feat: 开发编辑工作台
This commit is contained in:
R524809
+9
-1
@@ -1,5 +1,13 @@
|
||||
from fastapi import APIRouter
|
||||
|
||||
from schemas.image_edit import ImageEditRequest, ImageEditResponse
|
||||
from services.image_edit import edit_image
|
||||
|
||||
router = APIRouter(prefix="/api/image", tags=["image"])
|
||||
|
||||
# Phase 2: watermark / white background / img2img proxy
|
||||
|
||||
@router.post("/edit", response_model=ImageEditResponse, summary="AI 图生图（wanx2.1-imageedit）")
|
||||
async def create_image_edit(body: ImageEditRequest) -> ImageEditResponse:
|
||||
return await edit_image(body)
|
||||
|
||||
# 后续扩展点：白底 / 水印等服务端处理可继续在此文件下新增路由。
|
||||
|
||||
@@ -20,8 +20,12 @@ class Settings(BaseSettings):
|
||||
|
||||
deepseek_api_key: str = ""
|
||||
openai_api_key: str = ""
|
||||
dashscope_api_key: str = ""
|
||||
# 华北2（北京）业务空间时需填：https://{WorkspaceId}.cn-beijing.maas.aliyuncs.com/api/v1
|
||||
# 普通 API Key 调用留空即可。
|
||||
dashscope_base_http_api_url: str = ""
|
||||
host: str = "127.0.0.1"
|
||||
port: int = 8000
|
||||
port: int = 8800
|
||||
cors_origins: str = ""
|
||||
|
||||
@property
|
||||
|
||||
@@ -4,3 +4,4 @@ httpx>=0.27.0
|
||||
pydantic-settings>=2.6.0
|
||||
python-dotenv>=1.0.0
|
||||
PyYAML>=6.0.0
|
||||
dashscope>=1.23.8
|
||||
|
||||
@@ -0,0 +1,125 @@
|
||||
"""图生图 / 图像编辑请求响应模型（多模型）。"""
|
||||
from __future__ import annotations
|
||||
|
||||
from typing import Literal
|
||||
|
||||
from pydantic import BaseModel, Field, field_validator
|
||||
|
||||
# 支持的模型白名单
|
||||
SUPPORTED_MODELS: frozenset[str] = frozenset(
|
||||
{
|
||||
"wanx2.1-imageedit",
|
||||
"wan2.6-image",
|
||||
"qwen-image-edit",
|
||||
"qwen-image-edit-plus",
|
||||
"qwen-image-edit-plus-2025-10-30",
|
||||
}
|
||||
)
|
||||
|
||||
# 万相 2.1 通用图像编辑支持的 function 白名单
|
||||
# description_edit ：文本指令编辑（无掩码，即“图生图”）
|
||||
# description_edit_with_mask ：局部重绘（需 mask）
|
||||
# stylization_local ：局部风格化
|
||||
# stylization_all ：整体风格化
|
||||
ImageEditFunction = Literal[
|
||||
"description_edit",
|
||||
"description_edit_with_mask",
|
||||
"stylization_local",
|
||||
"stylization_all",
|
||||
]
|
||||
|
||||
# 单张 base64 图片（编码后）上限，超出直接拒绝，避免无谓占用内存。
|
||||
_MAX_BASE64_LENGTH = 15 * 1024 * 1024
|
||||
|
||||
|
||||
class ImageEditRequest(BaseModel):
|
||||
"""图生图请求。图片以 data URL（data:image/...;base64,...）或公网 URL 传入。"""
|
||||
|
||||
base_image: str = Field(..., description="待编辑原图：data URL 或公网 http(s) URL")
|
||||
prompt: str = Field(..., description="编辑指令，例如“把背景换成纯白色摄影棚”")
|
||||
model: str = Field(
|
||||
default="wanx2.1-imageedit",
|
||||
description="模型 code，见 SUPPORTED_MODELS",
|
||||
)
|
||||
mask_image: str | None = Field(
|
||||
default=None, description="掩码图（仅 function=description_edit_with_mask 时使用）"
|
||||
)
|
||||
function: ImageEditFunction = Field(
|
||||
default="description_edit", description="编辑能力类型，见 ImageEditFunction"
|
||||
)
|
||||
n: int = Field(default=1, ge=1, le=4, description="生成数量")
|
||||
size: str = Field(default="", description="输出尺寸，如 1024*1024；留空由模型决定")
|
||||
seed: int | None = Field(default=None, ge=0, description="随机种子，可选")
|
||||
style: str = Field(default="", description="输出风格，如 auto/photography/anime 等，可选")
|
||||
prompt_extend: bool = Field(default=True, description="是否自动扩写提示词（官方示例默认开启）")
|
||||
strength: float | None = Field(
|
||||
default=None,
|
||||
ge=0.0,
|
||||
le=1.0,
|
||||
description=(
|
||||
"修改幅度 0.0~1.0，官方默认 0.5；值越大改动越明显、越偏离原图。"
|
||||
"新增文字、大幅修改等场景建议 0.8 左右"
|
||||
),
|
||||
)
|
||||
|
||||
@field_validator("prompt")
|
||||
@classmethod
|
||||
def prompt_not_empty(cls, value: str) -> str:
|
||||
text = (value or "").strip()
|
||||
if not text:
|
||||
raise ValueError("prompt 不能为空")
|
||||
return text
|
||||
|
||||
@field_validator("model")
|
||||
@classmethod
|
||||
def model_whitelist(cls, value: str) -> str:
|
||||
model = (value or "").strip()
|
||||
if model not in SUPPORTED_MODELS:
|
||||
raise ValueError(f"不支持的模型「{model}」，可选：{sorted(SUPPORTED_MODELS)}")
|
||||
return model
|
||||
|
||||
@field_validator("base_image")
|
||||
@classmethod
|
||||
def base_image_valid(cls, value: str) -> str:
|
||||
image = (value or "").strip()
|
||||
if not image:
|
||||
raise ValueError("base_image 不能为空")
|
||||
if image.startswith("data:image/"):
|
||||
if len(image) > _MAX_BASE64_LENGTH:
|
||||
raise ValueError("图片过大，请压缩后重试")
|
||||
elif not image.startswith(("http://", "https://")):
|
||||
raise ValueError("base_image 仅支持 data URL 或公网 http(s) URL")
|
||||
return image
|
||||
|
||||
@field_validator("mask_image")
|
||||
@classmethod
|
||||
def mask_image_valid(cls, value: str | None) -> str | None:
|
||||
if value is None or not value.strip():
|
||||
return None
|
||||
image = value.strip()
|
||||
if image.startswith("data:image/"):
|
||||
if len(image) > _MAX_BASE64_LENGTH:
|
||||
raise ValueError("mask_image 图片过大")
|
||||
elif not image.startswith(("http://", "https://")):
|
||||
raise ValueError("mask_image 仅支持 data URL 或公网 http(s) URL")
|
||||
return image
|
||||
|
||||
@field_validator("style")
|
||||
@classmethod
|
||||
def normalize_style(cls, value: str) -> str:
|
||||
return (value or "").strip()
|
||||
|
||||
|
||||
class ImageEditResult(BaseModel):
|
||||
url: str = Field(..., description="生成图片 URL（有效期 24 小时）")
|
||||
image_base64: str = Field(
|
||||
default="",
|
||||
description="生成图片的 data URL（服务端代理下载，便于前端直接合成/导出，规避跨域画布污染）",
|
||||
)
|
||||
|
||||
|
||||
class ImageEditResponse(BaseModel):
|
||||
task_id: str = ""
|
||||
results: list[ImageEditResult] = Field(default_factory=list)
|
||||
image_count: int = 0
|
||||
request_id: str = ""
|
||||
@@ -0,0 +1,185 @@
|
||||
"""多模型图生图 / 图像编辑服务。
|
||||
|
||||
按模型路由到不同的 DashScope 调用形态：
|
||||
- wanx2.1-imageedit → ImageSynthesis（function=description_edit）
|
||||
- wan2.6-image / qwen-image-edit(-plus) → MultiModalConversation（多模态生成）
|
||||
|
||||
职责单一：把请求翻译成 DashScope 调用并返回结果。当前不落盘、不存储图片；
|
||||
后续接入七牛云时，可在返回结果 URL 之后增加「下载并转存」步骤，而不改 api 层与前端契约。
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import asyncio
|
||||
import base64
|
||||
from http import HTTPStatus
|
||||
|
||||
import httpx
|
||||
from fastapi import HTTPException
|
||||
|
||||
from config import get_settings
|
||||
from schemas.image_edit import ImageEditRequest, ImageEditResponse, ImageEditResult
|
||||
|
||||
# 服务端代理下载结果图时的体积上限，超出则跳过 base64（前端回退用 url）
|
||||
_MAX_PROXY_BYTES = 20 * 1024 * 1024
|
||||
|
||||
# 走 ImageSynthesis 的模型（其余统一走 MultiModalConversation）
|
||||
_WANX_IMAGEEDIT = "wanx2.1-imageedit"
|
||||
|
||||
|
||||
def _download_to_data_url(url: str) -> str:
|
||||
"""把生成结果图下载并转成 data URL。
|
||||
|
||||
目的：阿里云结果 URL 未开放 CORS，前端直接绘制到 canvas 会导致画布被污染、
|
||||
无法 toDataURL 导出；由服务端代下载可彻底规避。失败时返回空串，前端回退用 url。
|
||||
"""
|
||||
try:
|
||||
with httpx.Client(timeout=60.0, follow_redirects=True) as client:
|
||||
resp = client.get(url)
|
||||
resp.raise_for_status()
|
||||
content_type = (resp.headers.get("content-type") or "image/png").split(";")[0].strip()
|
||||
if not content_type.startswith("image/"):
|
||||
content_type = "image/png"
|
||||
payload = resp.content
|
||||
if len(payload) > _MAX_PROXY_BYTES:
|
||||
return ""
|
||||
b64 = base64.b64encode(payload).decode("ascii")
|
||||
return f"data:{content_type};base64,{b64}"
|
||||
except Exception: # noqa: BLE001 - 下载失败不影响主流程
|
||||
return ""
|
||||
|
||||
|
||||
def _apply_base_url(settings) -> None:
|
||||
"""仅华北2（北京）业务空间需要配置 base_http_api_url。"""
|
||||
import dashscope
|
||||
|
||||
url = (settings.dashscope_base_http_api_url or "").strip()
|
||||
if url:
|
||||
dashscope.base_http_api_url = url
|
||||
|
||||
|
||||
def _fail(rsp, label: str) -> None:
|
||||
code = getattr(rsp, "code", None)
|
||||
message = getattr(rsp, "message", "")
|
||||
status = getattr(rsp, "status_code", "?")
|
||||
raise HTTPException(status_code=502, detail=f"{label}失败（HTTP {status}，code={code}）：{message}")
|
||||
|
||||
|
||||
def _to_results(urls: list[str]) -> list[ImageEditResult]:
|
||||
return [ImageEditResult(url=u, image_base64=_download_to_data_url(u)) for u in urls]
|
||||
|
||||
|
||||
def _call_wanx(req: ImageEditRequest, api_key: str) -> ImageEditResponse:
|
||||
"""wanx2.1-imageedit：ImageSynthesis 同步调用。"""
|
||||
from dashscope import ImageSynthesis
|
||||
|
||||
kwargs: dict = {
|
||||
"api_key": api_key,
|
||||
"model": "wanx2.1-imageedit",
|
||||
"function": req.function,
|
||||
"prompt": req.prompt,
|
||||
"base_image_url": req.base_image,
|
||||
"n": req.n,
|
||||
}
|
||||
if req.mask_image:
|
||||
kwargs["mask_image_url"] = req.mask_image
|
||||
if req.size:
|
||||
kwargs["size"] = req.size
|
||||
if req.seed is not None:
|
||||
kwargs["seed"] = req.seed
|
||||
if req.style:
|
||||
kwargs["style"] = req.style
|
||||
if req.prompt_extend:
|
||||
kwargs["prompt_extend"] = True
|
||||
if req.strength is not None:
|
||||
# strength 走 SDK 的 **kwargs，最终落入请求体的 parameters
|
||||
kwargs["strength"] = req.strength
|
||||
|
||||
try:
|
||||
rsp = ImageSynthesis.call(**kwargs)
|
||||
except Exception as exc: # noqa: BLE001 - SDK 抛错类型不统一，统一转为 502
|
||||
raise HTTPException(status_code=502, detail=f"DashScope 调用异常：{exc}") from exc
|
||||
|
||||
if rsp.status_code != HTTPStatus.OK:
|
||||
_fail(rsp, "图生图")
|
||||
|
||||
urls: list[str] = []
|
||||
for item in getattr(rsp.output, "results", None) or []:
|
||||
url = item.get("url") if isinstance(item, dict) else getattr(item, "url", None)
|
||||
if url:
|
||||
urls.append(url)
|
||||
|
||||
if not urls:
|
||||
raise HTTPException(status_code=502, detail="模型未返回结果图片 URL")
|
||||
|
||||
usage = getattr(rsp, "usage", None)
|
||||
image_count = int(getattr(usage, "image_count", 0) or 0) if usage else len(urls)
|
||||
return ImageEditResponse(
|
||||
task_id=getattr(rsp.output, "task_id", ""),
|
||||
results=_to_results(urls),
|
||||
image_count=image_count or len(urls),
|
||||
request_id=getattr(rsp, "request_id", ""),
|
||||
)
|
||||
|
||||
|
||||
def _call_multimodal(req: ImageEditRequest, api_key: str) -> ImageEditResponse:
|
||||
"""wan2.6-image / qwen-image-edit(-plus[-快照])：MultiModalConversation。"""
|
||||
from dashscope import MultiModalConversation
|
||||
|
||||
messages = [{"role": "user", "content": [{"image": req.base_image}, {"text": req.prompt}]}]
|
||||
kwargs: dict = {
|
||||
"api_key": api_key,
|
||||
"model": req.model,
|
||||
"messages": messages,
|
||||
"prompt_extend": req.prompt_extend,
|
||||
}
|
||||
if req.n:
|
||||
kwargs["n"] = req.n
|
||||
if req.size:
|
||||
kwargs["size"] = req.size
|
||||
|
||||
try:
|
||||
rsp = MultiModalConversation.call(**kwargs)
|
||||
except Exception as exc: # noqa: BLE001
|
||||
raise HTTPException(status_code=502, detail=f"DashScope 调用异常：{exc}") from exc
|
||||
|
||||
if rsp.status_code != HTTPStatus.OK:
|
||||
_fail(rsp, "图生图")
|
||||
|
||||
urls: list[str] = []
|
||||
try:
|
||||
content = rsp.output.choices[0].message.content
|
||||
except (AttributeError, IndexError, KeyError, TypeError):
|
||||
content = []
|
||||
if isinstance(content, list):
|
||||
for item in content:
|
||||
img = item.get("image") if isinstance(item, dict) else None
|
||||
if img:
|
||||
urls.append(img)
|
||||
|
||||
if not urls:
|
||||
raise HTTPException(status_code=502, detail="模型未返回结果图片")
|
||||
|
||||
return ImageEditResponse(
|
||||
task_id="",
|
||||
results=_to_results(urls),
|
||||
image_count=len(urls),
|
||||
request_id=getattr(rsp, "request_id", ""),
|
||||
)
|
||||
|
||||
|
||||
def _call_dashscope(req: ImageEditRequest, api_key: str) -> ImageEditResponse:
|
||||
"""同步调用 DashScope（在线程池中执行），按模型路由。"""
|
||||
if req.model == _WANX_IMAGEEDIT:
|
||||
return _call_wanx(req, api_key)
|
||||
return _call_multimodal(req, api_key)
|
||||
|
||||
|
||||
async def edit_image(req: ImageEditRequest) -> ImageEditResponse:
|
||||
"""入口：校验密钥后在独立线程中执行同步 SDK 调用。"""
|
||||
settings = get_settings()
|
||||
api_key = (settings.dashscope_api_key or "").strip()
|
||||
if not api_key:
|
||||
raise HTTPException(status_code=500, detail="未配置 DASHSCOPE_API_KEY，请写入 .env")
|
||||
|
||||
_apply_base_url(settings)
|
||||
return await asyncio.to_thread(_call_dashscope, req, api_key)
|
||||
Reference in New Issue
Block a user