IMA AI Creation

⚠️ 重要：模型 ID 参考

CRITICAL: When calling the script, you MUST use the exact model_id (second column), NOT the friendly model name. Do NOT infer model_id from the friendly name (e.g., ❌ nano-banana-pro is WRONG; ✅ gemini-3-pro-image is CORRECT).

Quick Reference Table:

图像模型 (Image Models)

视频模型 (Video Models)

音乐模型 (Music Models)

语音模型 (Speech/TTS Models)

How to get the correct model_id:

1. Check this table first
2. Use --list-models --task-type <type> to query available models
3. Refer to command examples in this SKILL.md

Example:

# ❌ WRONG: Inferring from friendly name
--model-id nano-banana-pro

# ✅ CORRECT: Using exact model_id from table
--model-id gemini-3-pro-image

⚠️ MANDATORY PRE-CHECK: Read Knowledge Base First!

If ima-knowledge-ai is not installed: Skip all "Read …" steps below; use only this SKILL's 📥 User Input Parsing (media type → task_type) and the Recommended Defaults / model tables for each media type.

BEFORE executing ANY multi-media generation task, you MUST:

1. Check for workflow complexity — Read ima-knowledge-ai/references/workflow-design.md if:

   • User mentions: "MV"、"宣传片"、"完整作品"、"配乐"、"soundtrack"
   • Task spans multiple media types (image + video, video + music, etc.)
   • Complex multi-step workflows that need task decomposition

2. Check for visual consistency needs — Read ima-knowledge-ai/references/visual-consistency.md if:

   • User mentions: "系列"、"多张"、"同一个"、"角色"、"续"、"series"、"same"
   • Task involves: multiple images/videos, character continuity, product shots
   • Second+ request about same subject (e.g., "旺财在游泳" after "生成旺财照片")

3. Check video modes — Read ima-knowledge-ai/references/video-modes.md if:

   • Any video generation task
   • Need to understand: image_to_video vs reference_image_to_video difference

4. Check model selection — Read ima-knowledge-ai/references/model-selection.md if:

   • Unsure which model to use
   • Need cost/quality trade-off guidance
   • User specifies budget or quality requirements

Why this matters:

• Multi-media workflows need proper task sequencing (e.g., video duration → matching music duration)
• AI generation defaults to 独立生成 each time — without reference images, results will be inconsistent
• Wrong video mode = wrong result (image_to_video ≠ reference_image_to_video)
• Model choice affects cost and quality significantly

Example multi-media workflow:

User: "帮我做个产品宣传MV，有背景音乐，主角是旺财小狗"

❌ Wrong: 
  1. Generate dog image (random look)
  2. Generate video (different dog)
  3. Generate music (unrelated)

✅ Right:
  1. Read workflow-design.md + visual-consistency.md
  2. Generate Master Reference: 旺财小狗图片
  3. Generate video shots using image_to_video with 旺财 as first frame
  4. Get video duration (e.g., 15s)
  5. Generate BGM with matching duration and mood

How to check:

# Step 0: Determine media type first (image / video / music / speech)
# From user request: "画"/"生成图"/"image" → image; "视频"/"video" → video; "音乐"/"歌"/"music"/"BGM" → music; "语音"/"朗读"/"TTS"/"speech" → speech
# Then choose task_type and model from the corresponding section (image: text_to_image/image_to_image; video: text_to_video/...; music: text_to_music; speech: text_to_speech)

# Step 1: Read knowledge base based on task type
if multi_media_workflow:
    read("~/.openclaw/skills/ima-knowledge-ai/references/workflow-design.md")

if "same subject" or "series" or "character":
    read("~/.openclaw/skills/ima-knowledge-ai/references/visual-consistency.md")

if video_generation:
    read("~/.openclaw/skills/ima-knowledge-ai/references/video-modes.md")

# Step 2: Execute with proper sequencing and reference images
# (see workflow-design.md for specific patterns)

No exceptions — for simple single-media requests, you can proceed directly. For complex multi-media workflows, read the knowledge base first.

📥 User Input Parsing (Media Type & Task Routing)

Purpose: So that any agent parses user intent consistently, first determine the media type from the user's request, then choose task_type and model.

1. User phrasing → media type (do this first)

If the request mixes media (e.g. "宣传片+配乐"), treat as multi-media workflow: read workflow-design.md, then plan image → video → music steps and use the correct task_type for each step.

2. Model and parameter parsing

• Image: For model name → model_id and size/aspect_ratio parsing, follow the same rules as in ima-image-ai skill (User Input Parsing section).

• Video: For task_type (t2v / i2v / first_last / reference), model alias → model_id, and duration/resolution/aspect_ratio, follow ima-video-ai skill (User Input Parsing section).

• Music: Suno (sonic) vs DouBao BGM/Song — infer from "BGM"/"背景音乐" → BGM; "带歌词"/"人声" → Suno or Song. Use model_id sonic, GenBGM, GenSong per "Recommended Defaults" and "Music Generation" tables below.

• Speech (TTS): Get model_id from GET /open/v1/product/list?category=text_to_speech or run script with --task-type text_to_speech --list-models. Map user intent to parameters using product form_config:

  User intent / phrasing	Parameter (if in form_config)	Notes
  女声 / 女声朗读 / female voice	voice_id / voice_type	Use value from form_config options
  男声 / 男声朗读 / male voice	voice_id / voice_type	Use value from form_config options
  语速快/慢 / speed up/slow	speed	e.g. 0.8–1.2
  音调 / pitch	pitch	If supported
  大声/小声 / volume	volume	If supported

  If the user does not specify, use form_config defaults. Pass extra params via --extra-params '{"speed":1.0}'. Only send parameters present in the product’s credit_rules/attributes or form_config (script reflection strips others on retry).

⚙️ How This Skill Works

For transparency: This skill uses a bundled Python script (scripts/ima_create.py) to call the IMA Open API. The script:

• Sends your prompt to two IMA-owned domains (see "Network Endpoints" below)
• Uses --user-id only locally as a key for storing your model preferences
• Returns image/video/music URLs when generation is complete

What gets sent to IMA servers:

• ✅ Your prompt/description (image/video/music)
• ✅ Model selection (SeeDream/Wan/Suno/etc.)
• ✅ Generation parameters (size, duration, style, etc.)
• ❌ NO API key in prompts (key is used for authentication only)
• ❌ NO user_id (it's only used locally)

What's stored locally:

• ~/.openclaw/memory/ima_prefs.json - Your model preferences (< 1 KB)
• ~/.openclaw/logs/ima_skills/ - Generation logs (auto-deleted after 7 days)

🌐 Network Endpoints Used

Key Points:

• Music tasks (text_to_music) and TTS tasks (text_to_speech) only use api.imastudio.com.
• Image/video tasks require imapi.liveme.com to obtain presigned URLs for uploading input images.
• Your API key is sent to both api.imastudio.com and imapi.liveme.com (both owned by IMA Studio).
• Verify network calls: tcpdump -i any -n 'host api.imastudio.com or host imapi.liveme.com'. See this document: 🌐 Network Endpoints Used and ⚠️ Credential Security Notice for full disclosure.

⚠️ Credential Security Notice

Your API key is sent to both IMA-owned domains:

1. Authorization: Bearer ima_xxx... → api.imastudio.com (main API)
2. Query param appUid=ima_xxx... → imapi.liveme.com (upload service)

Security best practices:

• 🧪 Use test keys for experiments: Generate a separate API key for testing.
• 🔍 Monitor usage: Check https://imastudio.com/dashboard for unauthorized activity.
• ⏱️ Rotate keys: Regenerate your API key periodically (monthly recommended).
• 📊 Review logs: Check ~/.openclaw/logs/ima_skills/ for unexpected API calls.

Why two domains? IMA Studio uses a microservices architecture:

• api.imastudio.com: Core AI generation API
• imapi.liveme.com: Specialized image/video upload service (shared infrastructure)

Both domains are operated by IMA Studio. The same API key grants access to both services.

Agent Execution (Internal Reference)

Note for users: You can review the script source at scripts/ima_create.py anytime.
The agent uses this script to simplify API calls. Music tasks use only api.imastudio.com, while image/video tasks also call imapi.liveme.com for file uploads (see "Network Endpoints" above).

Use the bundled script internally for all task types — it ensures correct parameter construction:

# ─── Image Generation ──────────────────────────────────────────────────────────

# Basic text-to-image (default model)
python3 {baseDir}/scripts/ima_create.py \
  --api-key $IMA_API_KEY --task-type text_to_image \
  --model-id doubao-seedream-4.5 --prompt "a cute puppy on grass, photorealistic" \
  --user-id {user_id} --output-json

# Text-to-image with size override (Nano Banana2)
python3 {baseDir}/scripts/ima_create.py \
  --api-key $IMA_API_KEY --task-type text_to_image \
  --model-id gemini-3.1-flash-image --prompt "city skyline at sunset, 4K" \
  --size 2k --user-id {user_id} --output-json

# Image-to-image with input URL
python3 {baseDir}/scripts/ima_create.py \
  --api-key $IMA_API_KEY --task-type image_to_image \
  --model-id doubao-seedream-4.5 --prompt "turn into oil painting style" \
  --input-images https://example.com/photo.jpg --user-id {user_id} --output-json

# ─── Video Generation ──────────────────────────────────────────────────────────

# Basic text-to-video (default model, 5s 720P)
python3 {baseDir}/scripts/ima_create.py \
  --api-key $IMA_API_KEY --task-type text_to_video \
  --model-id wan2.6-t2v --prompt "a puppy dancing happily, cinematic" \
  --user-id {user_id} --output-json

# Text-to-video with extra params (10s 1080P)
python3 {baseDir}/scripts/ima_create.py \
  --api-key $IMA_API_KEY --task-type text_to_video \
  --model-id wan2.6-t2v --prompt "dramatic ocean waves, sunset" \
  --extra-params '{"duration":10,"resolution":"1080P","aspect_ratio":"16:9"}' \
  --user-id {user_id} --output-json

# Image-to-video (animate static image)
python3 {baseDir}/scripts/ima_create.py \
  --api-key $IMA_API_KEY --task-type image_to_video \
  --model-id wan2.6-i2v --prompt "camera slowly zooms in, gentle movement" \
  --input-images https://example.com/photo.jpg --user-id {user_id} --output-json

# First-last frame video (two images)
python3 {baseDir}/scripts/ima_create.py \
  --api-key $IMA_API_KEY --task-type first_last_frame_to_video \
  --model-id kling-video-o1 --prompt "smooth transition between frames" \
  --input-images https://example.com/frame1.jpg https://example.com/frame2.jpg \
  --user-id {user_id} --output-json

# ─── Music Generation ──────────────────────────────────────────────────────────

# Basic text-to-music (Suno default)
python3 {baseDir}/scripts/ima_create.py \
  --api-key $IMA_API_KEY --task-type text_to_music \
  --model-id sonic --prompt "upbeat electronic music, 120 BPM, no vocals" \
  --user-id {user_id} --output-json

# Music with custom lyrics (Suno custom mode)
python3 {baseDir}/scripts/ima_create.py \
  --api-key $IMA_API_KEY --task-type text_to_music \
  --model-id sonic --prompt "pop ballad, emotional" \
  --extra-params '{"custom_mode":true,"lyrics":"Your custom lyrics here...","vocal_gender":"female"}' \
  --user-id {user_id} --output-json

# Background music (DouBao BGM)
python3 {baseDir}/scripts/ima_create.py \
  --api-key $IMA_API_KEY --task-type text_to_music \
  --model-id GenBGM --prompt "relaxing ambient music for meditation" \
  --user-id {user_id} --output-json

# ─── Text-to-Speech (TTS) ─────────────────────────────────────────────────────

# List TTS models first to get model_id, then generate speech
python3 {baseDir}/scripts/ima_create.py \
  --api-key $IMA_API_KEY --task-type text_to_speech --list-models

# TTS: use model_id from list above (prompt = text to speak)
python3 {baseDir}/scripts/ima_create.py \
  --api-key $IMA_API_KEY --task-type text_to_speech \
  --model-id <model_id from list> --prompt "Text to be spoken here." \
  --user-id {user_id} --output-json

The script outputs JSON with url, model_name, credit — use these values in the UX protocol messages below. The script internals (product list query, parameter construction, polling) are invisible to users.

Overview

Call IMA Open API to create AI-generated content. All endpoints require an ima_* API key. The core flow is: query products → create task → poll until done.

🔒 Security & Transparency Policy

This skill is community-maintained and open for inspection.

✅ What Users CAN Do

Full transparency:

• ✅ Review all source code: Check scripts/ima_create.py and ima_logger.py anytime
• ✅ Verify network calls: Music tasks use api.imastudio.com only; image/video tasks also use imapi.liveme.com (see "Network Endpoints" section)
• ✅ Inspect local data: View ~/.openclaw/memory/ima_prefs.json and log files
• ✅ Control privacy: Delete preferences/logs anytime, or disable file writes (see below)

Configuration allowed:

• ✅ Set API key in environment or agent config:
  • Environment variable: export IMA_API_KEY=ima_your_key_here
  • OpenClaw/MCP config: Add IMA_API_KEY to agent's environment configuration
  • Get your key at: https://imastudio.com
• ✅ Use scoped/test keys: Test with limited API keys, rotate after testing
• ✅ Disable file writes: Make prefs/logs read-only or symlink to /dev/null

Data control:

• ✅ View stored data: cat ~/.openclaw/memory/ima_prefs.json
• ✅ Delete preferences: rm ~/.openclaw/memory/ima_prefs.json (resets to defaults)
• ✅ Delete logs: rm -rf ~/.openclaw/logs/ima_skills/ (auto-cleanup after 7 days anyway)

⚠️ Advanced Users: Fork & Modify

If you need to modify this skill for your use case:

1. Fork the repository (don't modify the original)
2. Update your fork with your changes
3. Test thoroughly with limited API keys
4. Document your changes for troubleshooting

Note: Modified skills may break API compatibility or introduce security issues. Official support only covers the unmodified version.

❌ What to AVOID (Security Risks)

Actions that could compromise security:

• ❌ Sharing API keys publicly or in skill files
• ❌ Modifying API endpoints to unknown servers
• ❌ Disabling SSL/TLS certificate verification
• ❌ Logging sensitive user data (prompts, IDs, etc.)
• ❌ Bypassing authentication or billing mechanisms

Why this matters:

1. API Compatibility: Skill logic aligns with IMA Open API schema
2. Security: Malicious modifications could leak credentials or bypass billing
3. Support: Modified skills may not be supported
4. Community: Breaking changes affect all users

📋 Privacy & Data Handling Summary

What this skill does with your data:

Privacy recommendations:

1. Use test/scoped API keys for initial testing
2. Note: --user-id is never sent to IMA servers - it's only used locally as a key for storing preferences in ~/.openclaw/memory/ima_prefs.json
3. Review source code at scripts/ima_create.py to verify network calls (search for create_task function)
4. Rotate API keys after testing or if compromised

Get your IMA API key: Visit https://imastudio.com to register and get started.

🔧 For Skill Maintainers Only

Version control:

• All changes must go through Git with proper version bumps (semver)
• CHANGELOG.md must document all changes
• Production deployments require code review

File checksums (optional):

# Verify skill integrity
sha256sum SKILL.md scripts/ima_create.py

If users report issues, verify file integrity first.

🧠 User Preference Memory (Image)

User preferences have highest priority when they exist. But preferences are only saved when users explicitly express model preferences — not from automatic model selection.

Storage: ~/.openclaw/memory/ima_prefs.json

Single file, shared across all IMA skills:

{
  "user_{user_id}": {
    "text_to_image":  { "model_id": "doubao-seedream-4.5", "model_name": "SeeDream 4.5", "credit": 5,  "last_used": "2026-02-27T03:07:27Z" },
    "image_to_image": { "model_id": "doubao-seedream-4.5", "model_name": "SeeDream 4.5", "credit": 5,  "last_used": "2026-02-27T03:07:27Z" },
    "text_to_speech": { "model_id": "<from product list>", "model_name": "...", "credit": 2, "last_used": "..." }
  }
}

Model Selection Flow (Image Generation)

Step 1: Get knowledge-ai recommendation (if installed)

knowledge_recommended_model = read_ima_knowledge_ai()  # e.g., "SeeDream 4.5"

Step 2: Check user preference

user_pref = load_prefs().get(f"user_{user_id}", {}).get(task_type)  # e.g., {"model_id": "midjourney", ...}

Step 3: Decide which model to use

if user_pref exists:
    use_model = user_pref["model_id"]  # Highest priority
else:
    use_model = knowledge_recommended_model or fallback_default

Step 4: Check for mismatch (for later hint)

if user_pref exists and knowledge_recommended_model != user_pref["model_id"]:
    mismatch = True  # Will add hint in success message

When to Write (User Explicit Preference ONLY)

✅ Save preference when user explicitly specifies a model:

❌ Do NOT save when:

• Agent auto-selects from knowledge-ai → not user preference
• Agent uses fallback default → not user preference
• User says generic quality requests (see "Clear Preference" below) → clear preference instead

When to Clear (User Abandons Preference)

🗑️ Clear preference when user wants automatic selection:

Implementation:

del prefs[f"user_{user_id}"][task_type]
save_prefs(prefs)

⭐ Model Selection Priority (Image)

Selection flow:

1. User preference (if exists) → Highest priority, always respect
2. ima-knowledge-ai skill (if installed) → Professional recommendation based on task
3. Fallback defaults → Use table below (only if neither 1 nor 2 exists)

Important notes:

• User preference is only saved when user explicitly specifies a model (see "When to Write" above)
• Knowledge-ai is always consulted (even when user pref exists) to detect mismatches
• When mismatch detected → add gentle hint in success message (does NOT interrupt generation)

The defaults below are FALLBACK only. User preferences have highest priority, then knowledge-ai recommendations.

When using user preference for image generation, show a line like:

🎨 根据你的使用习惯，将用 [Model Name] 帮你生成…
• 模型：[Model Name]（你的常用模型）
• 预计耗时：[X ~ Y 秒]
• 消耗积分：[N pts]

Preference Change Confirmation

When user switches to a different model than their saved preference:

💡 你之前喜欢用 [Old Model]，这次换成了 [New Model]。
要把 [New Model] 设为以后的默认吗？
回复「是」保存 / 回复「否」仅本次使用

⭐ Recommended Defaults

These are fallback defaults — only used when no user preference exists.
Always default to the newest and most popular model. Do NOT default to the cheapest.

Premium options:

• Image: Nano Banana Pro — Highest quality with size control (1K/2K/4K), higher cost (10-18 pts for text_to_image, 10 pts for image_to_image)
• Video: Kling O1, Sora 2 Pro, Google Veo 3.1 — Premium quality with longer duration options

Quick selection guide (production as of 2026-02-27, sorted by popularity):

• Image (4 models available) → SeeDream 4.5 (5, default); artistic → Midjourney 🎨 (8-10); budget → Nano Banana2 (4, 512px); premium → Nano Banana Pro (10-18)
• 🔥 Video from text (most popular) → Wan 2.6 (25, balanced); premium → Hailuo 2.3 (38); budget → Vidu Q2 (5)
• 🔥 Video from image (most popular) → Wan 2.6 (25)
• Music → Suno (25); DouBao BGM/Song (30 each)
• Cheapest → Nano Banana2 512px (4) for image; Vidu Q2 (5) for video

Selection guide by use case:

Image Generation:

• General image generation → SeeDream 4.5 (5pts)
• Custom aspect ratio (16:9, 9:16, 4:3, etc.) → SeeDream 4.5 🌟 or Nano Banana Pro/2 🆕 (native support)
• Budget-conscious / fast generation → Nano Banana2 (4pts)
• Highest quality with size control (1K/2K/4K) → Nano Banana Pro (text_to_image: 10-18pts, image_to_image: 10pts)
• Artistic/creative styles, illustrations, paintings → Midjourney 🎨 (8-10pts)
• Style transfer / image editing → SeeDream 4.5 (5pts) or Midjourney 🎨 (artistic)

Video Generation:

• General video generation → Wan 2.6 (25pts, most popular)
• Premium cinematic quality → Google Veo 3.1 (70-330pts) or Sora 2 Pro (122+pts)
• Budget video → Vidu Q2 (5pts) or Hailuo 2.0 (5pts)
• With audio support → Kling O1 (48+pts) or Google Veo 3.1 (70+pts)
• First/last frame animation → Kling O1 (48+pts)
• Reference image consistency → Kling O1 (48+pts) or Google Veo 3.1 (70+pts)

Music Generation:

• Custom song with lyrics, vocals, style → Suno sonic-v5 (25pts, default, ~2min)
  • Full control: custom_mode, lyrics, vocal_gender, tags, negative_tags
  • Best for: complete songs, vocal tracks, artistic compositions
• Background music / ambient loop → DouBao BGM (30pts, ~30s)
  • Simplified: prompt-only, no advanced parameters
  • Best for: video backgrounds, ambient music, short loops
• Simple song generation → DouBao Song (30pts, ~30s)
  • Simplified: prompt-only
  • Best for: quick song generation, structured vocal compositions
• User explicitly asks for cheapest → DouBao BGM/Song (6pts each) — only if explicitly requested

Speech (TTS) Generation:

• Text-to-speech / 语音合成 / 朗读 → text_to_speech. Always query GET /open/v1/product/list?category=text_to_speech (or --list-models) to get current model_id and credit. No fixed default; use first available or user preference. Voice/speed/format parameters: see "Model and parameter parsing" (TTS table) and "Speech (TTS) — text_to_speech" in this document.

⚠️ Technical Note for Suno:

model_version inside parameters.parameters (e.g., "sonic-v5") is different from the outer model_version field (which is "sonic"). Always set both correctly when creating Suno tasks.

⚠️ Production Image Models (4 available):

• SeeDream 4.5 (doubao-seedream-4.5) — 5 pts, default
• Midjourney 🎨 (midjourney) — 8/10 pts for 480p/720p, artistic styles
• Nano Banana2 (gemini-3.1-flash-image) — 4/6/10/13 pts for 512px/1K/2K/4K
• Nano Banana Pro (gemini-3-pro-image) — 10/10/18 pts for 1K/2K/4K

All other image models mentioned in older documentation are no longer available in production.

🌟 Parameter Support Notes (All Task Types):

Image Models (text_to_image / image_to_image)

🆕 MAJOR UPDATE: Nano Banana series now has NATIVE aspect_ratio support!

• Nano Banana Pro: ✅ Supports aspect_ratio (1:1, 16:9, 9:16, 4:3, 3:4) NATIVELY
• Nano Banana2: ✅ Supports aspect_ratio (1:1, 16:9, 9:16, 4:3, 3:4) NATIVELY
• SeeDream 4.5: ✅ Supports 8 ratios via virtual params (1:1, 16:9, 9:16, 4:3, 3:4, 2:3, 3:2, 21:9)
• Midjourney: ❌ 1:1 only (fixed 1024x1024)

aspect_ratio support details:

• ✅ aspect_ratio:
  • SeeDream 4.5: ✅ Supports 8 ratios via virtual params (1:1, 16:9, 9:16, 4:3, 3:4, 2:3, 3:2, 21:9)
  • Nano Banana2: ✅ Native support for 5 ratios (1:1, 16:9, 9:16, 4:3, 3:4)
  • Nano Banana Pro: ✅ Native support for 5 ratios (1:1, 16:9, 9:16, 4:3, 3:4)
  • Midjourney: ❌ 1:1 only (fixed 1024x1024)
• ✅ size:
  • Nano Banana2: 512px, 1K, 2K, 4K (via different attribute_ids, 4-13 pts)
  • Nano Banana Pro: 1K, 2K, 4K (via different attribute_ids, 10-18 pts)
  • SeeDream 4.5: Adaptive default (5 pts)
  • Midjourney: 480p/720p (via attribute_id, 8/10 pts)
• ❌ 8K: No model supports 8K (max is 4K via Nano Banana Pro)
• ❌ Non-standard aspect ratios (7:3, 8:5, etc.): Not supported. Use closest supported ratio or video models.
• ✅ n: Multiple outputs supported (1-4), credit × n

When user requests unsupported combinations for images:

• Midjourney + aspect_ratio (16:9, etc.): Recommend SeeDream 4.5 or Nano Banana series instead

  ❌ Midjourney 暂不支持自定义 aspect_ratio（仅支持 1024x1024 方形）
  
  ✅ 推荐方案：
    1. SeeDream 4.5（支持虚拟参数 aspect_ratio）
       • 支持比例：1:1, 16:9, 9:16, 4:3, 3:4, 2:3, 3:2, 21:9
       • 成本：5 积分（性价比最佳）
    2. Nano Banana Pro/2（原生支持 aspect_ratio）
       • 支持比例：1:1, 16:9, 9:16, 4:3, 3:4
       • 成本：4-18 积分（按尺寸）
  
  需要我帮你用 SeeDream 4.5 生成吗？
  

• Any model + 8K: Inform user no model supports 8K, max is 4K (Nano Banana Pro)
• Any model + non-standard ratio (7:3, 8:5, etc.): Non-standard ratio, not supported. Suggest closest supported ratio (e.g., 21:9 for ultra-wide, 2:3 for portrait)

Video Models (text_to_video / image_to_video / first_last_frame / reference_image)

• ✅ resolution: 540P, 720P, 1080P, 2K, 4K (model-dependent, higher res = higher cost)
• ✅ aspect_ratio: 16:9, 9:16, 1:1, 4:3 (model-dependent, check form_config)
• ✅ duration: 4s, 5s, 10s, 15s (model-dependent, longer = higher cost)
• ⚠️ generate_audio: Supported by Veo 3.1, Kling O1, Hailuo (check form_config)
• ✅ prompt_extend: AI-powered prompt enhancement (most models support)
• ✅ negative_prompt: Content exclusion (most models support)
• ✅ shot_type: Single/multi-shot control (model-dependent)
• ✅ seed: Reproducibility control (most models support, -1 = random)
• ✅ n: Multiple outputs (1-4), credit × n

🆕 Special Case: Pixverse Model Parameter (v1.0.7+)

Auto-Inference Logic for Pixverse V5.5/V5/V4:

• Problem: Pixverse V5.5, V5, V4 lack model field in form_config from Product List API
• Backend Requirement: Backend requires model parameter (e.g., "v5.5", "v5", "v4")
• Auto-Fix: System automatically extracts version from model_name and injects it
  • Example: model_name: "Pixverse V5.5" → auto-inject model: "v5.5"
  • Example: model_name: "Pixverse V4" → auto-inject model: "v4"
• Note: V4.5 and V3.5 include model in form_config (no auto-inference needed)
• Relevant Task Types: All video modes (text_to_video, image_to_video, first_last_frame_to_video, reference_image_to_video)

Error Prevention:

• Without auto-inference: err_code=400017 err_msg=Invalid value for model
• With auto-inference (v1.0.7+): Pixverse V5.5/V5/V4 work seamlessly ✅

Music Models (text_to_music)

Suno sonic-v5 (Full-Featured):

• ✅ custom_mode: Suno only (enables vocal_gender, lyrics, tags support)
• ✅ vocal_gender: Suno only (male/female/mixed, requires custom_mode=True)
• ✅ lyrics: Suno only (custom lyrics support, requires custom_mode=True)
• ✅ make_instrumental: Suno only (force instrumental, no vocals)
• ✅ auto_lyrics: Suno only (AI-generated lyrics)
• ✅ tags: Suno only (genre/style tags)
• ✅ negative_tags: Suno only (exclude unwanted styles)
• ✅ title: Suno only (song title)
• ❌ duration: Fixed-length output (DouBao ~30s, Suno ~2min, not user-controllable)
• ✅ n: Multiple outputs supported (1-2), credit × n

DouBao BGM/Song (Simplified):

• ✅ prompt: Text description only
• ❌ No advanced parameters (no custom_mode, lyrics, vocal control)
• ❌ duration: Fixed ~30s output

🎵 Suno Prompt Writing Guide (for gpt_description_prompt):

When using Suno, structure your prompt with these elements:

1. Genre/Style:

   • Examples: "lo-fi hip hop", "orchestral cinematic", "upbeat pop", "dark ambient", "indie folk", "electronic dance"

2. Tempo/BPM:

   • Examples: "80 BPM", "fast tempo", "slow ballad", "moderate pace 110 BPM"

3. Vocals Control:

   • No vocals: "no vocals" → set make_instrumental=true
   • With vocals: "female vocals" → set vocal_gender="female"
   • Male vocals: "male vocals" → set vocal_gender="male"
   • Mixed: Set vocal_gender="mixed"

4. Mood/Emotion:

   • Examples: "happy and energetic", "melancholic", "tense and dramatic", "peaceful and calming"

5. Negative Tags (exclude styles):

   • Use negative_tags: "heavy metal, distortion, screaming" to exclude unwanted elements

6. Duration Hint:

   • Examples: "60 seconds", "30 second loop", "2 minute track"
   • Note: Suno typically generates ~2min, not strictly controllable

Example Suno prompts:

"upbeat lo-fi hip hop, 90 BPM, no vocals, relaxed and chill"
→ Set: make_instrumental=true

"emotional pop ballad, slow tempo, female vocals, melancholic"
→ Set: vocal_gender="female"

"orchestral cinematic trailer music, epic and dramatic, 120 BPM, no vocals"
→ Set: make_instrumental=true, tags="orchestral,cinematic,epic"

"acoustic indie folk, gentle guitar, male vocals, warm and nostalgic"
→ Set: vocal_gender="male", tags="acoustic,indie,folk"

⚠️ Technical Note for Suno:

model_version inside parameters.parameters (e.g., "sonic-v5") is different from the outer model_version field (which is "sonic"). Always set both correctly.

Common Parameter Patterns

• n (batch generation): Supported by ALL models. Cost = base_credit × n. Creates n independent resources.
• seed: Supported by most models (-1 = random, >0 = reproducible results)
• prompt_extend: AI-powered prompt enhancement (video models only)

Decision Tree: When User Requests Unsupported Features

User asks for custom aspect ratio image (e.g. "7:3 landscape")
  → ❌ Image models don't support custom ratios
  → ✅ Solution: "图片模型不支持自定义比例。建议用视频模型(Wan 2.6 t2v)生成16:9视频，然后截取首帧作为图片。"

User asks for 8K image
  → ❌ No model supports 8K
  → ✅ Solution: "当前最高支持4K分辨率(Nano Banana Pro，18积分)。要使用吗？"

User asks for video with audio
  → Check model: Veo 3.1 / Kling O1 / Hailuo have generate_audio
  → ✅ Solution: "Veo 3.1 和 Kling O1 支持音频生成(需在参数中设置 generate_audio=True)。要用哪个？"

User asks for long music (e.g. "5 minute track")
  → ❌ Duration not user-controllable
  → ✅ Solution: "Suno 生成约2分钟音乐。需要更长时长可以生成多段后拼接。"

User asks for 30s video
  → Check model: Most models max 15s
  → ✅ Solution: "当前最长15秒。可选模型：Wan 2.6(15s, 75积分), Kling O1(10s, 96积分)。"

When user requests unsupported combinations:

• Video + audio (unsupported model) → "该模型不支持音频。建议用 Veo 3.1 或 Kling O1 (支持 generate_audio 参数)"
• Music + custom duration → "音乐时长由模型固定(Suno约2分钟,DouBao约30秒),无法自定义"
• Video duration > 15s → "当前最长15秒。可选模型：Wan 2.6(15s, 75积分), Kling O1(10s, 96积分)"

Note: Image-specific unsupported combinations (Midjourney + aspect_ratio, 8K, non-standard ratios) are documented in the "Image Models" section above.

🧠 User Preference Memory (Video)

User preferences have highest priority when they exist. But preferences are only saved when users explicitly express model preferences — not from automatic model selection.

Storage: ~/.openclaw/memory/ima_prefs.json

{
  "user_{user_id}": {
    "text_to_video":              { "model_id": "wan2.6-t2v",      "model_name": "Wan 2.6",  "credit": 25, "last_used": "..." },
    "image_to_video":             { "model_id": "wan2.6-i2v",      "model_name": "Wan 2.6",  "credit": 25, "last_used": "..." },
    "first_last_frame_to_video":  { "model_id": "kling-video-o1", "model_name": "Kling O1", "credit": 48, "last_used": "..." },
    "reference_image_to_video":   { "model_id": "kling-video-o1", "model_name": "Kling O1", "credit": 48, "last_used": "..." }
  }
}

Model Selection Flow (Video Generation)

Step 1: Get knowledge-ai recommendation (if installed)

knowledge_recommended_model = read_ima_knowledge_ai()  # e.g., "Wan 2.6"

Step 2: Check user preference

user_pref = load_prefs().get(f"user_{user_id}", {}).get(task_type)  # e.g., {"model_id": "kling-video-o1", ...}

Step 3: Decide which model to use

if user_pref exists:
    use_model = user_pref["model_id"]  # Highest priority
else:
    use_model = knowledge_recommended_model or fallback_default

Step 4: Check for mismatch (for later hint)

if user_pref exists and knowledge_recommended_model != user_pref["model_id"]:
    mismatch = True  # Will add hint in success message

When to Write (User Explicit Preference ONLY)

✅ Save preference when user explicitly specifies a model:

❌ Do NOT save when:

• Agent auto-selects from knowledge-ai → not user preference
• Agent uses fallback default → not user preference
• User says generic quality requests (see "Clear Preference" below) → clear preference instead

When to Clear (User Abandons Preference)

🗑️ Clear preference when user wants automatic selection:

Implementation:

del prefs[f"user_{user_id}"][task_type]
save_prefs(prefs)

⭐ Model Selection Priority (Video)

Selection flow:

1. User preference (if exists) → Highest priority, always respect
2. ima-knowledge-ai skill (if installed) → Professional recommendation based on task
3. Fallback defaults → Use table below (only if neither 1 nor 2 exists)

Important notes:

• User preference is only saved when user explicitly specifies a model (see "When to Write" above)
• Knowledge-ai is always consulted (even when user pref exists) to detect mismatches
• When mismatch detected → add gentle hint in success message (does NOT interrupt generation)

The defaults below are FALLBACK only. User preferences have highest priority, then knowledge-ai recommendations.

💬 User Experience Protocol (IM / Feishu / Discord) v2.0 🆕

v2.0 Updates (aligned with ima-image-ai v1.3):

• Added Step 0 for correct message ordering (fixes group chat bug)
• Added Step 5 for explicit task completion
• Enhanced Midjourney support with proper timing estimates
• Now 6 steps total (0-5): Acknowledgment → Pre-Gen → Progress → Success/Failure → Done

This skill runs inside IM platforms (Feishu, Discord via OpenClaw).
Generation takes 10 seconds (music) up to 6 minutes (video). Never let users wait in silence.
Always follow all 6 steps below, every single time.

🚫 Never Say to Users

The following are internal implementation details. Never mention them in any user-facing message, under any circumstances:

User messages must only contain: model name, estimated/actual time, credits consumed, result URL, and natural language status updates.

Estimated Generation Time (All Task Types)

estimated_max_seconds = upper bound of the range (e.g. 60 for SeeDream 4.5, 40 for Nano Banana2, 120 for Nano Banana Pro, 90 for Midjourney, 180 for Kling 2.6, 360 for Kling O1).

Step 0 — Initial Acknowledgment Reply (Normal Reply) 🆕

⚠️ CRITICAL: This step is essential for correct message ordering in IM platforms (Feishu, Discord).

Before doing anything else, reply to the user with a friendly acknowledgment message using your normal reply (not message tool). This reply will automatically appear FIRST in the conversation.

Example acknowledgment messages:

For images:

好的!来帮你画一只萌萌的猫咪 🐱

收到！马上为你生成一张 16:9 的风景照 🏔️

OK! Starting image generation with SeeDream 4.5 🎨

For videos:

好的!来帮你生成一段视频 🎬

收到！开始用 Wan 2.6 生成视频 🎥

For music:

好的!来帮你创作一首音乐 🎵

Rules:

• Keep it short and warm (< 15 words)
• Match the user's language (Chinese/English)
• Include relevant emoji (🐱/🎨/🎬/🎵/✨)
• This is your ONLY normal reply — all subsequent updates use message tool

Why this matters:

• Normal replies automatically appear FIRST in the conversation thread
• message tool pushes appear in chronological order AFTER your initial reply
• This ensures users see: "好的!" → "🎨 开始生成..." → "⏳ 进度..." → "✅ 成功!" (correct order)
• Without Step 0, the confirmation might appear LAST, confusing users

Step 1 — Pre-Generation Notification (Push via message tool)

After Step 0 reply, use the message tool to push a notification immediately:

[Emoji] 开始生成 [内容类型]，请稍候…
• 模型：[Model Name]
• 预计耗时：[X ~ Y 秒]
• 消耗积分：[N pts]

Emoji by content type:

• 图片 → 🎨
• 视频 → 🎬（加注:视频生成需要较长时间，我会定时汇报进度）
• 音乐 → 🎵

Cost transparency (new requirement):

• Always show credit cost with model tier context
• For expensive models (>50 pts), offer cheaper alternative proactively
• Examples:
  • Balanced (default): "使用 Wan 2.6（25 积分，最新 Wan）"
  • Premium (user explicit): "使用高端模型 Kling O1（48-120 积分），质量最佳"
  • Premium (auto-selected): "使用 Wan 2.6（25 积分）。若需更高质量可选 Kling O1（48 积分起）"
  • Budget (user asked): "使用 Vidu Q2（5 积分，最省钱）"

Adapt language to match the user (Chinese / English). For video, always add a note that it takes longer. For expensive models, always mention cheaper alternatives unless user explicitly requested premium.

Step 2 — Progress Updates

Poll the task detail API every [Poll Every] seconds per the table.
Send a progress update every [Send Progress Every] seconds.

⏳ 正在生成中… [P]%
已等待 [elapsed]s，预计最长 [max]s

Progress formula:

P = min(95, floor(elapsed_seconds / estimated_max_seconds * 100))

• Cap at 95% — never reach 100% until the API confirms success
• If elapsed > estimated_max: freeze at 95%, append 「快了，稍等一下…」
• For video with max=360s: at 120s → 33%, at 250s → 69%, at 400s → 95% (frozen)

Step 3 — Success Notification

When task status = success:

For Video Tasks (text_to_video / image_to_video / first_last_frame / reference_image)

3.1 Send video player first (IM platforms like Feishu will render inline player):

# Get result URL from script output or task detail API
result = get_task_result(task_id)
video_url = result["medias"][0]["url"]

# Build caption
caption = f"""✅ 视频生成成功！
• 模型：[Model Name]
• 耗时：预计 [X~Y]s，实际 [actual]s
• 消耗积分：[N pts]

[视频描述]"""

# Add mismatch hint if user pref conflicts with knowledge-ai recommendation
if user_pref_exists and knowledge_recommended_model != used_model:
    caption += f"""

💡 提示：当前任务也许用 {knowledge_recommended_model} 也会不错（{reason}，{cost} pts）"""

# Send video with caption (use message tool if available)
message(
    action="send",
    media=video_url,  # ⚠️ Use HTTPS URL directly, NOT local file path
    caption=caption
)

Important:

• Hint is non-intrusive — does NOT interrupt generation
• Only shown when user pref conflicts with knowledge-ai recommendation
• User can ignore the hint; video is already delivered

3.2 Then send link as text (for copying/sharing):

# Send link message immediately after video
message(action="send", text=f"🔗 视频链接（可复制分享）：\n{video_url}")

⚠️ Critical for video:

• Send video player FIRST (inline preview)
• Send text link SECOND (for copying)
• Include first-frame thumbnail URL if available: result["medias"][0]["cover"]

For Image Tasks (text_to_image / image_to_image)

# Build caption
caption = f"""✅ 图片生成成功！
• 模型：[Model Name]
• 耗时：预计 [X~Y]s，实际 [actual]s
• 消耗积分：[N pts]

🔗 原始链接：{image_url}"""

# Add mismatch hint if user pref conflicts with knowledge-ai recommendation
if user_pref_exists and knowledge_recommended_model != used_model:
    caption += f"""

💡 提示：当前任务也许用 {knowledge_recommended_model} 也会不错（{reason}，{cost} pts）"""

# Send image with caption
message(
    action="send",
    media=image_url,
    caption=caption
)

Important:

• Hint is non-intrusive — does NOT interrupt generation
• Only shown when user pref conflicts with knowledge-ai recommendation
• User can ignore the hint; image is already delivered

For Music Tasks (text_to_music)

Send audio file with player:

✅ 音乐生成成功！
• 模型：[Model Name]
• 耗时：预计 [X~Y]s，实际 [actual]s
• 消耗积分：[N pts]
• 时长：约 [duration]

[音频URL或直接发送音频文件]

For TTS Tasks (text_to_speech) — Full UX Protocol (Steps 0–5)

Step 0 — Initial acknowledgment (normal reply)
First reply with a short acknowledgment, e.g.: 好的，正在帮你把这段文字转成语音。 / OK, converting this text to speech.

Step 1 — Pre-generation (message tool)
Push once:

🔊 开始语音合成，请稍候…
• 模型：[Model Name]
• 预计耗时：[X ~ Y 秒]
• 消耗积分：[N pts]

Step 2 — Progress
Poll every 2–5s. Every 10–15s send: ⏳ 语音合成中… [P]%，已等待 [elapsed]s，预计最长 [max]s. Cap progress at 95% until API returns success.

Step 3 — Success (message tool)
When resource_status == 1 and status != "failed", send media = medias[0].url and caption:

✅ 语音合成成功！
• 模型：[Model Name]
• 耗时：实际 [actual]s
• 消耗积分：[N pts]
🔗 原始链接：[url]

Use the URL from the API (do not use local file paths).

Step 4 — Failure (message tool)
On failure, send user-friendly message. TTS error translation (do not expose raw API errors):

Links: API key — https://www.imaclaw.ai/imaclaw/apikey ；Credits — https://www.imaclaw.ai/imaclaw/subscription

Step 5 — Done
After Step 0–4, no further reply needed. Do not send duplicate confirmations.

Step 4 — Failure Notification

When task status = failed or any API/network error, send:

❌ [内容类型]生成失败
• 原因：[natural_language_error_message]
• 建议改用：
  - [Alt Model 1]（[特点]，[N pts]）
  - [Alt Model 2]（[特点]，[N pts]）

需要我帮你用其他模型重试吗？

⚠️ CRITICAL: Error Message Translation

NEVER show technical error messages to users. Always translate API errors into natural language.
API key & credits: 密钥与积分管理入口为 imaclaw.ai（与 imastudio.com 同属 IMA 平台）。Key and subscription management: imaclaw.ai (same IMA platform as imastudio.com).

Generic fallback (when error is unknown):

• Chinese: 生成过程遇到问题，请稍后重试或换个模型试试
• English: Generation encountered an issue, please try again or use another model

Best Practices:

1. Focus on user action: Tell users what to do next, not what went wrong technically
2. Be reassuring: Use phrases like "建议换个模型试试" instead of "失败了"
3. Avoid blame: Never say "你的提示词有问题" → say "提示词需要调整一下"
4. Provide alternatives: Always suggest 1-2 alternative models in the failure message
5. 🆕 Include actionable links (v1.0.8+): For 401/4008 errors, provide clickable links to API key generation or credit purchase pages
6. 🎵 Music-specific (v1.2.0+):
   • For Suno lyrics errors, suggest simplifying lyrics or using auto-generated lyrics (auto_lyrics=true)
   • For prompt length errors, give example length (e.g., "建议20-100字")
   • For BGM requests, recommend DouBao BGM over Suno
7. 🔊 TTS-specific: Use the TTS error translation table in "For TTS Tasks (text_to_speech)" above; suggest another model via --list-models or shortening text.

Step 5 — Done (No Further Action Needed) 🆕

After sending Step 3 (success) or Step 4 (failure):

1. DO NOT send any additional messages unless the user asks a follow-up question
2. The task is complete — wait for the user's next request
3. User preference has been saved (if generation succeeded)
4. The conversation is ready for the next generation request

Why this step matters:

• Prevents unnecessary "anything else?" messages that clutter the chat
• Allows users to naturally continue the conversation when ready
• Respects the asynchronous nature of IM platforms

Exception: If the user explicitly asks "还有别的吗？" or similar, then respond naturally.

🆕 Enhanced Error Handling (v1.0.8):

The Reflection mechanism (3 automatic retries) now provides specific, actionable suggestions for common errors:

• 401 Unauthorized: System suggests generating a new API key with clickable link
• 4008 Insufficient Points: System suggests purchasing credits with clickable link
• 500 Internal Server Error: Automatic parameter degradation (size, resolution, duration, quality)
• 6009 No Rule Match: Automatic parameter completion from credit_rules
• 6010 Attribute Mismatch: Automatic credit_rule reselection
• Timeout: Helpful info with dashboard link for background task status

All error handling is automatic and transparent — users receive natural language explanations with next steps.

Failure fallback by task type:

Music-specific failure guidance:

• If Suno fails → Recommend DouBao BGM (for background music) or DouBao Song (for songs)
• If DouBao BGM fails → Try DouBao Song first (similar pricing), then Suno (more powerful)
• If DouBao Song fails → Try DouBao BGM first (similar pricing), then Suno (more powerful)
• For lyrics errors in Suno → Suggest simplifying lyrics or using auto_lyrics=true
• For prompt length errors → Recommend 20-100 characters

TTS-specific failure guidance:

• If TTS fails → Run --task-type text_to_speech --list-models and suggest another model_id; or shorten text / simplify content. Use the TTS error translation table in "For TTS Tasks" above for user-facing messages.

Supported Models at a Glance

Source: production GET /open/v1/product/list (2026-02-27). Model count reduced significantly. Always query product list API at runtime.

Image Generation (4 models each)

Midjourney attribute_ids: 5451/5452 (text_to_image), 5453/5454 (image_to_image)
Nano Banana2 size options: 512px (4pts), 1K (6pts), 2K (10pts), 4K (13pts)
Nano Banana Pro size options: 1K (10pts), 2K (10pts), 4K (18pts for t2i / 10pts for i2i)

Image Model Capabilities (Parameter Support)

⚠️ Critical: Models have varying parameter support. Custom aspect ratios are now supported by multiple models.

Key Capabilities:

• ✅ Aspect ratio control: SeeDream 4.5 (virtual params), Nano Banana Pro/2 (native support)
• ❌ 8K: Not supported by any model (max is 4K)
• ✅ Size control: Nano Banana2, Nano Banana Pro, and Midjourney support multiple size options via different attribute_ids
• ✅ Budget option: Nano Banana2 is the cheapest at 4 pts for 512px, but 4K costs 13pts
• 🎨 Artistic styles: Midjourney excels at creative, artistic, and illustration styles
• 💡 Best value: SeeDream 4.5 at 5pts offers aspect ratio flexibility; Nano Banana2 512px at 4pts for fastest/cheapest

Video Generation

| text_to_video | SeeDance 1.5 Pro / 1.0 Pro | doubao-seedance-1.5-pro / doubao-seedance-1.0-pro | 16 / 15 pts | | text_to_video | Sora 2 Pro / Sora 2 | sora-2-pro / sora-2 | 120 / 35 pts | | text_to_video | Kling O1 / 2.6 / 2.5 Turbo / 1.6 | kling-video-o1 / kling-v2-6 / kling-v2-5-turbo / kling-v1-6 | 48 / 80 / 24 / 32 pts | | text_to_video | Google Veo 3.1 Fast / 3.1 / 3.0 | veo-3.1-fast-generate-preview / veo-3.1-generate-preview / veo-3.0-generate-preview | 55 / 140 / 280 pts | | text_to_video | Pixverse V3.5–V5.5 | pixverse | 12 pts | | image_to_video | Wan 2.6 / 2.6 Flash / 2.5 / 2.2 Plus | wan2.6-i2v / wan2.6-i2v-flash / wan2.5-i2v-preview / wan2.2-i2v-plus | 25 / 12 / 12 / 10 pts | | image_to_video | Kling 2.1 Master | kling-v2-1-master | 150 pts | | first_last_frame_to_video | Kling O1 | kling-video-o1 | 70 pts | | reference_image_to_video | Kling O1 / Vidu Q2 / Q1 | kling-video-o1 / viduq2 / viduq1 | 48 / 10 / 25 pts |

Music Generation

Speech (TTS) — text_to_speech

Models and credits are not fixed. Always call GET /open/v1/product/list?category=text_to_speech (or run the script with --task-type text_to_speech --list-models) to get current model_id, attribute_id, and credit.

ima-all-ai has complete TTS capability: This document and the bundled ima_create.py provide full TTS support (routing, parameters, create/poll, UX protocol Steps 0–5, error translation). The ima-tts-ai skill is an optional standalone package with the same specification.

TTS Task Detail — Response Shape

Poll POST /open/v1/tasks/detail until completion. For TTS, medias[] uses the same structure as other IMA audio tasks:

Success example: When all medias have resource_status == 1 and status != "failed", read medias[0].url (or watermark_url). Example: {"medias":[{"resource_status":1,"status":"success","url":"https://cdn.../output.mp3","duration_str":"12s","format":"mp3"}]}.

TTS Create Task — Request Shape

task_type: "text_to_speech". No image input: src_img_url: [], input_images: []. prompt (text to speak) must be inside parameters[].parameters, not at top level. Extra fields (e.g. voice_id, speed) come from product form_config; pass via --extra-params and only include params present in the product’s credit_rules/form_config.

TTS Common Mistakes

Always call GET /open/v1/product/list?category=<type> first to get the live attribute_id and form_config defaults required for task creation.

There are two equivalent route systems serving the same backend logic:

This skill documents the /open/v1/ Open API. All business logic (credit validation, N-flattening, risk control) runs identically on both paths.

Environment

Base URL: https://api.imastudio.com

Required/recommended headers for all /open/v1/ endpoints:

Authorization: Bearer ima_your_api_key_here
x-app-source: ima_skills
x_app_language: en

📤 When to Upload Images (Quick Reference)

The IMA Open API does NOT accept raw bytes or base64 images. All image inputs must be public HTTPS URLs.

Upload flow:

1. User provides local file path or bytes → call prepare_image_url() (see section below)
2. User provides HTTPS URL → use directly, no upload needed
3. Use the returned CDN URL (fdl) as the value for input_images / src_img_url

Example workflow (image_to_image):

# User provides local file
image_url = prepare_image_url("/path/to/photo.jpg", api_key)
# → Returns: https://ima-ga.esxscloud.com/webAgent/privite/2026/02/27/..._uuid.jpeg

# Then create task with this URL
create_task(
    task_type="image_to_image",
    input_images=[image_url],  # Use uploaded URL
    prompt="turn into oil painting"
)

⚠️ MANDATORY: Always Query Product List First

CRITICAL: You MUST call /open/v1/product/list BEFORE creating any task.
The attribute_id field is REQUIRED in the create request. If it is 0 or missing, you get:
"Invalid product attribute" → "Insufficient points" → task fails completely.
NEVER construct a create request from the model table alone. Always fetch the product first.

How to get attribute_id (all task types)

# Query product list with the correct category
GET /open/v1/product/list?app=ima&platform=web&category=<task_type>
# task_type: text_to_image | image_to_image | text_to_video | image_to_video |
#            first_last_frame_to_video | reference_image_to_video | text_to_music | text_to_speech

# Walk the V2 tree to find your target model (type=3 leaf nodes only)
for group in response["data"]:
    for version in group.get("children", []):
        if version["type"] == "3" and version["model_id"] == target_model_id:
            attribute_id  = version["credit_rules"][0]["attribute_id"]
            credit        = version["credit_rules"][0]["points"]
            model_version = version["id"]    # = version_id / model_version
            model_name    = version["name"]
            form_defaults = {f["field"]: f["value"] for f in version["form_config"]}
            break

Quick Reference: Known attribute_ids

Pre-queried values for convenience. Always call the product list at runtime for accuracy.

⚠️ Production warning: attribute_id and credit values change frequently in production. Always call /open/v1/product/list at runtime; above table is pre-queried reference only (2026-02-27).

Common Mistakes (and resulting errors)

⚠️ Critical for Google Veo 3.1 and multi-rule models:

Models like Google Veo 3.1 have multiple credit_rules, each with a different attribute_id for different parameter combinations:

• 720p + 4s + optimized → attribute_id A
• 720p + 8s + optimized → attribute_id B
• 4K + 4s + high → attribute_id C

The script automatically selects the correct attribute_id by matching your parameters (duration, resolution, compression_quality, generate_audio) against each rule's attributes. If the match fails, you get error 6010.

Fix: The bundled script now checks these video-specific parameters for smart credit_rule selection. Always use the script, not manual API construction.

Core Flow

1. GET /open/v1/product/list?app=ima&platform=web&category=<type>
   → REQUIRED: Get attribute_id, credit, model_version, model_name, form_config defaults

[If input image required]
2. Upload image → get public HTTPS URL
   → See "Image Upload" section below

3. POST /open/v1/tasks/create
   → Must include: attribute_id, model_name, model_version, credit, cast, prompt (nested!)

4. POST /open/v1/tasks/detail  {"task_id": "..."}
   → Poll until medias[].resource_status == 1
   → Extract url from completed media

Image Upload (Required Before Image Tasks)

The IMA Open API does NOT accept raw bytes or base64 images. All image inputs must be public HTTPS URLs.

When a user provides an image (local file, bytes, base64), you must upload it first and get a URL. This is exactly what the IMA frontend does before every image task.

Real Upload Flow (from IMA Frontend Source)

The frontend uses a two-step presigned URL flow via the IM platform:

Step 1: GET /api/rest/oss/getuploadtoken   → returns { ful, fdl }
          ful = presigned PUT URL (upload destination, expires ~7 days)
          fdl = final CDN download URL (use this as input_images value)

Step 2: PUT {ful}  with raw image bytes + Content-Type header
          → image is stored in Aliyun OSS: zhubite-imagent-bot.oss-us-east-1.aliyuncs.com
          → accessible via CDN: https://ima-ga.esxscloud.com/...

Step 1: Get Upload Token

GET https://imapi-qa.liveme.com/api/rest/oss/getuploadtoken

Required query parameters (11 total — sourced directly from frontend generateUploadInfo):

简化认证：直接使用 IMA API key 填充 appUid 和 cmimToken 参数，无需单独获取凭证。

Response:

{
  "ful": "https://zhubite-imagent-bot.oss-us-east-1.aliyuncs.com/webAgent/privite/2026/02/26/..._uuid.jpeg?Expires=...&OSSAccessKeyId=...&Signature=...",
  "fdl": "https://ima-ga.esxscloud.com/webAgent/privite/2026/02/26/..._uuid.jpeg",
  "ful_expire": "...",
  "fdl_expire": "...",
  "fdl_key": "..."
}

Step 2: Upload Image via Presigned URL

PUT {ful}
Content-Type: image/jpeg
Body: [raw image bytes]

No auth headers needed — the presigned URL already encodes the credentials.

Step 3: Use fdl as the Image URL

After the PUT succeeds, use fdl (the CDN URL) as the value for input_images / src_img_url.

Python Implementation

import hashlib, time, uuid, requests, mimetypes

# ── 🌐 IMA Upload Service Endpoint (IMA-owned, for image/video uploads) ──────
IMA_IM_BASE = "https://imapi-qa.liveme.com"   # prod: https://imapi.liveme.com

# ── 🔑 Hardcoded APP_KEY (Public, Shared Across All Users) ──────────────────
# This APP_KEY is a PUBLIC identifier used by IMA Studio's image/video upload 
# service. It is NOT a secret—it's intentionally shared across all users and 
# embedded in the IMA web frontend. This key is used to generate HMAC signatures 
# for upload token requests, but your IMA API key (ima_xxx...) is the ACTUAL 
# authentication credential. Think of APP_KEY as a "client ID" rather than a 
# "client secret."
#
# ⚠️ Security Note: Your ima_xxx... API key is the sensitive credential. It is 
# sent to imapi.liveme.com as query parameters (appUid, cmimToken). Always use 
# test keys for experiments and rotate your API key regularly.
#
# 📖 See SECURITY.md for complete disclosure and network verification guide.
APP_ID    = "webAgent"
APP_KEY   = "32jdskjdk320eew"   # Public shared key (used for HMAC sign generation)
APP_UID   = "<your_app_uid>"    # POST /api/v3/login/app → data.user_id
APP_TOKEN = "<your_app_token>"  # POST /api/v3/login/app → data.token


def _gen_sign() -> tuple[str, str, str]:
    """Generate per-request (sign, timestamp, nonce)."""
    nonce = uuid.uuid4().hex[:21]
    ts    = str(int(time.time()))
    raw   = f"{APP_ID}|{APP_KEY}|{ts}|{nonce}"
    sign  = hashlib.sha1(raw.encode()).hexdigest().upper()
    return sign, ts, nonce


def get_upload_token(app_uid: str, app_token: str,
                     suffix: str, content_type: str) -> dict:
    """Step 1: Get presigned upload URL from IMA's upload service.
    
    Calls GET imapi.liveme.com/api/rest/oss/getuploadtoken with exactly 11 params.
    Returns: { "ful": "<presigned PUT URL>", "fdl": "<CDN download URL>" }
    
    Args:
        app_uid: Your IMA API key (ima_xxx...), used as appUid parameter
        app_token: Your IMA API key (ima_xxx...), used as cmimToken parameter
        suffix: File extension (jpeg, png, mp4, mp3)
        content_type: MIME type (image/jpeg, video/mp4, etc.)
    
    Security Note:
        Your IMA API key (ima_xxx...) is sent to imapi.liveme.com as query 
        parameters (appUid, cmimToken). This is IMA Studio's image/video upload 
        service, separate from the main api.imastudio.com API. Both domains are 
        owned by IMA Studio—this is part of IMA's microservices architecture.
        
        Why two domains?
        - api.imastudio.com: Core AI generation API (product list, task creation)
        - imapi.liveme.com: Specialized upload service (presigned URL generation)
        
        Your API key grants access to both services. For security verification, 
        see SECURITY.md section "Network Traffic Verification."
    """
    sign, ts, nonce = _gen_sign()
    r = requests.get(
        f"{IMA_IM_BASE}/api/rest/oss/getuploadtoken",
        params={
            # App Key params
            "appUid":       app_uid,       # APP_UID
            "appId":        APP_ID,
            "appKey":       APP_KEY,
            "cmimToken":    app_token,     # APP_TOKEN
            "sign":         sign,
            "timestamp":    ts,
            "nonce":        nonce,
            # File params
            "fService":     "privite",     # fixed
            "fType":        "picture",     # picture / video / audio
            "fSuffix":      suffix,        # jpeg / png / mp4 / mp3
            "fContentType": content_type,
        },
    )
    r.raise_for_status()
    return r.json()["data"]


def upload_image_to_oss(image_bytes: bytes, content_type: str, ful: str) -> None:
    """Step 2: PUT image bytes to the presigned OSS URL. No auth needed."""
    resp = requests.put(ful, data=image_bytes, headers={"Content-Type": content_type})
    resp.raise_for_status()


def prepare_image_url(source, api_key: str) -> str:
    """
    Full workflow: upload any image and return the CDN URL (fdl).
    
    Args:
        source: file path (str), raw bytes, or already-public HTTPS URL
        api_key: IMA API key for upload authentication
    
    Returns: public HTTPS CDN URL ready to use as input_images value
    """
    # Already a public URL → use directly, no upload needed
    if isinstance(source, str) and source.startswith("https://"):
        return source
    
    # Read file bytes
    if isinstance(source, str):
        ext = source.rsplit(".", 1)[-1].lower() if "." in source else "jpeg"
        with open(source, "rb") as f:
            image_bytes = f.read()
        content_type = mimetypes.guess_type(source)[0] or "image/jpeg"
    else:
        image_bytes = source
        ext = "jpeg"
        content_type = "image/jpeg"

    # Step 1: Get presigned URL using API key directly
    token_data = get_upload_token(api_key, ext, content_type)
    ful = token_data["ful"]
    fdl = token_data["fdl"]

    # Step 2: Upload to OSS
    upload_image_to_oss(image_bytes, content_type, ful)

    # Step 3: Return CDN URL
    return fdl   # use this as input_images / src_img_url value

OSS path format: webAgent/privite/{YYYY}/{MM}/{DD}/{timestamp}_{uid}_{uuid}.{ext} CDN base: https://ima-ga.esxscloud.com/ OSS bucket: zhubite-imagent-bot.oss-us-east-1.aliyuncs.com

Quick Reference

Task Types (category values)

Detail API status values

Each media in medias[] has two fields:

Poll on resource_status first, then ensure status is not "failed":

Important: (1) Treat resource_status: null as 0. (2) Stop only when all medias have resource_status == 1. (3) When resource_status=1, still check status != "failed".

API 1: Product List

GET /open/v1/product/list?app=ima&platform=web&category=text_to_image

Internally calls downstream /v1/products/listv2. Returns a V2 tree structure: type=2 nodes are model groups, type=3 nodes are versions (leaves). Only type=3 nodes contain credit_rules and form_config.

webAgent is auto-converted to ima by the gateway — you can use either value for app.

[
  {
    "id": "SeeDream",
    "type": "2",
    "name": "SeeDream",
    "model_id": "",
    "children": [
      {
        "id": "doubao-seedream-4-0-250828",
        "type": "3",
        "name": "SeeDream 4.0",
        "model_id": "doubao-seedream-4.0",
        "credit_rules": [
          { "attribute_id": 332, "points": 5, "attributes": { "default": "enabled" } }
        ],
        "form_config": [
          { "field": "size", "type": "tags", "value": "1K",
            "options": [{"label":"1K","value":"1K"}, {"label":"2K","value":"2K"}] }
        ]
      }
    ]
  }
]

How to pick a version for task creation:

1. Traverse nodes to find type=3 leaves (versions)
2. Use model_id and id (= model_version) from the leaf
3. Pick credit_rules[].attribute_id matching your desired quality/size (attributes field shows the config)
4. Use form_config[].value as default parameters values

credit_rules[].attribute_id → required for task creation as attribute_id. credit_rules[].points → required for task creation as credit and cast.points.

API 2: Create Task

POST /open/v1/tasks/create

Request Structure

{
  "task_type": "text_to_image",
  "enable_multi_model": false,
  "src_img_url": [],
  "upload_img_src": "",
  "parameters": [
    {
      "attribute_id": 8538,
      "model_id":      "doubao-seedream-4.5",
      "model_name":    "SeeDream 4.5",
      "model_version": "doubao-seedream-4-5-251128",
      "app":           "ima",
      "platform":      "web",
      "category":      "text_to_image",
      "credit":        5,
      "parameters": {
        "prompt":       "a beautiful mountain sunset, photorealistic",
        "size":         "4k",
        "n":            1,
        "input_images": [],
        "cast":         {"points": 5, "attribute_id": 8538}
      }
    }
  ]
}

Field Reference

N-Field Flattening (Gateway Internal Logic)

When n > 1, the gateway automatically:

1. Generates n independent resourceBizId values
2. Deducts credits n times (one per resource)
3. Creates n separate tasks in the downstream service

Response medias[] will contain n items. Poll until all have resource_status == 1.

Response

{
  "code": 0,
  "data": {
    "id": "task_abc123",
    "biz_id": "biz_xxx",
    "task_type": "text_to_image",
    "medias": [],
    "generate_count": 1,
    "created_at": 1700000000000,
    "timeout_at": 1700000300000
  }
}

data.id = task ID for polling. timeout_at = Unix ms deadline.

API 3: Task Detail (Poll)

POST /open/v1/tasks/detail
{"task_id": "<id from create response>"}

Poll every 2–5s (8s+ for video). Completed response:

{
  "id": "task_abc",
  "medias": [{
    "resource_status": 1,
    "status": "success",
    "url": "https://cdn.../output.jpg",
    "cover": "https://cdn.../cover.jpg",
    "format": "jpg",
    "width": 1024,
    "height": 1024
  }]
}

Polling stop condition (must implement exactly):

• Treat resource_status: null (or missing) as 0 (processing). Do not stop when you see null; backend may serialize Go *int as null.
• Stop only when ALL medias[].resource_status == 1 and no status == "failed". If you return on the first media with resource_status == 1 while others are still 0, the task is not fully done and you will keep polling or get inconsistent state.
• Stop immediately if any status == "failed" or resource_status == 2 or resource_status == 3.

Task Type Examples

text_to_image ✅ Verified

No image input. src_img_url: [], input_images: []. See API 2 for full example.

text_to_video ✅ Verified

Extra fields vs text_to_image — all from form_config defaults:

{
  "task_type": "text_to_video",
  "src_img_url": [],
  "parameters": [{
    "attribute_id":  4838,
    "model_id":      "wan2.6-t2v",
    "model_name":    "Wan 2.6",
    "model_version": "wan2.6-t2v",
    "category":      "text_to_video",
    "credit":        3,
    "app": "ima", "platform": "web",
    "parameters": {
      "prompt":          "a puppy dancing happily, sunny meadow",
      "negative_prompt": "",
      "prompt_extend":   false,
      "duration":        5,
      "resolution":      "1080P",
      "aspect_ratio":    "16:9",
      "shot_type":       "single",
      "seed":            -1,
      "n":               1,
      "input_images":    [],
      "cast":            {"points": 3, "attribute_id": 4838}
    }
  }]
}

Video-specific fields from form_config: duration (seconds), resolution, aspect_ratio, shot_type, negative_prompt, prompt_extend. Poll every 8s (video generation is slower). Response medias[].cover = first-frame thumbnail.

text_to_music

No image input. src_img_url: [], input_images: [].

image_to_image ✅ Verified

{
  "task_type": "image_to_image",
  "src_img_url": ["https://...input.jpg"],
  "parameters": [{
    "attribute_id":  8560,
    "model_id":      "doubao-seedream-4.5",
    "model_version": "doubao-seedream-4-5-251128",
    "category":      "image_to_image",
    "credit":        5,
    "app": "ima", "platform": "web",
    "parameters": {
      "prompt":       "turn into oil painting style",
      "size":         "4k",
      "n":            1,
      "input_images": ["https://...input.jpg"],
      "cast":         {"points": 5, "attribute_id": 8560}
    }
  }]
}

⚠️ size must be from form_config options (e.g. "2k", "4k", "2048x2048"). "adaptive" is NOT valid for SeeDream 4.5 i2i — causes error 400. Top-level src_img_url and parameters.input_images must both contain the input image URL. Some i2i models (e.g. doubao-seededit-3.0-i2i) may not be available in test environments — fall back to SeeDream 4.5.

image_to_video / first_last_frame_to_video / reference_image_to_video

{
  "src_img_url": ["https://first-frame.jpg", "https://last-frame.jpg"]
}

Index 0 = first frame (or reference), index 1 = last frame (first_last_frame only).

Common Mistakes

Complete Python Example

See the Python example sections throughout this documentation for implementation guidance covering all 7 task types.

Supported Models & Search Terms

Image: SeeDream 4.5 (see dream), Midjourney (MJ), Nano Banana 2, Nano Banana Pro Video: Wan 2.6, Kling O1, Kling 2.6, Google Veo 3.1 (veo), Sora 2 Pro, Pixverse V5.5, Hailuo 2.0, Hailuo 2.3, MiniMax Hailuo, SeeDance 1.5 Pro, Vidu Q2 Music: Suno sonic v4, Suno sonic v5, DouBao BGM (GenBGM), DouBao Song (GenSong) TTS: seed-tts-2.0 (seed tts, text-to-speech)

Capabilities: multimodal AI creation, all-in-one, image generation, video generation, music generation, text-to-speech, text-to-image, image-to-video, text-to-music