First and Last Frame AI Video: How It Works, Which Models Support It (2026)

First and last frame AI video is image-to-video with two anchors: you upload the picture the clip must start on and the picture it must end on, and the model generates the motion in between. It is the most reliable way to control where a shot lands — for transitions, morphs, product reveals and loops — as long as both frames describe one believable shot. As of 23 October 2026, 30 SFW video models in Kenerate's studio accept an end frame, including Seedance 2.5, Kling 3.0, Veo 3.1, Wan 3 and Hailuo 02.

This guide explains how the technique works, which models support it, how to build a frame pair that actually connects, how to prompt the gap, and how to fix the usual failures.

What does "first and last frame" mean in AI video?

A normal image-to-video run gives the model one picture and a prompt. The model decides everything after frame one: where the camera goes, how the subject moves and what the final second looks like. That freedom is fine for b-roll, but it makes the ending a lottery.

First-and-last-frame generation (also called start/end frame, keyframe-to-video or frame interpolation) fixes both ends. Replicate's write-up of Veo 3.1 describes it simply: the model "interpolates between these two points based on your text prompt guidance." Your prompt no longer has to describe the whole scene — only the journey from A to B.

What it gives you:

• A guaranteed landing. The last frame is the picture you chose, not the model's guess.
• Less drift. With both ends pinned, the subject, outfit and setting have less room to wander mid-clip.
• Shots that cut together. The end frame of shot 1 can be the start frame of shot 2, so a sequence joins without a jump.

What it does not give you: control over the middle. Two matching endpoints do not guarantee a believable transition — if the frames are too far apart, the model invents something to bridge them, and that is where morphing faces and melting objects come from.

Which AI video models support a first and last frame?

Support is per model, and sometimes per version. Veo 3.1 takes an end frame; Veo 3 does not. Kling 2.6 Pro does; Kling 2.6 Standard does not. Here is the split in Kenerate's video studio as of 23 October 2026, read from the studio's own model list:

Other models with an end frame in the studio: Seedance 2.0, Seedance 2 Fast, Seedance 1 / 1.5 Pro, Kling 3.0 Pro and 4K, Kling 2.0 Master, Kling 2.6 Pro, Veo 3.1, Wan 2.2, Wan 2.7, Wan 3 Prime, MiniMax H3 and H3 Pro, Hailuo 02 Pro, LTX 2.5, FLUX 3 Video and Gemini Omni 1.1 Flash.

Two details worth knowing:

• Kling 2.6 Pro can't make sound with an end frame. That run is sent and billed silent, so add audio afterwards or pick another model if you need native sound.
• The studio shows you which is which. In image-to-video mode, a model with end-frame support shows two slots labelled Start frame (required) and End frame (optional). A model without it shows a single slot with the note "no end frame on this model".

How do you make a first and last frame AI video? (step by step)

1. Decide the one change. Write the shot as a single sentence: "the camera pushes in until the watch fills the frame", "the empty room fills with furniture", "day turns to night over the harbour". If the sentence needs "and then", it is two shots.
2. Make the start frame. Use a photo or generate one with an AI image generator. Pick the final aspect ratio now (16:9, 9:16, 1:1) — the clip follows the frame's shape.
3. Make the end frame from the start frame. Don't generate the second picture from scratch. Edit the first one (same subject, same lens, same light) so the model only has to explain what changed. An image editor that takes a reference picture is the easiest way to keep the identity and the room geometry.
4. Compare them side by side. Same face, same product details, same horizon line, same light direction? Fix contradictions before you spend a credit on video.
5. Open the studio in image-to-video mode at kenerateai.com/app/video and choose a model that shows both frame slots.
6. Drop the pictures in the Start frame and End frame slots and leave the aspect ratio matching your images so nothing gets cropped.
7. Prompt only the motion in between (see the next section), pick a short length (4–6 s) and a low resolution for the first try.
8. Review the middle, not just the ending. Scrub through frame by frame. If the landing is right but the journey morphs, change the frames or shrink the change — not the prompt length.
9. Re-run the keeper at full quality — higher resolution, sound on if the model supports it.

How should you prompt the motion between the two frames?

The frames already describe what is on screen, so the prompt only has to describe how you get from one to the other. Keep it to three parts: the subject's action, the camera, and the ending.

Good prompts look like this:

• "Slow dolly-in, the camera glides toward the wristwatch on the table, shallow depth of field, ends on a tight close-up of the dial."
• "Time-lapse: the sun sets behind the harbour, lights in the windows switch on one by one, static tripod shot."
• "The woman turns from the window toward the camera and smiles, handheld, natural light."
• "The empty living room fills with furniture piece by piece, static wide shot, soft daylight."

Things that help:

• Name the camera move. "Static camera", "slow push-in", "orbit left", "crane up". An unnamed camera tends to wander.
• Use time words for big changes. "Time-lapse", "gradually", "as the seconds pass" give the model permission to change light or season without a hard cut.
• Don't repeat what the pictures show. Re-describing the outfit or the room in detail can fight the frames.
• Match the length to the change. A small move in 4–5 s; a full transformation in 8–10 s. Most guides agree that first-and-last works best on shots a few seconds long — the further apart the frames, the harder they are to bridge.

What can you use first and last frame video for?

The loop trick deserves its own mention: give the model the same image as start and end frame and ask for motion that can come back to where it began — water, clouds, swaying lanterns, a turntable spin. The clip closes on itself and can play forever. The full method, including how to fix a visible seam, is in our seamless loop video AI guide.

Product reveals are the other big commercial use: you control the hero frame exactly, so the packaging and label look right on the last frame even when the middle moves. For full product ads, see the AI product video generator.

Is first and last frame the same as reference-to-video?

No. They solve different problems:

• First and last frame pins where the clip starts and ends. The pictures become actual frames of the video.
• Reference-to-video pins who and what appears — a character, an outfit, a product — without fixing any particular frame. The model is free to compose the shot. See reference to video AI for that workflow.

Many tools won't let you combine them in a single run; getimg.ai's guide, for example, says frames can't be mixed with reference images in the same generation. In Kenerate's studio they are separate modes too (image-to-video with frame slots, and reference-to-video with a set of reference images). A common pipeline uses both: make consistent stills with references, then animate between pairs of those stills with first-and-last-frame runs.

Why does my first and last frame video look wrong? (common failures)

A rule that solves most of these: change one thing per shot. Location or outfit or camera or lighting. Two changes at once is where the in-between frames fall apart.

How much does first and last frame AI video cost on Kenerate?

You pay the same as a normal image-to-video run on the same model — the end frame is not an extra fee on the models in the table above. As of 23 October 2026, a silent 5-second test costs 15 credits on Wan 2.2 Ultra Fast at 480p, 75 on Wan 3 at 480p, 126 on Kling 3.0 Standard and 270 on Seedance 2.5 at 480p. Sound costs more on some models (Kling 3.0 Standard goes from 126 to 189 credits for 5 s with sound; Wan 3 is the same price with or without).

The cheap way to work: test the frame pair at 480p and 4–5 seconds on a low-cost model, and only re-render the pair that works at 1080p or 4K. Video generation needs a paid plan; plans and credit packs are on the pricing page. For a broader look at every model, start at the AI video generator.

Key tips before you generate

• Make the end frame by editing the start frame, not by generating a new picture.
• Keep both frames the same aspect ratio, lens and light direction.
• Prompt the motion and the camera, not the scene.
• One change per shot, 4–10 seconds long.
• Chain shots by using the last frame of one clip as the first frame of the next.
• Check the middle of the clip frame by frame before you upscale.

Sources

• Replicate — Veo 3.1 on Replicate (first and last frame interpolation)
• Magic Hour — First-frame and last-frame AI video: shot design guide
• getimg.ai — Guide to first and last frame control
• Intangible — First and last frame
• Kenerate video studio model list and credit prices, checked 23 October 2026