MiniMax H3 Simple I2V · compétence Open Skill
MiniMax H3 Simple I2V
MiniMax H3 Simple I2V est une Open Skill publique sur Digen. Ouvrez-la pour voir ce que ce workflow crée.
Beauty
2 0 2
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284863/2026-10-20/vd47bb3w.jpg
---
allowed-tools:
- recognize_image
- generate_video
---
# MiniMax H3 Simple I2V
Turn one uploaded still image into a short, single-shot video that keeps the source photo's
subject, styling, and atmosphere intact while adding only gentle, believable motion.
There is no project state and no storyboard set: the uploaded image is the first frame, the
text (if any) is a hint about motion and mood, and the output is one short clip.
## When to use
- The user uploads a photo (portrait, street shot, product shot, scene) and asks for a video.
- The user says "make it move", "animate this", "bring this to life", "simple image-to-video".
- The user uploads an image plus a short description of motion/mood/camera feel.
- The user asks for a second attempt or a variation of a clip already made from the same image.
Do not use this skill for text-only video requests that need a new scene invented from
scratch, for multi-scene editing, or for cutting and joining existing clips.
## Steps
1. **Confirm the inputs.** You need at least one image in this turn (an uploaded asset id or
`s3://` uri). Read the user's text, if any, as the motion and mood direction. If no image
is present, ask for one instead of inventing a scene.
2. **Read the visual rules.**
`read_skill_file(name="minimax-h3-simple-i2v", path="references/style-and-prompt.md")`
This file defines the palette, materials, lighting, subject handling, layout, must-keep and
must-avoid rules, and the English prompt template you will fill in later.
3. **Read the motion rules.**
`read_skill_file(name="minimax-h3-simple-i2v", path="references/video-content-constraints.md")`
This file fixes shot scale, allowed camera movement, motion amplitude, duration, and
everything that must never appear in the finished clip.
4. **Inspect the source image.** Read the attached image directly. If details are unclear
(small face, busy background, text on the image, more than one uploaded image), call
`recognize_image(image_url=<the uploaded image>, prompt="Describe the main subject, face,
hair, exact clothing items and colors, pose, framing, background elements, light direction,
and color palette in detail.")` and use that description. Do not guess details that are not
visible.
5. **Write the script for this request.**
`read_skill_file(name="minimax-h3-simple-i2v", path="references/video-script.md")`
Following that structure, write a concrete single-shot script for *this* user request: logline,
duration, framing, camera move, subject micro-actions per beat, background action, light and
atmosphere, and the must-keep / must-avoid lists derived from the actual image. Do not reuse a
generic template as the finished script — every line must come from this image and this request.
6. **Build the prompt.** Compress the script into one English prompt paragraph using the template
in `references/style-and-prompt.md`, and obey the limits in
`references/video-content-constraints.md`. Keep it specific: describe the real outfit, real
hairstyle, real setting, and the exact motion you want. When you need finer checks, branches,
or parameter defaults, use
`read_skill_file(name="minimax-h3-simple-i2v", path="references/step-details.md")`.
7. **Generate the video.**
`generate_video(prompt=<final English prompt>, input_urls=[<uploaded image asset_id or s3:// uri>], model='i2v', orientation='portrait', resolution='720P', duration=5, name=<short clip title>, entity_type='final_videos')`
- `model='i2v'` is the image-to-video channel: the uploaded image is the first frame.
- `orientation` follows the source image (`portrait` for a vertical photo, `landscape` for a
horizontal one, `square` only if the source is square and the channel accepts it).
- `duration` stays short (about 5 seconds); if the user asks for longer, stay within the short
clip range rather than inventing a long sequence.
- If the user supplies two images (firs