fashion-video-creator

# Skill: fashion-video-creator ## Overview Generate complete assets for fashion e-commerce videos from scratch: a virtual model image (Seedream text-to-image) + a Seedance 2.0 video prompt with dialogue, actions, and scene + an operator SOP for the Jimeng platform. Supports single and batch modes. ## When to Activate User query contains any of: - "穿搭视频", "outfit video", "生成模特", "虚拟模特" - "seedance prompt", "穿搭创作", "带货视频", "带货短视频" - "帮我做个穿搭视频", "批量生成穿搭", "batch prompt" - "帮我拍个带货视频", "我有件衣服想做个展示视频", "做个穿搭短视频" - "generate fashion video", "create outfit content" Do NOT activate for: - Analyzing / reverse-engineering existing videos -> use `viral-video-replicator` - "这个视频怎么拍的", "帮我复刻这个视频" -> use `viral-video-replicator` - Pure image generation without video prompt -> direct Seedream API call - Non-fashion video generation -> not applicable ## Prerequisites This skill requires a Volcano Engine (火山方舟) account with Seedream model access. ``` Required: ARK_API_KEY + Seedream model/endpoint ID Optional: None (all other logic is local) ``` ## Clarification Flow ### Phase 1: API Key Acquisition Ask these questions IN ORDER. Use plain language — the user may not be technical. **Q1: Image Generation Service** > "你需要先有一个能生成图片的 AI 服务账号。目前这个工具使用的是火山方舟平台的 Seedream 模型。 > 你有火山方舟的账号吗？如果有，请提供你的 API Key（在火山方舟控制台 -> API Key 管理中可以找到）。" If user has no account -> STOP. Guide them to register at 火山方舟. Do NOT proceed without API key. **Q2: Model Version** > "火山方舟上有两个版本的图片生成模型： > - **Seedream 5.0**（推荐）：理解能力更强，生成的人像更自然，每张约 0.22 元 > - **Seedream 4.5**：可以更精细地控制细节，每张约 0.32 元 > > 你开通了哪个？需要提供模型ID或接入点ID。" ### Phase 2: Mandatory Recommendations MUST show before proceeding. These are not optional tips — violating them degrades output quality: ``` ============================================================ API Configuration — Mandatory Recommendations ============================================================ [REQUIRED] Seedream 5.0 endpoint preferred WHY: 5.0 uses internal Chain-of-Thought reasoning, making natural language prompts significantly more effective. 4.5 relies on keyword stacking which loses semantics in complex body descriptions. If only 4.5: use model ID doubao-seedream-4-5-251128 or newer. [REQUIRED] Realism scale 40-60 WHY: Below 30 the output looks cartoonish (Pixar-style), unsuitable for product videos that need believable human models. Above 80 triggers uncanny valley artifacts and skin texture glitches. [WARNING] Using older models or non-recommended endpoints will produce noticeably worse portrait realism and clothing detail accuracy. ============================================================ ``` ### Phase 3: Mode Selection > "你要生成几条视频？单条还是批量？" ### Phase 4: Creative Parameter Collection **Collection strategy — ask in 2 rounds, never more than 4 questions per round:** **Round 1 (required — must have before executing):** 1. Model gender (female/male) — default: female 2. Garment type (dress/top/pants/jacket/suit/casual) — NO default, must ask. WHY: "default" type produces generic dialogue without product-specific selling points. **Defensive fallback:** If user declines to specify after 2 attempts, fall back to "default" with warning: "Using generic dialogue — conversion effectiveness will be lower than type-specific scripts." 3. Scene (11 presets or custom text) — default: modern_apartment **Round 2 (enhanced — use defaults if not stated):** 4. Model appearance: preset (9 templates) / body params / custom desc — default: asian_female_slim 5. Realism (0-100) — default: 40 6. Style: realism / pixar / anime_realistic — default: realism 7. Dialogue style: natural / professional / enthusiastic — default: natural 8. Camera: vlog / pro / static — default: vlog 9. Target audience: auto / female / male — default: auto 10. Duration (5-40s) — default: 10 11. Language: zh / en — default: zh 12. Task label (batch only) **Batch shortcut:** If all tasks share config, ask once and apply to all. Only collect per-task differences. ### Batch-Specific Recommendations When batch mode selected, ADDITIONALLY show: ``` ============================================================ Batch Mode — Additional Recommendations ============================================================ [REQUIRED] Specify exact garment type per task — do NOT use "default" WHY: Each garment type has 500+ chars of specialized dialogue with product-specific hand gestures (e.g., dress: pull hem to show fabric; pants: flip waistband to show grip strip). "default" loses all this. [RECOMMENDED] Same gender per batch WHY: Mixing genders creates inconsistent dialogue tone — female uses filler words and emotional reactions, male is direct and factual. [REQUIRED] Duration 10-15 seconds WHY: Under 5s cannot complete a product showcase. Over 15s triggers multi-segment auto-chaining which requires manual "extend" operations. ============================================================ ``` ## Core Workflow ### Step 0: Verify API Key (mandatory, never skip) Platform-adaptive verification: **If bash/Python available (Claude Code, terminal):** ```python import httpx resp = httpx.get(f"{ARK_API_BASE}/api/v3/models", headers={"Authorization": f"Bearer {ARK_API_KEY}"}, timeout=10) # 200 -> proceed. 401/403 -> key invalid. Timeout -> network issue. ``` **If no code execution (Claude.ai web, chat-only):** Trust the user-provided key and proceed. Mark internally: `api_verified: false`. If Step 3 (image generation) fails with 401 -> surface the error then. - Key valid -> proceed to Step 1 - Key invalid (401/403) -> report error, ask user to verify. Do NOT proceed. - No code execution -> proceed with unverified key, validate on first API call. ### Single Mode ``` Step 1: Collect parameters (2 rounds max) Step 2: Build model prompt (version-aware: 4.5 keyword / 5.0 natural language) Step 3: Call Seedream API -> generate model image (720x1280, 9:16) Step 4: Build Seedance prompt -> compose dialogue + actions + scene + camera Read references/dialogue-library.md for garment-specific scripts Read references/prompt-assembly.md for assembly rules Step 5: Generate SOP Step 6: If duration > 15s -> build chained multi-segment extend plan Step 7: Validate output (see below) ``` ### Batch Mode ``` Step 0: Verify API key Step 1: Collect batch size + per-task configs (2 rounds) Step 2: Validate all configs (garment type != "default"?) Step 3: For each task (sequential): a. Build model prompt b. Call Seedream API (if image needed) c. Compose dialogue (opening + core + closing) from dialogue-library.md d. Compose actions (garment-specific choreography) e. Assemble Seedance prompt f. If duration > 15s: build chained segments g. Generate SOP h. Mark task: completed / failed Step 4: Validate all outputs Step 5: Return results with progress summary Progress: queued -> processing -> completed/failed Partial success: batch completes even if some tasks fail. ``` ### Output Validation (mandatory, never skip) Before delivering results, verify ALL: - [ ] Model image is 9:16 portrait with full body visible (head to feet)? - [ ] Prompt contains @image1 (garment) and @image2 (model ref)? - [ ] Dialogue matches the selected garment type (not generic "default")? - [ ] SOP specifies correct image upload order (image1=garment FIRST)? - [ ] If chained: every segment has continuation prompt + dialogue preview? - [ ] Brand constraint present: "Do not alter clothing pattern, color, texture or style"? **Any NO -> fix before delivering. Do NOT send unvalidated output.** ## Error Handling | Failure | Detection | Action | |---------|-----------|--------| | No API key | ARK_API_KEY empty or missing | **STOP.** Guide user to 火山方舟 console. Do NOT proceed. | | Invalid API key | 401/403 from API | Report error. Ask user to verify key. Do NOT retry with guessed keys. | | Seedream timeout | No response in 300s | Retry once with same prompt. Still fails -> report with the prompt used so user can try manually. | | No image in response | API returns empty data | Simplify prompt (remove extras), retry. Still empty -> report error. | | Batch task fails | Exception during prompt assembly | Mark task as failed with error message. Continue remaining tasks. Report partial results at end. | | Invalid garment type | User provides unknown type | Map to closest valid type or ask for clarification. Valid: dress/top/pants/jacket/suit/casual/default. | ### Degraded Mode: Prompt-Only (when image generation fails) If Seedream API is unreachable or fails after retry, do NOT block the entire workflow. **Skip image generation, still deliver prompt + SOP.** Mark output clearly: ``` [Note] Model image generation unavailable (API error). Prompt and SOP generated successfully. Use your own model photo as @image2 when operating on the Jimeng platform. ``` This ensures the user gets the valuable prompt + SOP even without the AI-generated model image. ## Usage Example **Input:** "帮我做一个女性穿搭视频，裙子类型，现代公寓场景" **Resolved params:** gender=female, garment=dress, scene=modern_apartment, realism=40, style=realism, dialogue=natural, camera=vlog, audience=auto->female, duration=10, lang=zh **Output 1 — Model Image:** 720x1280 PNG (asian_female_slim preset, realism=40) **Output 2 — Seedance Prompt:** ``` 一位面容和身材参考@图片2的年轻女性穿着@图片1中的服装，在简约现代的白色公寓客厅中，自然光从落地窗照入。她面朝镜头表情生动自然地展示服装，右手拉起裙摆展示面料（始终用右手），右手翻开裙摆内侧展示车线做工，小幅转身让裙摆飘动，右手捏腰部展示松紧设计，对着镜头说：「姐妹们你们快看...哇，不是，我真的没想到这条裙子上身效果这么好。你看这个面料，是那种...嗯...醋酸缎面的...（右手轻轻拉起裙摆展示）滑滑的凉凉的，而且有一定的厚度，不是那种廉价的薄纱感。然后你看这个车线...（右手翻开裙子内侧给镜头看）全部是包边走线的，没有毛边，做工真的很扎实。裙摆是A字的微微伞摆，你看我转一下...（转了一圈）你看它飘起来那个弧度，而且腰线这里有一个隐藏的松紧设计...（右手捏了捏腰部）不会勒但又收腰。你们猜多少钱？不到两百！超显腿长，闭眼入。」语气自然亲切，像在跟闺蜜视频通话。说话有停顿、有喘息、偶尔磕巴自我纠正，真实感强。Do not alter clothing pattern, color, texture or style. 手持vlog镜头感，竖屏9:16构图。音频要求：全程只有模特一个人的清晰人声，不出现第二个人的声音或对白。背景有符合场景的自然环境音，音量不超过人声的10%。视觉要求：全程保持中景到中近景，不要切到手部或脚部的特写镜头，避免手指变形问题。模特展示服装细节时始终使用右手操作，避免左右手切换导致的镜像翻转。全程画面中只有模特一人，不出现其他人物。 ``` **Output 3 — Operator SOP:** Step-by-step for 即梦 platform (upload order, settings, quality checklist) ## Domain Knowledge Role Declaration > The reference files contain dialogue scripts, model presets, API specs, and prompt templates. > Their role is to **assist prompt assembly** — providing the raw materials that get composed into Seedance prompts. > They do NOT replace the execution workflow. Never output reference content directly as the final answer. > Always assemble through the workflow: collect params -> build prompt -> validate -> deliver. ## References | File | Purpose | When to read | |------|---------|-------------| | [references/model-presets.md](references/model-presets.md) | 9 model presets, body params, 17-level realism anchors, scene/camera/style presets | Step 2: building model prompt | | [references/seedream-api.md](references/seedream-api.md) | Seedream 4.5/5.0 API endpoint, request format, cost, crop logic | Step 3: calling API | | [references/prompt-assembly.md](references/prompt-assembly.md) | Prompt composition order, multi-segment chaining constants, SOP templates | Step 4-6: assembling prompt and SOP | | [references/dialogue-library.md](references/dialogue-library.md) | Complete dialogue scripts: 7 garment types x 2 genders x 2 languages (28 scripts total), openings, closings, actions | Step 4: composing dialogue content |