Introduction
MiniMax H3 Video extend continues a clip you already have. Send it as video_url with a prompt for what happens next. The final frame of your clip becomes the opening frame of a new MiniMax H3 image-to-video shot of 4 to 15 seconds, and the two are joined into one MP4: your original first, then the continuation, with the repeated seam frame removed. The new segment also comes back as its own file, so you can keep extending or edit the pieces separately.
Key features & highlights
Continuity comes from the last frame
Wardrobe, set, light and where everyone stands carry over because the new segment literally starts on your final frame. Motion does not carry over the same way: the model sees one still, not the whole clip, so describe the movement you want to continue instead of assuming it will keep its speed or direction.
One file back, plus the new piece on its own
The joined MP4 keeps your clip’s frame rate and its framing, with no black bars. To line the seam up, the picture may be trimmed or adjusted by a few percent at the edges. If the model returns a noticeably different shape, the new segment is scaled to fill your frame and centered, cropping what spills over. The new segment comes with sound generated in the same pass; sound from your source stays, and the seam is blended over a fraction of a second. If one side is silent, the other is padded with silence. The second asset in the output is the new segment alone, with the role segment, ready for another extend or for an editor.
An optional closing frame
last_image_url is optional. Send an image and the new segment comes to rest on it; leave it out and the ending is free. The closing frame tightens the ending only, and the longer the segment, the further the middle can drift from it. A 4-second segment is only available at 768p, and 2k needs at least 5 seconds.
Everything else is MiniMax H3 image-to-video
The new segment runs on the same routes, with the same content range and the same per-second price as the image-to-video endpoint, and a failed generation is refunded the same way. If the generation succeeds but the join itself cannot be built, the task fails and the charge is refunded.
Request parameters
| Parameter | Required | Description |
|---|---|---|
prompt | Yes | Required. What happens next: the action that follows, the camera move and the sound. at least 1 characters · at most 7000 characters |
video_url | Yes | The clip to continue: MP4, MOV or WebM, up to 60 seconds. Its final frame becomes the opening frame of the new segment, so cut it on a clean, sharp frame. at least 1 characters |
last_image_url | No | Optional target end frame: the new segment closes on this image. It shapes the ending only; the middle is still free. |
resolution | No | Resolution of the new segment: 768p or 2k, default 768p. Both are native renders. |
duration_seconds | No | Length of the new segment: whole seconds, 4 to 15; defaults to 5. A 4-second segment is only available at 768p, and 2k needs at least 5 seconds. |
Pricing
Billed per second of the new segment at the resolution you choose; the source clip you send is not billed. A failed generation, or a join that cannot be built, is refunded in full.
| Specification | Price |
|---|---|
| $0.037 / second |
| $0.058 / second |
Use cases
Lengthen a shot that ended too early
A product turntable that stops before the label faces camera, or a walk that cuts before the subject reaches the door. Extend it by a few seconds and describe only the remaining motion.
Build a longer sequence step by step
One request produces at most 15 seconds. To go further, generate a first shot, extend it, then extend the joined result again; each step bills only the seconds it adds. Because every step starts from a single frame, watch each seam before stacking the next one.
Prompt guide & pro tips
Describe the next beat, not the whole scene
The frame already fixes who and where, so spend the prompt on the action that follows, the camera move and the sound. Repeat the camera language your clip ended on, such as a slow push in or a locked shot, so the seam reads as one take.
Trim before you extend
The last frame is the only thing the model sees, so a blurred, mid-cut or heavily compressed ending carries straight into the continuation. Cut your source on a clean, sharp frame first. Sources of up to 60 seconds are accepted as MP4, MOV or WebM.
Keep the resolution consistent
The resolution only decides how sharp the new segment is generated, so a lower setting makes the new part look softer than the original. Match the resolution of your source, and set duration_seconds to the length you actually need: the charge follows the new segment only.
API quick start
# Generate the idempotency key once; reuse it when retrying after a lost response.
export SPICY_IDEMPOTENCY_KEY="${SPICY_IDEMPOTENCY_KEY:-$(uuidgen)}"
curl -X POST https://api.spicyapi.ai/api/v1/jobs/createTask \
-H "Authorization: Bearer $SPICY_API_KEY" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: $SPICY_IDEMPOTENCY_KEY" \
-d '{
"model": "minimax/h3/video-extend",
"input": {
"duration_seconds": 5,
"enable_prompt_expansion": false,
"prompt": "A slow aerial shot drifting over a misty pine forest at sunrise",
"resolution": "768p",
"video_url": "https://example.com/reference.mp4"
}
}'
# Copy data.taskId from the response into TASK_ID, then poll the status.
TASK_ID="TASK_ID_FROM_RESPONSE"
curl "https://api.spicyapi.ai/api/v1/jobs/recordInfo?taskId=$TASK_ID" \
-H "Authorization: Bearer $SPICY_API_KEY"
See the API reference for authentication, the complete schema, polling, and error handling.
