AI Video From a Start and End Frame: Which Models Take Both, Tested (November 2026)

Live figures last changed · 7 min read · The GenMagic team

Diese Seite auf Deutsch

Most of today's video models can make a clip that opens on one picture and ends on another: Veo 3.1, Kling 3.0, Seedance 2.0 and 2.5, Wan 3.0, Hailuo 3 (MiniMax H3), FLUX 3 Video and Gemini Omni all take a start frame and an end frame. Grok Imagine, Runway Gen-4.5, HappyHorse and Hailuo 2.3 open on a picture but cannot end on one. In our test, Hailuo 3, Kling 3.0 and Veo 3.1 Lite all ended exactly on the end frame; Seedance turned down an AI-made photorealistic woman as a possible real person. Make the end frame by editing the start frame, so the person, the clothes and the light match.

This guide covers the test, which models take a start and end frame, making keyframes that work, writing the prompt, start and end frames or references, how to do it, what a clip costs, and questions.

A note on who wrote this: GenMagic runs every model named here from one balance. The clips below were made through GenMagic's API on November 11, 2026, each once, unedited; the woman in them was made with an AI image model, so nobody real is shown.

Table of contents

  1. The test
  2. Which models take a start and end frame
  3. Making keyframes that work
  4. Writing the prompt
  5. Start and end frames, or references?
  6. How to do it
  7. What a clip costs
  8. Questions

The test

Two keyframes, made with an AI image model: the start frame on a 1920s station platform, then the same picture edited into the end frame inside the train, so the woman, her coat, hat and suitcase stay the same. One prompt, sent unchanged with both pictures to four models:

One continuous shot: the woman in the red coat picks up her suitcase, turns and steps up into the train, walks into the compartment and sits down by the window, the suitcase on the seat beside her, and looks out at the platform. The camera follows her. Soft morning light, drifting steam, station sounds. No text.

ModelEnded on the end frameHow it got thereReady inCost in the test (Nov 11)Price today
Hailuo 3 (MiniMax H3), 6 sYesShe walks to the carriage door, then inside puts the suitcase down and sits: the whole action239 s$1.17$1.17
Kling 3.0 Standard, 5 s with soundYesShe walks along the platform, then a cut to a view through the window, then seated130 s$0.95no longer listed
Veo 3.1 Lite, 8 s at 720pYesShe walks towards the train, then a dissolve into the compartment; black bars at the sides (see below)58 s$0.60$0.60
Seedance 2.0 MiniRefusedTurned down the pictures as a possible photo of a real person, although she was made with an AI image model; not charged

All three that ran opened exactly on the start frame and ended exactly on the end frame; they differ in what happens between. Two lessons from the test:

  • Seedance treats a photorealistic person as a possible real person, even an AI-made one. For a photoreal character, use Hailuo, Kling, Veo or Wan; Seedance takes start and end frames of drawn, painted or 3D characters ( why Seedance refuses).
  • Match the clip's shape. Our keyframes came out 3:2, not 16:9. Kling and Hailuo kept the pictures' 3:2 shape; Veo made a 16:9 clip and filled the sides with black. Make both keyframes in the shape of the clip you want.

Which models take a start and end frame

What each video model on GenMagic takes today, with the price of a 5-second clip:

ModelStart frameEnd frame5 s clip today
Veo 3.1 Lite, Veo 3.1 Fast, Veo 3.1YesYesfrom $0.38
Kling 3.0 Standard and Pro, Kling O1YesYesfrom no longer listed
Seedance 2.0, 2.0 Fast, 2.0 Mini, 2.5YesYesfrom $0.25
Wan 3.0, Wan 3.0 Prime, Wan 2.7YesYesfrom $0.23
Hailuo 3 (MiniMax H3), Hailuo 3 MaxYesYesfrom $0.98
FLUX 3 VideoYesYes
Gemini Omni 1.1 and 1.2 FlashYesYes
Grok Imagine Video, Runway Gen-4.5, HappyHorse, Hailuo 2.3, Wan 2.6YesNo

The full, current list is in GenMagic's model catalog: each video model says whether it opens on an image and whether it ends on one, and the API returns the same as capabilities.frame_images (first_frame, last_frame).

Making keyframes that work

  • Make the end frame from the start frame. Generate the start picture, then edit it into the end picture ("the same woman, same coat and hat, now seated inside the train by the window"). Two pictures made separately differ in face, clothes and light, and the clip has to morph between them.
  • Same shape. Both pictures in the clip's aspect ratio (16:9 for a landscape clip, 9:16 for a vertical one). A square end frame on a widescreen start gets cropped or stretched.
  • A distance the clip can cover. In 5 to 8 seconds a person can walk a few steps, turn, sit down or open a door; a jump across a city or a change of season needs a longer clip or a cut.
  • The same light and lens. Morning stays morning, a wide shot stays roughly wide; a sudden change of light reads as a cut, not a movement.

Writing the prompt

The two pictures say where the clip starts and ends; the prompt says what happens in between:

  • Describe the motion, not the pictures. "She picks up her suitcase, steps into the train and sits down by the window", not a description of the platform and the compartment again.
  • Say "one continuous shot" and how the camera moves (follows her, stays still, pushes in).
  • Keep it to what fits the length: one or two actions for 5 seconds, three or four for 8 to 10.
  • Add the sound if the model makes it: steam, footsteps, a line of dialogue in quotes.

Start and end frames, or references?

Two pictures can do two different jobs:

  • Start and end frames pin the first and last moment of the clip exactly. Use them for a transition you have designed: a walk from one place to another, a reveal, a before and after.
  • References keep a character, an object or a place without fixing any frame: the model draws a new shot with them in it. Use them for a series with recurring characters ( Seedance character sheet, consistent characters in AI video).

A clip takes one or the other, not both.

How to do it

In the studio: choose Video, attach your two pictures, then click Use as start and end frames under them: the first becomes the start frame, the second the end frame. Write what happens in between and generate; the price shows before you do.

Through the API: send the start frame as image and the end frame as last_frame (https URLs or data:image URLs):

curl https://genmagic.co/api/v1/videos \
  -H "Authorization: Bearer $GENMAGIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kwaivgi/kling-v3.0-std",
    "prompt": "One continuous shot: she picks up her suitcase, steps into the train and sits down by the window. The camera follows her.",
    "image": "https://example.com/platform.jpg",
    "last_frame": "https://example.com/compartment.jpg",
    "duration": 5,
    "resolution": "720p"
  }'

A model that cannot end on a picture is refused before anything is charged, with its name in the error. Every parameter is in the video API documentation.

Longer scenes: chain clips. Under a finished clip, Continue starts the next one on its last frame, so a scene runs on across takes ( long AI videos).

What a clip costs

A start and end frame clip costs the same as any clip of that model, length and resolution. Today:

ClipPrice today
Veo 3.1 Lite, 8 s at 720p$0.60
Kling 3.0 Standard, 5 s at 720pno longer listed
Wan 3.0, 5 s at 720p$0.45
Hailuo 3, 6 s$1.17
Seedance 2.0 Mini, 5 s at 480p (drawn or 3D characters)$0.25

Every clip draws its own price from your GenMagic balance, shown before you generate; a clip that fails is not charged, top-ups start at $5 and never expire, a new account gets $0.50 of free credit once its email is verified, and nothing renews unless you choose a monthly plan. Every video model and price: the pay-as-you-go AI video generator.

Questions

Which AI video generators take a start and an end frame? Veo 3.1, Kling 3.0, Seedance 2.0 and 2.5, Wan 3.0, Hailuo 3 (MiniMax H3), FLUX 3 Video and Gemini Omni. Grok Imagine, Runway Gen-4.5 and HappyHorse open on a picture but cannot end on one.

How do I keep the same person in the start and end frame? Make the end frame by editing the start frame with an image model, describing only what changes (where they are, what they do), so the face, clothes and light stay the same.

Can I use a start frame, an end frame and reference images together? No. A clip either opens and ends on pictures or follows reference images; to keep a character from a frame, use that picture as a reference instead.

Do the two pictures need the same aspect ratio? Yes, and the same as the clip: both 16:9 for a landscape clip, both 9:16 for a vertical one.

Can I make a start and end frame video with a photo of a real person? On Wan 3.0 and Kling 3.0, yes. Seedance does not take photos of real people ( what to do instead).