feat: 增加gpt-image-2 模型

2026-08-17 17:28:02 +08:00
parent 9d15d8f784
commit e9d1eef07e
11 changed files with 162 additions and 17 deletions
@@ -27,6 +27,13 @@ DASHSCOPE_API_KEY=
 DASHSCOPE_BASE_URL=
 DASHSCOPE_MODEL=wan2.7-image-pro

+# --- RightAPI（gpt-image，OpenAI 兼容中转）---
+RIGHTAPI_API_KEY=
+RIGHTAPI_BASE_URL=https://rightapi.ai/draw
+RIGHTAPI_IMAGE_MODEL=gpt-image-2
+# 出图质量：auto | low | medium | high（high 单张约 1-5 分钟，超时自动兜底 600s）
+RIGHTAPI_IMAGE_QUALITY=high
+
 # --- DeepSeek（出图方案规划器）---
 DEEPSEEK_API_KEY=
 DEEPSEEK_BASE_URL=https://api.deepseek.com/v1
@@ -41,6 +41,8 @@ Chrome 插件（WXT + React + antd）          Python 后端（FastAPI + SQLite
 - `generator.py`：图像 provider（图生图，参考图 = 采集主图）
  - `doubao`：火山方舟 Seedream（默认，`ARK_API_KEY`）
  - `tongyi`：通义万相/千问（`DASHSCOPE_API_KEY`，wan* 异步轮询 / qwen* 同步）
+  - `rightapi`：gpt-image-2（OpenAI 兼容中转 `RIGHTAPI_API_KEY`，edits 参考图 / generations 同步）
+  - 插件只传模型名，服务端按模型名自动路由到对应 provider

 ## 快速开始

@@ -0,0 +1,35 @@
+# ===== 服务 =====
+# 注意：不要用 5000/7000（macOS 隔空播放接收器占用）
+HOST=127.0.0.1
+PORT=3300
+# 拼给插件/前端的媒体访问地址（默认本机）
+APP_BASE_URL=http://127.0.0.1:3300
+
+# ===== 存储 =====
+# 图片/数据库落盘目录（默认 <项目根>/data）
+# DATA_DIR=/Users/joey/sites/seller-store/image-suite-studio/data
+
+# ===== 图像生成 =====
+# 默认 provider：doubao（火山方舟 Seedream）| tongyi（阿里 DashScope）| rightapi（gpt-image）
+IMAGE_PROVIDER=doubao
+# 单张生图请求超时（秒）
+REQUEST_TIMEOUT=300
+# 异步任务轮询上限（秒）
+POLL_MAX_WAIT=600
+
+# --- 豆包 / 火山方舟 ---
+ARK_API_KEY=
+ARK_BASE_URL=https://ark.cn-beijing.volces.com/api/v3/images/generations
+ARK_IMAGE_MODEL=doubao-seedream-4-5-251128
+
+# --- 通义 / DashScope ---
+DASHSCOPE_API_KEY=
+DASHSCOPE_BASE_URL=
+DASHSCOPE_MODEL=wan2.7-image-pro
+
+# --- RightAPI（gpt-image，OpenAI 兼容中转）---
+RIGHTAPI_API_KEY=sk-b6b9ffba28b64ca795be31f6dee842c6
+RIGHTAPI_BASE_URL=https://rightapi.ai/draw
+RIGHTAPI_IMAGE_MODEL=gpt-image-2
+# 出图质量：auto | low | medium | high（high 单张约 1-5 分钟，超时自动兜底 600s）
+RIGHTAPI_IMAGE_QUALITY=high
@@ -20,5 +20,6 @@
    "@types/react-dom": "^18.3.0",
    "typescript": "^5.5.3",
    "wxt": "^0.19.0"
-  }
+  },
+  "packageManager": "pnpm@10.32.1+sha512.a706938f0e89ac1456b6563eab4edf1d1faf3368d1191fc5c59790e96dc918e4456ab2e67d613de1043d2e8c81f87303e6b40d4ffeca9df15ef1ad567348f2be"
 }
@@ -102,7 +102,7 @@ export const PLATFORM_OPTIONS = [

 export type PlatformId = (typeof PLATFORM_OPTIONS)[number]['value'];

-/** 生图模型（通义 DashScope）：下拉可选 + 中文特点说明 */
+/** 生图模型（下拉可选 + 中文特点说明；服务端按模型名路由 provider） */
 export const IMAGE_MODEL_OPTIONS = [
  {
    value: 'qwen-image-3.0-pro',
@@ -114,6 +114,11 @@ export const IMAGE_MODEL_OPTIONS = [
    label: 'wan2.7-image-pro',
    desc: '异步精修，质感与细节更强，适合高质量电商大片',
  },
+  {
+    value: 'gpt-image-2',
+    label: 'gpt-image-2',
+    desc: 'GPT 图像模型，构图与图内文案渲染最强，参考图高保真，单张 1-5 分钟',
+  },
 ] as const;

 export const PLATFORM_SPECS: Record<string, { lang: string; ratio: string; label: string }> = {
@@ -7,7 +7,8 @@ from config import get_settings
 from db import get_db
 from models import Suite
 from schemas import (
-    GenerateRequest, PLATFORM_SPECS, SUPPORTED_TYPES, TONGYI_MODELS, SuiteCreateResponse, TextMaterial,
+    GenerateRequest, PLATFORM_SPECS, SUPPORTED_TYPES, TONGYI_MODELS, RIGHTAPI_MODELS,
+    SuiteCreateResponse, TextMaterial, resolve_provider,
    PlanRequest, PlanResponse, PlanItemOut,
 )
 from services.generator import run_suite
@@ -84,10 +85,13 @@ async def generate_suite(
    spec = PLATFORM_SPECS[req.platform]

    settings = get_settings()
-    provider_name = req.provider or settings.image_provider
+    # 插件只传模型名：已知模型直接路由到对应 provider（gpt-image-2 → rightapi）
+    provider_name = resolve_provider(req.model, req.provider, settings.image_provider)
    model = req.model
    if provider_name == "tongyi" and model and model not in TONGYI_MODELS:
        raise HTTPException(status_code=400, detail=f"不支持的模型: {model}（tongyi 支持: {TONGYI_MODELS}）")
+    if provider_name == "rightapi" and model and model not in RIGHTAPI_MODELS:
+        raise HTTPException(status_code=400, detail=f"不支持的模型: {model}（rightapi 支持: {RIGHTAPI_MODELS}）")
    suite = Suite(
        product_id=None,
        style_set=req.style_set,
@@ -13,7 +13,10 @@ from sqlalchemy.ext.asyncio import AsyncSession
 from config import get_settings
 from db import get_db
 from models import Product, ProductAsset, Suite, SuiteImage, STATUS_OK
-from schemas import PLATFORM_SPECS, SUPPORTED_TYPES, TONGYI_MODELS, SuiteCreateRequest, SuiteCreateResponse, SuiteImageOut, SuiteOut
+from schemas import (
+    PLATFORM_SPECS, SUPPORTED_TYPES, TONGYI_MODELS, RIGHTAPI_MODELS,
+    SuiteCreateRequest, SuiteCreateResponse, SuiteImageOut, SuiteOut, resolve_provider,
+)
 from services import storage
 from services.generator import run_suite

@@ -76,9 +79,12 @@ async def create_suite(
    spec = PLATFORM_SPECS[req.platform]

    settings = get_settings()
-    provider_name = req.provider or settings.image_provider
+    # 模型名优先路由：已知模型直接定位 provider（gpt-image-2 → rightapi）
+    provider_name = resolve_provider(req.model, req.provider, settings.image_provider)
    if provider_name == "tongyi" and req.model and req.model not in TONGYI_MODELS:
        raise HTTPException(status_code=400, detail=f"不支持的模型: {req.model}（tongyi 支持: {TONGYI_MODELS}）")
+    if provider_name == "rightapi" and req.model and req.model not in RIGHTAPI_MODELS:
+        raise HTTPException(status_code=400, detail=f"不支持的模型: {req.model}（rightapi 支持: {RIGHTAPI_MODELS}）")
    suite = Suite(
        product_id=product.id,
        style_set=req.style_set,
@@ -41,6 +41,12 @@ class Settings(BaseSettings):
    dashscope_base_url: str = ""        # 留空按模型自动选择万象异步/千问同步端点
    dashscope_model: str = "wan2.7-image-pro"

+    # RightAPI（OpenAI 兼容中转，gpt-image 系列）
+    rightapi_api_key: str = ""
+    rightapi_base_url: str = "https://rightapi.ai/draw"
+    rightapi_image_model: str = "gpt-image-2"
+    rightapi_image_quality: str = "high"  # auto | low | medium | high
+
    # DeepSeek（出图方案规划器）
    deepseek_api_key: str = ""
    deepseek_base_url: str = "https://api.deepseek.com/v1"
@@ -50,6 +50,7 @@ async def health():
        "provider": settings.image_provider,
        "ark_configured": bool(settings.ark_api_key),
        "dashscope_configured": bool(settings.dashscope_api_key),
+        "rightapi_configured": bool(settings.rightapi_api_key),
    }


@@ -12,6 +12,22 @@ SUPPORTED_TYPES = [
 # 通义（DashScope）生图模型白名单：插件下拉可选的模型
 TONGYI_MODELS = ["qwen-image-3.0-pro", "wan2.7-image-pro"]

+# RightAPI 生图模型白名单
+RIGHTAPI_MODELS = ["gpt-image-2"]
+
+# 模型 → provider 推断表：插件只传模型名，服务端据此路由（模型名优先于 provider 字段）
+MODEL_PROVIDERS: dict[str, str] = {
+    **{m: "tongyi" for m in TONGYI_MODELS},
+    **{m: "rightapi" for m in RIGHTAPI_MODELS},
+}
+
+
+def resolve_provider(model: str | None, requested: str | None, default: str) -> str:
+    """已知模型名直接定位 provider；未知模型回退到请求指定的 provider 或默认值。"""
+    if model and model in MODEL_PROVIDERS:
+        return MODEL_PROVIDERS[model]
+    return requested or default
+

 # ── 采集上传 ──

@@ -46,6 +46,9 @@ def _image_size(provider: str, ratio: str, is_wan: bool = True) -> str:
    """平台比例 → provider 尺寸参数。3:4 竖版（Ozon/WB），1:1 方图（国内）。"""
    if provider == "doubao":
        return "1536x2048" if ratio == "3:4" else "2048x2048"
+    if provider == "rightapi":
+        # gpt-image 自定义尺寸约束：16 的倍数、长短边比 ≤ 3:1（1536x2048 合法）
+        return "1536x2048" if ratio == "3:4" else "2048x2048"
    # tongyi：万象与千问的 size 语法相同（* 分隔），档位不同
    if ratio == "3:4":
        return "1536*2048" if is_wan else "768*1024"
@@ -70,27 +73,34 @@ def _bytes_to_data_uri(data: bytes, mime: str) -> str:
    return f"data:{mime};base64,{base64.b64encode(data).decode()}"


-async def _resolve_ref(url: str) -> str:
-    """参考图 URL → data URI。本地 media 文件直读磁盘；远程 URL 带 Referer 下载。
+async def _resolve_ref_bytes(url: str) -> tuple[bytes, str]:
+    """参考图 URL → (bytes, mime)。本地 media 文件直读磁盘；远程 URL 带 Referer 下载。

    生图 API 的服务器无法访问 127.0.0.1，代理 URL 也不能直接透传，
-    所以统一在本地解析成 base64 data URI 再进请求体。
+    所以统一在本地解析成原始字节再进请求体（data URI 或 multipart）。
    """
    if url.startswith("data:"):
-        return url
+        head, _, b64 = url.partition(",")
+        mime = head[5:].split(";", 1)[0] or "image/jpeg"
+        return base64.b64decode(b64), mime
    path = storage.local_path(url)
    if path is not None:
        mime = mimetypes.guess_type(path.name)[0] or "image/jpeg"
-        return _bytes_to_data_uri(path.read_bytes(), mime)
+        return path.read_bytes(), mime
    if url.startswith(("http://", "https://")):
        from api.proxy import guess_referer
        data, ctype = await storage.download_bytes(url, referer=guess_referer(url))
        if not ctype.startswith("image/"):
            ctype = "image/jpeg"
-        return _bytes_to_data_uri(data, ctype)
+        return data, ctype
    raise FileNotFoundError(f"无法解析参考图: {url}")


+async def _resolve_ref(url: str) -> str:
+    data, mime = await _resolve_ref_bytes(url)
+    return _bytes_to_data_uri(data, mime)
+
+
 # ── Provider：豆包 Seedream（火山方舟）────────────────────────────────────

 async def generate_doubao(prompt: str, ref_images: list[str], size: str = "2048x2048", model: str | None = None) -> bytes:
@@ -198,7 +208,56 @@ async def generate_tongyi(prompt: str, ref_images: list[str], size: str = "2048*
        return dl.content


-GENERATORS = {"doubao": generate_doubao, "tongyi": generate_tongyi}
+# ── Provider：RightAPI（gpt-image，OpenAI 兼容中转）──────────────────────
+
+async def generate_rightapi(prompt: str, ref_images: list[str], size: str = "2048x2048", model: str | None = None) -> bytes:
+    """gpt-image 系列：有参考图走 /v1/images/edits（multipart），无参考图走 /v1/images/generations。
+
+    OpenAI 兼容协议：响应固定 b64_json（不支持 response_format 参数，传了报 400）；
+    同步调用无任务轮询，高质量档单张 1-5 分钟，超时按文档建议兜底 600s。
+    """
+    s = get_settings()
+    if not s.rightapi_api_key:
+        raise RuntimeError("未配置 RIGHTAPI_API_KEY（.env）")
+    model = model or s.rightapi_image_model
+    base = s.rightapi_base_url.rstrip("/")
+    headers = {"Authorization": f"Bearer {s.rightapi_api_key}"}
+
+    async with httpx.AsyncClient(timeout=max(s.request_timeout, 600), verify=False) as client:
+        common = {
+            "model": model,
+            "prompt": prompt,
+            "size": size,
+            "quality": s.rightapi_image_quality,
+            "output_format": "jpeg",  # 与落盘 .jpg 后缀一致
+            "n": 1,
+        }
+        if ref_images:
+            files = []
+            for i, u in enumerate(ref_images):
+                data, mime = await _resolve_ref_bytes(u)
+                files.append(("image[]", (f"ref-{i + 1}.{mime.split('/')[-1]}", data, mime)))
+            resp = await client.post(f"{base}/v1/images/edits", headers=headers, files=files, data=common)
+        else:
+            resp = await client.post(
+                f"{base}/v1/images/generations",
+                headers={**headers, "Content-Type": "application/json"},
+                json=common,
+            )
+        _raise_api_error(resp, "RightAPI")
+        item = resp.json()["data"][0]
+        b64 = item.get("b64_json") or ""
+        if b64:
+            return base64.b64decode(b64.split(",", 1)[-1] if "," in b64 else b64)
+        img_url = item.get("url") or ""
+        if not img_url:
+            raise RuntimeError(f"RightAPI 响应里没有图片数据: {item}")
+        dl = await client.get(img_url, timeout=s.request_timeout)
+        dl.raise_for_status()
+        return dl.content
+
+
+GENERATORS = {"doubao": generate_doubao, "tongyi": generate_tongyi, "rightapi": generate_rightapi}


 # ── 任务执行器 ────────────────────────────────────────────────────────────
@@ -280,9 +339,10 @@ async def run_suite(suite_id: str) -> None:

        raw = suite.context if not product else (product.raw or {})
        ctx = build_context(raw or {}, fallback_name=product.name if product else "")
-        model = suite.model or (
-            settings.dashscope_model if provider_name == "tongyi" else settings.ark_image_model
-        )
+        model = suite.model or {
+            "tongyi": settings.dashscope_model,
+            "rightapi": settings.rightapi_image_model,
+        }.get(provider_name, settings.ark_image_model)
        is_wan = provider_name == "tongyi" and _is_wan_model(model)
        size = _image_size(provider_name, suite.ratio, is_wan=is_wan)

@@ -317,7 +377,9 @@ async def run_suite(suite_id: str) -> None:
                else:
                    refs = _refs_for_job(list(suite.ref_images or []), job)
                data = await generator(prompt, refs, size=size, model=model)
-                key = storage.write_bytes(data, key_prefix=f"suites/{suite.id}", ext=".jpg")
+                # 部分中转不遵守 output_format（要 jpeg 回 PNG），按魔数定扩展名
+                ext = ".png" if data[:8] == b"\x89PNG\r\n\x1a\n" else ".jpg"
+                key = storage.write_bytes(data, key_prefix=f"suites/{suite.id}", ext=ext)
                image_row.stored_url = storage.public_url(key)
                image_row.status = STATUS_OK
                ok += 1