Start and End Frames in AI Video: Which Platforms Support What

A start frame and an end frame are the most direct control you have over an AI video shot. You supply the image the clip opens on and the image it closes on, and the model invents the journey between them. It is the right tool for a product that turns to show its label, a day-to-night change on the same framing, a character who walks from the door to the window, or a transition that has to land exactly on the next shot. Platforms call it different things (Start/End Frame, Frames to Video, keyframes, first-and-last-frame), but the idea is the same. This guide sets out who supports what, then explains how to make frame pairs that actually work.

The platforms at a glance

Exact labels and limits change often. If a mode disappears from a menu, look for the function described here rather than the name.

Two frames or many keyframes

Most platforms give you two anchors: where the shot begins and where it ends. That covers the majority of briefs, because most shots have one clear movement. The model decides everything in between, which is both the appeal and the risk. If the two frames are far apart in content, it has to invent a lot, and that is where morphing and odd shortcuts appear.

Luma Ray3.2 is the exception. It accepts up to 16 keyframes within one clip, so you can pin down the middle of a move as well as its ends. That is useful for a slow orbit that must pass a specific angle, or a choreographed product reveal. The method is explained in Luma Keyframes, Loops and Extend. Pika’s equivalent is covered in Pika Prompting and Pikaframes.

Frames plus references and sound

Start and end frames fix composition, but not necessarily identity. A face seen from behind in the first frame and in profile in the last can still drift in the middle. Some platforms now let you combine frames with other controls. Seedance 2.0 takes up to nine reference images, three reference videos and three audio files, and you tell it which image is the first frame. Luma’s Ray3 Modify lets you add a character reference alongside start and end frames, which helps when the face is not clear in either. In Flow, Google brought audio to Frames to Video, so a transition can carry sound without a separate pass. For keeping a character steady across several such shots, see Kling Elements and Flow Ingredients and Scenebuilder.

How to make frame pairs that work

The quality of the result depends far more on the two images than on the platform. A few rules hold everywhere.

• Make the end frame from the start frame. Edit a copy of the first image rather than generating the second from scratch. Instruction editors such as Flux Kontext or Gemini image editing are ideal for this, because they change one thing and keep the rest.
• Keep the lens and the light. Same focal length, same lighting direction, same colour temperature. Change these and the model has to reconcile two different photographs.
• Match the aspect ratio and resolution. Mismatched frames are cropped or stretched, and the motion suffers.
• Describe the movement, not the frames. The images already show the start and end, so the prompt should say how you get from one to the other: 'the cup rotates slowly clockwise', 'the camera pushes in'.
• Keep the change small. One clear movement per clip. For a bigger journey, chain several clips, using the end frame of one as the start frame of the next.

Chaining clips means tracking which frame came from which generation. Keeping every image and clip with its prompt and settings, as Dreamdrive does, saves you hunting for the exact end frame you used three shots ago.

Try it

Generate or photograph a product on a plain surface, label facing away. Then edit a copy so the label faces camera, with nothing else changed. Use the pair as start and end frames on two platforms with this prompt.

The bottle rotates slowly on the spot to reveal its label, soft studio light, static camera, plain background, no other movement

Compare three things: did the bottle keep its shape through the turn, did the label arrive legible and undistorted, and did anything else in the frame move? Then try the same pair on Luma with a third keyframe at the halfway point, and see how much steadier the middle of the turn becomes. Write down the model and version you used, because these modes change with almost every release.