MiniMax H3 First and Last Frame: How to Write the Transition Between Two Images

How first and last frame works

In first and last frame mode, the video starts exactly on your first image and ends exactly on your last image; the model invents everything in between. It is the most precise way to control where a shot ends: a product landing in the hand, a character arriving at a mark, a door closing on a specific frame. MiniMax's earlier Hailuo 02 model offered start and end frames too, and MiniMax H3 supports them natively.

On XXT it is image-to-video (i2v) with a --last-frame, 4–60 seconds. The platform writes the official alignment sentences that pin the first image to the start and the last image to the final frame, so you only describe what happens in between.

Choose two images that can connect

• Same scene, same person, same light. The model moves between them; it doesn't teleport.
• Keep them close. If the two images are far apart, in pose, position or camera angle, the motion tends to stall in the middle and then rush to match the last frame at the end.
• Same aspect ratio as the video: 9:16, 16:9 or 1:1. An image in another ratio is scaled and center-cropped, which can cut off the subject.
• Short side ideally 720 px or more.
• A good way to get a matching last frame: edit the first image (move the hand, open the umbrella, turn the head) instead of generating a new one from scratch.

The command

h3 submit --mode i2v --first-frame start.png --last-frame end.png --aspect 9:16 --seconds 8 \
  --prompt "A medium-wide shot frames the rain-soaked cyclist beside a silver bicycle, holding a closed black umbrella. In one continuous shot, the camera pulls out with small amplitude at slow speed as she releases the handlebar, raises the umbrella and pushes it open, steps beneath it, and settles into the pose and composition of the final frame at the end of the shot." \
  --sound "Steady rain on the pavement, the click of the umbrella opening, distant traffic." \
  --wait --out ./out

This follows the official MiniMax H3 first and last frame example: one sentence anchoring the first image, then the action path, then the landing on the last image.

Write the path, not the two pictures

• One continuous shot. Write in one continuous shot; a cut in the middle defeats the point of a fixed ending.
• Describe the action that connects the images, in order: who starts where, what they do, how it develops. Don't describe the first image and then the last image as two stops, or you get a slideshow feel or a dissolve.
• End on the last frame: settles into the pose and composition of the final frame at the end of the shot.
• Camera = movement + amplitude + speed, such as pulls out with small amplitude at slow speed. If the two images are framed differently, this sentence is what explains the change.
• Write in English, 60–150 words. Only write what you want to see; there is no negative prompt.

The general rules are in the MiniMax H3 prompt guide.

Seamless loops: the same image twice

Use the same image as first and last frame to make a loop. If you just describe the scene, the picture barely moves, because the model is told to start and end in the same place. Describe a closed-loop action instead: something happens in the first part, peaks in the middle and returns to the starting pose at the end, such as a dancer spinning once and landing back in her opening pose, or a cat stretching and curling up again.

Keep the face: add reference photos or a character

Between two frames a person turns, moves and changes size, and without more information the face can drift. On XXT you can add reference photos (--ref front=…) or a saved character (--character) to a first and last frame job, up to 7 images (8 without a last frame: the last frame takes one image slot). MiniMax's hosted API documents first and last frame and reference images as separate modes; here they combine in one job.

h3 submit --mode i2v --first-frame door_closed.png --last-frame door_open.png --aspect 16:9 --seconds 6 \
  --character Mia=front,three_quarter_left \
  --prompt "A medium shot frames Mia standing in a dim hallway in front of a closed white door, her hand on the handle. In one continuous shot, the camera holds a static shot as she turns the handle, pulls the door open toward her and steps half into the warm light from the room, and settles into the pose and composition of the final frame at the end of the shot." \
  --sound "A quiet hallway, the click of the door latch, a faint creak of hinges." \
  --wait --out ./out

How to prepare the photos and pick angles: character consistency.

Chain shots into a longer sequence

The last frame of one shot can be the first frame of the next. Chaining first and last frame shots this way gives you control over every join, which is the most reliable way past the 60-second limit of a single clip. See longer videos.

When it goes wrong

What it costs

A last frame costs nothing extra. Video is billed per second: $0.009 at night (22:00–10:00 Singapore Time) and $0.0135 during the day, so an 8-second 768P shot submitted at night is 48 credits ($0.072). 1080P costs double.

Videos, songs and uploaded files are kept for 8 hours after the job ends, then deleted automatically. Please download them in time.

FAQ