SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsmeitu-video-dance

meitu-video-dance

Design & creation

View repository

Updated 10-16 06:46

About this skill

Overview

Extract motion trajectories from a reference video, transfer them onto the target person/character image provided by the user, and generate a video of that character performing the same actions. The core tool is meitu video-motion-transfer(asynchronous task).

Before execution, the user should be clearly informed that: this Skill will read Meitu credentials, invoke the local meitu CLI, send the user-provided person image, reference action video, and related prompts to the Meitu OpenAPI for processing, and write the results to ./output/ or $VISUAL/output/meitu-video-dance/. When portrait and action materials are involved, confirm that you have authorization to process the people in the materials.

Dependencies

  • meitu-cli (>=2.0.6): npm install -g meitu-cli@latest
    • Preferred environment variables:MEITU_OPENAPI_ACCESS_KEY / MEITU_OPENAPI_SECRET_KEY
    • or preset credential file:~/.meitu/credentials.json
    • If manual initialization of local credentials is needed, you may explicitly run meitu config set-ak --value "..." + meitu config set-sk --value "..."(will write to a local file)
    • Verify: meitu auth verify --json

Path aliases: Below, $VISUAL = {OPENCLAW_HOME}/workspace/visual/

Core Workflow

Preflight → [Context] → Execute → Refine → Deliver → [Record]

Preflight

  1. meitu --version → if not installed, prompt npm install -g meitu-cli@latest
  2. meitu auth verify --json → if credentials are invalid, guide configuration
  3. Detect mode: cwd has openclaw.yaml → project mode; else → one-off Check $VISUAL directory → determine capabilities can_record = cwd has openclaw.yaml AND $VISUAL exists (either missing means false)
  4. output_dir resolution (MUST be completed within Preflight): Resolve output_dir: openclaw.yaml → ./output/ | else → $VISUAL/output/meitu-video-dance/ mkdir -p {output_dir}

Context (executed in project mode, skipped in one-off mode)

mode = one-off → skip this step, go directly to Execute.

Read step by step:

  1. Read DESIGN.md → extract Context References (quality, preferences, brand_refs)
  2. Try to read global assets: quality.yaml → global.md → scenes/{type}.md (scene file source: read the type from openclaw.yaml's project.types(array takes priority) or project.type )
  3. If unreadable, use inline fallback values (all skip if missing) → quality forbidden list filters prompt elements, preferences enhance the creative direction

Execute

This is the core of the skill. The quality of action transfer is 80% determined by the input materials, 15% by the prompt, and 5% by the parameters.

1. Collect and validate inputs

The user must provide two inputs:

Input Description Accepted formats
Target person image (image_url) The character to "bring to life" URL or local path (.jpg/.jpeg/.png)
Reference action video (video_url) Action source URL or local path (.mp4/.mov)

Input quality precheck — proactively assess before calling the API to avoid wasting quota:

Check the target image for:

  • Full body visible (at least down to the knees; dance types must include the feet) → if not satisfied, remind the user to change the image
  • The person's outline is clear and not blended with the background → if the background is cluttered, suggest first using meitu image-cutout to cut out the subject and replace it with a clean background
  • Hands visible and not covering the face → hands covering the face will cause severe artifacts
  • Front-facing or slightly angled (not pure side/back) → extreme angles produce poor generation quality
  • Resolution short side ≥ 300px → if too small, it is recommended to first use meitu image-superres-enhance to improve clarity
  • Tight clothing is better than loose clothing → loose clothing (long skirts, coats) blurs joints and causes distortion

Check the reference video:

  • Duration of 3-10 seconds is optimal → if too long, it is recommended to cut the core segment
  • Single-person shot → multi-person shots cause motion confusion
  • Stable camera (fixed position) → handheld shake/rapid zoom introduces noise
  • Full body within the frame → the person going out of frame causes motion loss
  • Moderate motion speed → extremely fast motion (such as breaking) easily produces limb ghosting
  • Simple background → complex backgrounds interfere with pose estimation

Detailed evaluation criteria: When a more detailed input evaluation is needed, read references/input-quality-guide.md

If the input quality has serious issues (such as a half-body image + a full-body dance video),directly inform the user of the problem and suggest replacing the material, do not force the API call.

2. Build the Prompt

Key principle: the prompt describes the character's appearance and scene, not the motion. The motion comes from the reference video; writing motion descriptions in the prompt will conflict with the video signal.

Prompt construction strategy:

{角色外观描述}, {场景/背景描述}, {画面质量修饰}

Character appearance — extracted from the target image or supplemented by the user:

  • Gender, age range, body type
  • Clothing description (color, style, material)
  • Hairstyle, hair color
  • Distinctive features (accessories, tattoos, etc.)

Scene/background — by default consistent with the target image; the user may specify:

  • "white studio background" (clean, all-purpose background)
  • "city street at sunset" (scenario-based)
  • Keep it concise, no more than one sentence

Image quality — fixed suffix:

  • high quality, detailed, natural proportions, realistic hands

Prompt example:

Scene Prompt
Girl dancing a young woman in a white crop top and jeans, long black hair, white studio background, high quality, detailed, natural proportions
Anime character anime character with blue hair and school uniform, rooftop scene, vibrant colors, high quality, consistent style
Business person a man in a dark blue suit, short hair, modern office background, high quality, professional lighting

Taboos:

  • Do not describe specific motions ("hip-hop dance", "waving hands") → the motion is determined by the video
  • Do not use platform-specific syntax (--no、(keyword:1.5)、<lora:xxx>）
  • Do not use negative descriptions → express positively ("clean background" rather than "no clutter")

3. Call meitu-cli

meitu video-motion-transfer \
  --skill_name skill_meitu-video-dance \
  --image_list "{image_url}" \
  --reference_video_list "{video_url}" \
  --prompt "{composed_prompt}" \
  --download-dir "{output_dir}"

Asynchronous task description:

  • This command is an asynchronous task; the CLI will automatically poll until completion (internally calling meitu task wait）
  • Generation time is usually 30s-120s, depending on video duration and server load
  • Return JSON:ok: true → the result is in downloaded_files[0].saved_path(because --download-dir）；ok: false was used → first check the CLI raw field code、hint、error_name、action_url, then classify it as meitu-tools according to error_type

4. Quality judgment and error degradation

On success, check the generation quality:

  • Character identity consistency — whether the face, clothing, and hairstyle match the original image
  • Motion completeness — whether the motion in the reference video is fully reproduced
  • Limb naturalness — whether there are extra limbs, distorted joints, or abnormal fingers
  • Temporal coherence — whether it is smooth between frames, and whether there is flickering/jumping

Error degradation strategy on failure:

Level Action Description
L1 Simplify prompt Remove decorative modifiers, keep the core appearance + quality suffix
L2 Check input match Confirm whether the body type in the image matches the body type of the person in the video; if not, suggest replacing the image
L3 Shorten the reference video Clip the 3-5 second segment with the clearest motion and retry
L4 Simplify the target image Use meitu image-cutout to remove the complex background and retry
L5 Stop and report an error 2 consecutive failures → report code and hint, do not retry

Common error code handling:

  • Original error classified by meitu-tools as ORDER_REQUIRED → insufficient balance, inform the user to top up, provide action_url
  • Original error classified by meitu-tools as CREDENTIALS_MISSING → AK/SK not configured, prioritize prompting to use environment variables or preset ~/.meitu/credentials.json; only when the user explicitly requests writing to local credentials, then prompt meitu config set-ak/set-sk
  • Timeout → video too long or server busy, suggest shortening the video or retrying later

Refine

Present the result: Show the generated video to the user, explaining:

  • "The motion comes from the reference video you provided"
  • Point out possible quality issues (such as hand details, facial consistency)
  • If there are obvious flaws, proactively propose fix suggestions

Feedback classification and corresponding adjustments:

Feedback type Adjustment method
"Motion is wrong/doesn't look like it" Check the reference video quality, suggest switching to a clearer video segment
"The person doesn't look like the original image" Enhance the appearance description in the prompt, add more details
"Background is wrong" Adjust the scene description in the prompt
"Image quality is poor" Check the input image resolution, upscale first if necessary
"Hands/face have problems" This is a common limitation of current technology, suggest fixing with post-processing tools; or try reference images from different angles
"Want to change the motion" The user needs to provide a new reference video

Iteration pace: Suggest a maximum of 3 rounds. After that, proactively suggest:

  • Change the input material (change the image or change the video)
  • Break down the requirements (e.g., first generate a satisfactory character image, then do motion transfer)
  • Lower expectations and explain the current technical limitations

Deliver

Directly use the output_dir parsed by Preflight.

mv {file} {output_dir}/{date}_{name}.mp4

Naming convention: {date}_{描述性名称}.mp4, e.g., 2026-03-23_girl-hip-hop-dance.mp4

Record (project mode MUST / one-off mode skip)

can_record = false → skip. In one-off mode, feedback is only valid for the current conversation.

can_record definition:can_record = cwd 有 openclaw.yaml AND $VISUAL 存在(if either is missing, it is false)

User approved style → read $VISUAL/memory/observations/observations.yaml → scan similar key → merge or append → write back. len(projects) >= 2 → propose promotion (non-blocking). Propose promotion (non-blocking: mention at the end of the reply, do not interrupt the main flow) → confirmed → write target + delete observation entry

User rejected ("don't want XX") → has openclaw.yaml → ask scope → project: DESIGN.md Constraints / universal: quality.yaml no openclaw.yaml → current task only

No feedback → skip entirely.

Output

Property Value
Format MP4 video
Naming {date}_{effect-name}.mp4
Duration Consistent with the reference video (usually 3-10 seconds)
Path Determined by the Deliver step