avatar-studio

Files (48)

• CHANGELOG.md
• README.md
• SKILL.md
• __init__.py
• avatar_comfy_client.py
• avatar_dashscope_client.py
• avatar_model_registry.py
• avatar_models.py
• avatar_pipeline.py
• __init__.py · avatar_studio_inline/
• dep_bootstrap.py · avatar_studio_inline/
• llm_json_parser.py · avatar_studio_inline/
• oss_uploader.py · avatar_studio_inline/
• parallel_executor.py · avatar_studio_inline/
• storage_stats.py · avatar_studio_inline/
• upload_preview.py · avatar_studio_inline/
• vendor_client.py · avatar_studio_inline/
• avatar_task_manager.py
• avatar_tts_edge.py
• icon.svg
• plugin.json
• plugin.py
• requirements.txt
• __init__.py · tests/
• conftest.py · tests/
• .gitkeep · tests/fixtures/
• .gitkeep · tests/integration/
• __init__.py · tests/integration/
• test_dashscope_smoke.py · tests/integration/
• test_comfy_client.py · tests/
• test_dashscope_client.py · tests/
• test_model_registry.py · tests/
• test_models.py · tests/
• test_pipeline.py · tests/
• test_plugin.py · tests/
• test_relay_integration.py · tests/
• test_smoke.py · tests/
• test_task_manager.py · tests/
• test_tts_edge.py · tests/
• bootstrap.js · ui/dist/_assets/
• i18n.js · ui/dist/_assets/
• icons.js · ui/dist/_assets/
• markdown-mini.js · ui/dist/_assets/
• styles.css · ui/dist/_assets/
• icon.svg · ui/dist/
• index.html · ui/dist/
• README.md · workflows/
• recommended.json · workflows/

SKILL.md (first lines)

---
name: avatar-studio
description: Generate digital-human speaking photos, relip/reface videos, voices, and avatar compositions through Avatar Studio.
risk_class: network_out
---

# avatar-studio · Cursor Skill Card

> 一线插件。DashScope 万相 + s2v + cosyvoice + qwen-vl 全链路。

## 1 · 何时触发我

代理在用户出现以下意图时**优先**调用 avatar-studio 的工具，而不是手写
ffmpeg / ComfyUI 脚本或调其他视觉插件：

- "让这张照片说话 / 帮我做一个数字人开场白"
- "我录了一段视频，能不能换个嘴型 / 配音口型"
- "这段视频里的人换成另一个人"
- "把这两张参考图融合成一个新的角色，让他说一段话"
- "估算一下做 5 秒数字人视频要多少钱"

**不**应触发 avatar-studio：

- 纯文生视频（无人脸）→ 调 `seedance-video`
- 纯文生图 → 调 `tongyi-image`
- 仅做 TTS、不出视频 → 直接调 `cosyvoice` 工具或主端 TTS

## 2 · 工具清单

| 工具名 | 用途 | 关键参数 |
|---|---|---|
| `avatar_cost_preview` | 估算费用，**不**实际提交 | `mode`, 各 mode 必需的素材或 `text_chars` / `audio_duration_sec` |
| `avatar_photo_speak` | 创建照片说话任务 | `image_url` 或 `assets.image`, `text` 或 `audio_url`, `voice_id`, `resolution` |
| `avatar_video_relip` | 创建视频换嘴任务 | `video_url` 或 `assets.video`, `text` 或 `audio_url`, `voice_id` |
| `avatar_video_reface` | 创建视频换人任务 | `image_url`, `video_url`, `mode_pro` (bool), `watermark` (bool) |
| `avatar_compose` | 创建数字人合成任务 | `ref_images_url[]` (1-3 张), `prompt`, `text`, `voice_id`, `resolution` |
| `avatar_voice_create` | 注册自定义克隆音色 | `label`, `source_audio_path`, `dashscope_voice_id` |
| `avatar_voice_delete` | 删除自定义音色 | `voice_id` |
| `avatar_figure_create` | 注册自定义形象 | `label`, `image_path`, `preview_url` |
| `avatar_figure_delete` | 删除自定义形象 | `figure_id` |

每个 mode 工具返回 `任务已创建：{id}（mode=...）`；后续状态走 SSE
`plugin:avatar-studio:task_update` 推送，UI 端 Tasks Tab 自动更新。

## 3 · 输入 schema 速查

所有 `POST` body 都用 Pydantic + `extra="forbid"` 严格校验（落实 Pixelle
C6，不会静默丢字段）。下面是 4 个生成 mode 在 `/tasks` 端点的差异：

```jsonc
// photo_speak
{
  "mode": "photo_speak",
  "image_url": "https://...",
  // 二选一：text + voice_id  OR  audio_url
  "text": "你好",
  "voice_id": "longxiaochun_v2",
  // "audio_url": "https://...",
  "resolution": "480P",                  // 480P | 720P
  "audio_duration_sec": 3.0,             // 估价用；s2v 自身由音频时长决定
  "cost_approved": false                 // > threshold 时需为 true
}

// video_relip
{
  "mode": "video_relip",
  "video_url": "https://...",
  "text": "你好", "voice_id": "longxiaochun_v2"
  // 或 "audio_url"
}

// video_reface（最贵 — 务必先 cost-preview）
{
  "mode": "video_reface",
  "image_url": "https://...",
  "video_url": "https://...",
  "mode_pro": false,                     // false=wan-std ¥0.6/s，true=wan-pro ¥1.2/s
  "watermark": true,
  "video_duration_sec": 5,
  "cost_approved": true
}

// avatar_compose（多图 + 可选 AI prompt 助手）
{
  "mode": "avatar_compose",
  "ref_images_url": ["https://...", "https://..."],   // 1-3 张
  "prompt": "把人物融入古风街景，保留人物五官",
  "text": "你好", "voice_id": "longxiaochun_v2",
  "resolution": "480P",
  "audio_duration_sec": 5
}
```

> Tip：让 `qwen-vl-max` 替你写 `prompt` —— 调
> `POST /api/plugins/avatar-studio/ai/compose-prompt`，body
> `{"ref_images_url":[...], "user_intent":"古风街景"}`，返回 `{prompt}`。

## 4 · 输出 schema 速查

```jsonc
// POST /tasks
{ "ok": true, "task": { "id": "tsk_xxx", "mode": "...", "status": "pending", ... } }

// GET /tasks/{id}（任务跑完后）
{
  "ok": true,
  "task": {
    "id": "tsk_xxx",
    "mode": "photo_speak",
    "status": "succeeded",                 // pending | submitted | polling | succeeded | failed | cancelled
    "output_path": "/.../output.mp4",
    "output_url": "/api/plugins/avatar-studio/uploads/preview/...",
    "video_duration_sec": 3.2,
    "cost_breakdown": {
      "currency": "CNY",
      "items": [
        { "name": "wan2.2-s2v-detect", "units": 1, "unit_price": 0.004, "subtotal": 0.004 },
        { "name": "wan2.2-s2v 480P", "units": 3.0, "unit_price": 0.10, "subtotal": 0.30 },
        { "name": "cosyvoice-v2", "units": 0.0018, "unit_price": 0.20, "subtotal": 0.0036 }

From LZM163126lzm/openakita, AGPL-3.0.

More from LZM163126lzm/openakita