally a last frame, turns it into a controlled move from A to B.

This guide explains what start and end frames do, how to create good ones and when the technique helps or hurts. Larund, the studio we build, is used for the examples, but the principles apply wherever the feature exists.

What start and end frames do

A start frame (also called a first frame) is an image the video begins from. An end frame (last frame) is an image the video should finish on. The model then invents everything in between: the movement of the subject, the camera and the light.

With only a start frame, the model controls the ending. With both, you control the beginning and the ending and the model only controls the path. That is why keyframes make shots both more predictable and easier to join together.

Support for end frames differs between video models, so it is worth testing before building a workflow around it. Larund’s video pipeline is keyframe-first: each shot starts from a start and an end keyframe you approve.

When keyframes help most

Keyframes are most useful when the outcome matters more than surprise.

  • Character consistency: the same face and outfit appear in the first and last frame of every shot.
  • Product shots: the product must be accurate at the start and at the end.
  • Transitions: one shot ends on a composition that the next shot starts from.
  • Storyboards: each frame is a planned panel that you animate rather than a hopeful description.
  • Cost control: a wrong still is cheaper to redo than a wrong clip.

How to make good keyframes

Good keyframes are made like good photographs: decide the subject, the position in the frame and the light, then make both images match in everything except what should change.

Keep the world identical between the two frames, meaning the same location, light direction, outfit and lens, and change only what the shot is about, such as the subject’s position or expression. If the two images disagree about the world, the model blends them into a warped morph.

  1. Write the shot in one sentence: what changes between the start and the end.
  2. Generate the start frame from your character and location references.
  3. Generate the end frame from the same references, changing only the subject’s pose or position or the camera framing.
  4. Compare the two side by side: same outfit, same light, same lens feel?
  5. Approve the pair, or regenerate whichever frame is out of line.

Keep the change between frames reasonable

The bigger the difference between the first and the last frame, the more the model has to guess, and the more likely it is to cheat with a dissolve or a distorted in-between. A person turning their head, taking a step or lifting an object is a good change. A person moving from one end of a room to the other, in a different outfit, from a different angle, is not.

If a move is large, split it. Add a middle keyframe so that the action becomes two shots of moderate change rather than one wild one. Short shots of a few seconds are easier to keep clean in any case.

Chaining shots with keyframes

The end frame of one shot can be the start frame of the next. This is the simplest way to build continuity: the film literally continues from the final image of the previous clip.

In Larund’s Cinema Studio scenes are chained in this way, with each scene rendered and the next picking up from it. For hard cuts instead of seamless continuity, give the next shot a new start frame with a different angle but the same location and cast references.

Prompting with keyframes

Once the frames fix the look, the prompt only needs to describe the path: the action, the camera and the pacing. Do not re-describe the subject in detail; contradictions between prompt and image invite the model to redraw. Something like “slowly turns toward the window, camera pushes in gently” is enough.

Lens and angle tokens still help to keep the movement faithful to the framing. In Larund, tokens such as &35mm and &low-angle can be added to the prompt to pin the shot.

SituationFrames to supplyPrompt focus
Simple actionStart and endAction and pace
Camera move onlyStart (end optional)Camera movement
Product revealStart and end of the productSlow move, light
Long actionStart, middle, end as separate shotsOne action per shot

Troubleshooting

If the clip warps, check the two frames for disagreement in lighting, outfit or lens and make them match. If the motion is stiff or barely there, the frames may be too similar, so increase the change slightly. If the model ignores the end frame, try a different model, since support and quality vary.

Because Larund shows the credit price before each generation and has 25+ video models on one balance, testing the same pair of keyframes on two models is a cheap experiment.

Is Larund the right pick?

A good fit when

  • You want to see and approve exactly how a shot starts and ends before motion is rendered.
  • You are chaining several shots and need each to land where the next begins.
  • You want to compare the same keyframes on different video models.

Probably not the best fit when

  • You want quick, loose, exploratory clips and do not mind surprises.
  • The shot has unpredictable motion, such as water or smoke, where an end frame adds little.
  • You already work in a tool whose keyframe controls you know well and have no reason to switch.

Frequently asked questions

What is a start and end frame in AI video?

They are two images that mark how a shot begins and how it ends. The AI video model generates the motion between them.

Do all AI video generators support end frames?

No. Support varies by model, so check before relying on it. Larund’s workflow is built around start and end keyframes that you approve before motion is rendered.

Why does my AI video warp between the first and last frame?

Usually the two frames disagree about the world, such as light, outfit or lens, or the change between them is too large. Make them match and split big moves into two shots.

Can I use the end frame of one clip as the start of the next?

Yes, and it is the simplest way to build continuity. Larund chains scenes this way in Cinema Studio.

Are keyframes cheaper than rendering video directly?

Stills cost less than clips, so checking looks on keyframes first saves credits on shots that would have been wrong. Larund shows the credit price before each generation.

Should I describe the subject again in the prompt when I use keyframes?

Keep it minimal. The images already define the look, so the prompt should focus on the action and camera movement.

Every top model. One studio.

Start free with Larund Studio: 25+ video and 30+ image models, keyframe films and motion graphics, with the price shown before every generation.

Try Larund free