First and Last Frame AI Video: How Start/End Frames and Keyframes Work (October 2026)

Quick answer: first and last frame AI video means you give the model the opening image and the closing image, and it generates the motion between them. Google calls this interpolation. Most current tools support it: Veo 3.1 and Gemini Omni Flash in Google Flow and the Gemini API, Kling 4.0, Luma Ray3.2, and several models in Runway’s API. Kling 4.0 and Luma go further with keyframes in the middle of a clip. All details below are from official docs checked on October 18, 2026.

How does first and last frame generation work?

The model treats your two images as fixed points and invents the frames between them. Your text prompt describes what should happen in between. Google’s Veo docs put it this way: “Veo 3.1 lets you create videos using interpolation, or specifying the first and last frames of the video” (Gemini API).

Compared with plain image-to-video, which fixes only the first frame, an end frame lets you control where the shot lands. Kling’s guide lists the best uses as “transformation processes, craftsmanship, and before-and-after comparisons” (Kling 4.0 guide).

Keyframes extend the same idea. Kling explains that “without keyframes, the model determines the order and timing of visual changes. With keyframes, you define the visuals at specific timestamps.”

Which tools support first and last frames?

Tool / model	Start + end frame	Middle keyframes	Clip length	Official source
Google Flow: Veo 3.1 Lite, Fast, Quality	Yes (“Frames to Video: First and last”)	No	4, 6 or 8 s	Flow models
Google Flow: Gemini Omni Flash 1.1	Yes, on Flow web only	No	4, 6, 8 or 10 s	Flow models
Gemini API: Veo 3.1	Yes: image plus lastFrame	No	4, 6 or 8 s	Veo docs
Gemini API: Gemini Omni Flash	Yes: two images in the input	No	See docs	Omni docs
Kling 4.0 (Keyframes entry)	Yes	Yes, at any position; up to 10 images	3–30 s	Kling 4.0 guide
Kling API: 3.0, 3.0 Omni, O1	Yes (2.6 and 2.5 Turbo at 1080p only)	No	3–15 s (3.0 family)	Kling API capability map
Luma Ray3.2	Through Multi-Keyframe direction	Yes, up to 16 keyframes per clip	5 or 10 s	Luma Ray3.2
Runway API	Model-dependent: veo3.1, veo3.1_fast and several Kling models accept first and last; gen4.5 accepts first only	No	Varies by model	Runway API reference

Luma’s clip length comes from its Ray3.2 pricing table, which lists 5- and 10-second text and image-to-video generations (Luma).

How do you make a first and last frame video in Google Flow?

From Google’s Flow help page:

1. In your project’s prompt box, click the model name.
2. Click Video > Frames.
3. Drag an image to “+ Add start frame” and another to “+ Add end frame”.
4. “In your text prompt, describe the action or transition that should happen between the frames.”
5. Select your video preferences (aspect ratio, outputs, model, length) and click Generate.

Google notes that “product feature availability may differ by country/region” and that “certain image types may not be supported”. Step-by-step Flow setup is in how to use Veo 3.1.

How do you set keyframes in Kling 4.0?

Kling 4.0 has two entry points, Omni Reference and Keyframes, and they “cannot be combined within a single generation task” (Kling 4.0 guide). For start and end frames, use Keyframes:

• Inputs: up to 10 images, each up to 30 MB, with aspect ratios between 2:5 and 5:2. Text prompts can run to 8,000 tokens.
• Length: 3–30 seconds in 1-second steps, default 10.
• Aspect ratio: after you upload images, “the aspect ratio follows the first keyframe”.
• Limit: “The Keyframes entry does not support uploading reference assets.”

Kling’s suggested prompt format for start and end frames is: “Use @Image1 as the first frame and @Image2 as the last frame, then describe the [process] in between.” For middle keyframes, Kling suggests tying changes to times, for example “raise a hand at 3 seconds, turn around at 7 seconds”.

How do you do it through an API?

• Gemini API (Veo 3.1): pass the starting frame as image and the ending frame as lastFrame in the config. Google says lastFrame “must be used in combination with the image parameter”. When the input includes images of people, personGeneration only allows "allow_adult" (Veo docs).
• Gemini API (Omni Flash): “provide two images in the input list and describe the desired transition in your prompt” (Omni docs).
• Runway API: in promptImage, set position to first or last. For some models, Runway says you “use position first/last for keyframe mode, or omit position for reference images. The two modes cannot be mixed” (Runway API reference).
• Kling API: first and last frames work on Kling 3.0, 3.0 Omni and O1; “End Frame Only” is listed as not supported on any model (capability map).

For a broader API comparison, see AI video generation APIs compared.

How do you get a smooth transition?

These are general practices, not results from our own testing:

• Match the basics. Use the same aspect ratio, a similar camera angle and similar lighting in both frames, unless the change is the point of the clip.
• Leave enough time. A big change, such as day to night or a full turnaround, needs a longer clip than a small one.
• Describe the path, not the frames. The images already show the start and end; the prompt should say how you get from one to the other: the action, the camera move and the pace.
• Name a single camera move. Kling’s examples ask for “one continuous lateral tracking shot from the side, with no cuts, angle changes, or scene changes” when that’s what they want.
• Use middle keyframes for multi-step changes if your tool supports them, instead of asking two frames to cover several beats.

For single-image animation basics and motion prompts, see how to turn a photo into a video with AI. To pick a model, see Veo 3.1 vs Kling vs Runway. All guides are in the AI video tools hub.