SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsmeitu-image-fix

meitu-image-fix

Design & creation

View repository

Updated 10-16 00:23

About this skill

Meitu Image Fix

Overview

Receive a problematic image, automatically diagnose the three types of issues—image quality/portrait/content—plan the optimal repair pipeline, and call the meitu-cli toolchain in order to complete the repair. The core value isdiagnosis + chaining—the user does not need to know which tool to use, only needs to say "help me fix it."

Dependencies

  • meitu-cli (>=2.0.6): npm install -g meitu-cli@latest
  • Credential: prefer env MEITU_OPENAPI_ACCESS_KEY / MEITU_OPENAPI_SECRET_KEY, or preset ~/.meitu/credentials.json; only when the user explicitly requests writing local credentials, then execute meitu config set-ak / meitu config set-sk

Path alias: below, $VISUAL = {OPENCLAW_HOME}/workspace/visual/

Core Workflow

Preflight → [Context: 跳过] → Execute (诊断 → 规划 → 逐步执行) → Deliver
              ↑ 工具型 skill，无创意上下文需求，始终跳过 Context

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 can_record = cwd has openclaw.yaml AND $VISUAL exists (if either is missing, false)
  4. Resolve output_dir: openclaw.yaml → ./output/ | else → $VISUAL/output/meitu-image-fix/ mkdir -p {output_dir}

Context

Tool-type skill, no creative context requirement, always skip Context and go directly to Execute.

Execute

Diagnosis: look at the image + listen to the user

Determine the problem from two information sources:

Visual analysis(use the Read tool to read the image, multimodal recognition of issues):

Observed phenomenon Tag
Low resolution / obvious blur / compression aliasing / heavy noise quality:blur
Visible watermark / semi-transparent text overlay content:watermark
Passersby/clutter in the frame that are not part of the subject content:unwanted-object
Typos in the text content content:text-error
Defects in the background (messy/damaged/continuity error) content:bg-flaw
Face visible + rough/dull skin with obvious acne marks portrait:skin
Face visible + facial features can be optimized portrait:feature

User intent parsing(what the user says maps directly to a tag):

User expression Tag
"make it clearer" "super-resolution" "high definition" "blurry" "poor image quality" "old photo" quality:blur
"remove watermark" "watermark" content:watermark
"remove passersby" "remove clutter" "erase XX" content:unwanted-object
"change text" "fix typo" "change X to Y" content:text-error
"clean up the background" "messy background" content:bg-flaw
"skin smoothing" "beauty retouching" "acne removal" "brighten" "skin" portrait:skin
"make it more refined" "facial feature adjustment" portrait:feature
"remove background" "cutout" "transparent background" cutout
"fix it" "help me process it" (vague instruction) Completely rely on visual analysis

Merge the two information sources to obtain the final problem list.

Plan the repair pipeline

Execution order (hard constraint—wrong order will severely affect the result):

Phase 1: image-superres-enhance 先提升画质，后续操作在清晰图上效果更好
Phase 2: image-edit / cutout    内容修复在高清底图上精度更高
Phase 3: image-edit(gummy_pro)  人像精修放最后，避免被前面步骤的重编码破坏

Pipeline routing table:

Problem tag Phase Tool Command
quality:blur 1 image-superres-enhance meitu image-superres-enhance
content:watermark 2 image-edit meitu image-edit --model praline_pro
content:unwanted-object 2 image-edit meitu image-edit --model praline_pro
content:text-error 2 image-edit meitu image-edit --model praline_pro
content:bg-flaw 2 image-edit meitu image-edit --model praline_pro
cutout 2 image-cutout meitu image-cutout
portrait:skin 3 portrait-retouch meitu image-edit --model gummy_pro
portrait:feature 3 portrait-retouch meitu image-edit --model gummy_pro

When there are multiple edit operations within Phase 2, execute them serially one by one (each uses the URL output by the previous step). When cutout coexists with other Phase 2 operations: edit first, then cutout (background removal should come after the content is fixed).

Confirm the pipeline with the user(for multi-step repair):

List the diagnosis results and plan steps, and proceed only after confirmation:

Diagnosis: image is blurry + has a watermark + facial skin texture is rough Repair plan: ① Super-resolution → ② Watermark removal → ③ Natural beautification Confirm to start?

The user raised only a single issue and the diagnosis is unambiguous → skip confirmation and execute directly.

Execute step by step

Chained invocation mechanism: Add --jsonto each step, parse the returned JSON's media_urls[0] as the next step's --image_list input. Add --download-dir {output_dir} only to the last step to save locally. After adding --download-dir , the JSON output will have an extra downloaded_files array; directly take downloaded_files[0].saved_path to get the full path of the downloaded file, then rename it in the Deliver stage.

URL passing fallback: if the next step fails when using media_urls[0] as the --image_list input (UPLOAD_ERROR), first download to the system temporary directory and then retry with the local path:

curl -sL -o /tmp/meitu-fix-step{N}.{ext} {url}
meitu <command> --image_list /tmp/meitu-fix-step{N}.{ext} ...

Intermediate filesmustbe stored in /tmp/ or the system temporary directory;it is forbiddento store them in the skill folder or the directory where the source file is located. Clean up after the pipeline is complete:rm -f /tmp/meitu-fix-step*.{ext}。

Phase 1: Super-resolution — image-superres-enhance

Construct the --prompt content description based on the image subject:

Subject type --prompt Basis for judgment
Mainly a portrait portrait photo The face is the visual focus
Product/object e-commerce product image Products, food, close-ups of physical objects
Graphics/screenshots text document or UI screenshot UI screenshots, text images, illustrations, icons
Uncertain general image Default description when it cannot be determined 
meitu image-superres-enhance --image_url {input} --prompt "{content_description}" --json --skill_name skill_meitu-image-fix

Parse:ok: true → media_urls[0] pass to the next step. ok: false → record the error,continue with the original imagefor subsequent steps (an upscale failure does not block the pipeline).

Phase 2: Content repair — image-edit / image-cutout

Each content issue corresponds to one edit instruction, executed serially one by one:

Issue --prompt example Note
Watermark "去除图片中的水印" A large watermark area may require 2 rounds
Passersby/clutter "消除画面中的路人" / "去掉右下角的杂物" Describing the specific location works better
Text errors "将文字'XXX'改为'YYY'" Specify the original text and the corrected text
Background flaws "修复背景中的破损区域" Describe the specific area 
meitu image-edit --image_list {prev_url} --prompt "{instruction}" --model praline_pro --json --skill_name skill_meitu-image-fix

For each step, take media_urls[0] and pass it to the next step.

If the issue is background removal (cutout):

meitu image-cutout --image_url {prev_url} --json --skill_name skill_meitu-image-fix

This step can only be performed on people, pets, products, icons, and seals; when the subject model is clearly known, you may append --model_type 0、--model_type 1 or --model_type 2, otherwise omit it and let the model be selected automatically.

Phase 3: Portrait retouching — image-edit (gummy_pro)

meitu image-edit --image_list {prev_url} --prompt "{portrait_retouch_prompt}" --model gummy_pro --json --skill_name skill_meitu-image-fix
User request portrait_retouch_prompt
Natural retouching / minor flaws / everyday photos natural portrait retouch, improve skin texture and facial details, keep facial structure unchanged
Explicitly requests strong effects / severe flaws / commercial use strong but realistic portrait retouch, smoother skin and cleaner facial details, keep facial structure unchanged
Not explicitly stated Use the natural retouching prompt by default

Constraint：portrait retouch Only applicable to real single-person portraits. Skip this step and inform the user in the following cases:

  • Multi-person images
  • Non-real portrait (cartoon, illustration, stick figure, AI-drawn style)
  • Face too small or unclear

When the server refuses, simply skip.

When there are both portrait:skin and portrait:feature → a single image-edit --model gummy_pro covers both (no need to call twice).

Final step download: add --download-dir {output_dir}to the last command of the pipeline to save the result file directly to local.

Error handling

Each step is handled independently; a single step's failure does not block the pipeline:

L1: 简化 prompt — 移除修饰语，保留核心指令（edit 步骤适用）
L2: 省略可选参数 — 如裁剪掉非必要修饰，让服务端按默认路由处理
L3: 跳过当前步骤 — 用上一步结果继续后续管线
L4: 连续 2 步失败 → 停止管线，将最后一个成功步骤的结果下载到 {output_dir} 交付，并说明哪些步骤未完成
L5: 首步即失败 → 检查凭证/余额，报错含 code + hint

Special error codes:

  • ORDER_REQUIRED → prompt the user to top up, provide action_url
  • CREDENTIALS_MISSING → prompt to configure AK/SK

Deliver

output_dir has already been pre-resolved in Preflight; the --download-dir of the last step in Execute has already been written directly to that directory, and Deliver only does rename.

mv {output_dir}/{task_id_file} {output_dir}/{date}_{effect}.{ext}

Record (MUST in project mode / skipped in one-off mode)

can_record = cwd has openclaw.yaml AND $VISUAL exists (missing either one means false) → skip when false. In one-off mode, feedback is only valid for the current conversation.

User approved style → read $VISUAL/memory/observations/observations.yaml → scan similar key → merge or append → write back. len(projects) >= 2 → propose promotion (non-blocking) → non-blocking promotion proposal (mention at the end of the reply: "You prefer X in N projects, do you want to save it?") → confirmed → write target + delete observation

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

No feedback → skip entirely (do not read observations.yaml, zero overhead).

Output

  • Format: consistent with input (jpg→jpg, png→png), cutout outputs png (with alpha channel)
  • Naming：{date}_{修复摘要}.{ext}
    • Single fix:2026-03-23_upscale.jpg
    • Chained fix:2026-03-23_upscale+remove-watermark+beauty.jpg
  • Location: determined by the Deliver step

Boundaries

This skill only doesfixing— making an existing image better. It does not do creative tasks:

Do not Hand off
Style transfer (turn into comic/oil painting/3D) image-style-transfer(or via meitu-tools direct call)
Background/scene replacement (creative replacement) image-background-replace or meitu-visual-me background replacement workflow
Poster layout/text layout meitu-poster
AI portrait/portrait generation image-portrait-generate or meitu-visual-me

Boundary judgment: if the user's intent is "fix the problems with this image" → this skill. If the user's intent is "turn this image into something else" → inform the user and suggest the corresponding skill.