"""AI 炫图路由：gpt-image 生图代理（Key 仅存后端），提示词由 drawing_data 结构化构建。 - quick 模式: /v1/images/generations 纯提示词生成 - photo 模式: /v1/images/edits 以用户照片为参考图，把窗型合成进照片 """ import base64 import json import httpx from fastapi import APIRouter from pydantic import BaseModel, Field from huamenchuang.core.config import settings from huamenchuang.core.deps import CurrentUser, DbSession from huamenchuang.core.response import fail, ok from huamenchuang.services.ai_prompt import build_prompt, describe_window, pick_size router = APIRouter(prefix="/ai-render", tags=["AI 炫图"]) class AiRenderBody(BaseModel): template_id: int | None = None drawing_data: str | dict | None = None mode: str = Field(default="quick", pattern="^(quick|photo|blueprint)$") style: str = Field(default="modern", pattern="^(modern|chinese|european|nordic)$") time: str = Field(default="day", pattern="^(day|dusk|night)$") # quick: indoor/outdoor/studio; photo: render/install/erase/indoor/villa/fence/blank/sunroom-* scene: str = "indoor" photo_b64: str | None = Field(default=None, description="照片模式参考图(base64, 不带 dataURL 前缀)") rect: dict | None = Field(default=None, description="照片模式框选窗洞 {x,y,w,h}(原始照片像素坐标)") sketch_b64: str | None = Field(default=None, description="大样图模式: 2D 图纸截图(base64, 不带 dataURL 前缀)") async def _resolve_drawing(db, body: AiRenderBody) -> str | dict | None: if body.drawing_data: return body.drawing_data if body.template_id: from sqlalchemy import select from huamenchuang.models.windoor_template import WindoorTemplate t = (await db.execute(select(WindoorTemplate).where( WindoorTemplate.id == body.template_id))).scalar_one_or_none() return t.drawing_data if t else None return None @router.post("/prompt") async def preview_prompt(body: AiRenderBody, user: CurrentUser, db: DbSession): """提示词预览（前端可展开查看优化后的完整提示词）。""" drawing = await _resolve_drawing(db, body) if drawing is None: return fail("未找到图纸数据") try: info = describe_window(drawing) prompt = (build_blueprint_prompt(drawing) if body.mode == "blueprint" else build_prompt(drawing, body.style, body.time, body.scene)) except Exception: return fail("图纸数据解析失败") return ok(data={"features": info, "prompt": prompt, "size": pick_size(info["width_mm"], info["height_mm"])}) def build_blueprint_prompt(drawing: str | dict) -> str: """工程大样图提示词：手绘草图/量尺登记表照片或 2D 图纸 → 黑白 CAD 大样图 （移植自 autocad 项目草图标准化资产，叠加 drawing_data 结构化特征增强分格保真）。""" feat = "" if drawing: try: info = describe_window(drawing) feat = f"参考窗型（若输入草图中有相似窗型可参照其分格风格）: {info['summary']}; 外框尺寸 {info['width_mm']}×{info['height_mm']} mm。" except Exception: feat = "" return f"""Use case: style-transfer Asset type: 门窗工程审图参考图 Primary request: 将输入的门窗手绘草图（现场量尺登记表照片）或 2D 图纸重绘为清晰、整洁的工程 CAD 大样图风格预览。 Input images: 第一张是原始手绘草图照片或 2D 图纸，仅作为几何、分格与标注位置的几何来源。输入可能是一页包含多个窗的手绘量尺登记表——逐窗转换，保持各窗的相对布局与各自分格。 Style/medium: 黑白二维门窗工程大样图、正投影立面、细黑线框、玻璃淡灰填充、外部尺寸线与分格尺寸标注、中文工程标注、每个窗下方标注"窗1/窗2…"编号、右下角标题栏（图名/比例/日期）。 Constraints: {feat} 方向铁律(最重要): 每个窗的横竖方向必须与原图完全一致——原图中横向更宽的窗，输出必须仍是横向(宽>高)；原图中竖向更高的窗，输出必须仍是竖向(高>宽)；严禁旋转、竖立或横放任何窗。宽度尺寸一律标注在窗的正下方水平尺寸线上，高度尺寸一律标注在窗的右侧竖直尺寸线上。登记表手写表格中"宽/mm"栏的数字对应草图的水平方向，"高/mm"栏对应竖直方向。 数量铁律: 输出大样图的数量必须与输入草图中的窗数量完全相同，一个不多一个不少；每个窗只绘制一次，严禁重复绘制同一个窗，严禁合并或省略任何窗。 尺寸铁律: 手写数字可能有涂改或划掉——被划掉的数字以改后数字为准；只誊写可清晰辨认的数字；无法辨认的尺寸处只标注"?"，严禁猜测或编造任何数字。 开向铁律: 每个开启扇的对角三角符号必须与原图的铰链侧、开启方向完全一致，不得镜像或反向。 斜角铁律: 手稿中的斜边、切角、梯形轮廓必须原样保留——外边界或玻璃分格中出现斜线时，输出必须以相同方向与相近比例的斜线呈现（斜线两端点的位置与原图一致），严禁把斜角窗、梯形窗拉直或画成矩形；手稿标注了斜边尺寸的照常誊写，未标注的按原图比例绘制；圆弧边同样原样保留。 结构铁律: 严格保留每个窗的分格数量、比例与位置；不要增删或合并分格；如果照片是旋转拍摄的，先将画面转正再逐窗转换；输出仅供人工审图与报价附件，不作为 CAD 几何或生产数据来源；无水印、无装饰性场景、无彩色、不保留登记表表格框线。 Avoid: 三维渲染、透视效果、写实照片、彩色渲染、虚构品牌、模糊文字、重复的窗、被旋转方向的窗。""" @router.post("/generate") async def generate(body: AiRenderBody, user: CurrentUser, db: DbSession): """云端 AI 生图（gpt-image）：quick=生成 / photo=参考图编辑 / blueprint=手绘草图或2D图纸转工程大样图。""" if not settings.ai_api_key: return fail("AI 生图服务未配置(HMC_AI_API_KEY)") drawing = await _resolve_drawing(db, body) # blueprint 纯草图模式(上传手绘登记表, 无样板)不要求图纸数据 if drawing is None and body.mode != "blueprint": return fail("未找到图纸数据") if body.mode == "photo" and not body.photo_b64: return fail("照片模式需要上传参考照片") if body.mode == "blueprint" and not body.sketch_b64: return fail("大样图模式需要草图照片或 2D 图纸截图") try: info = describe_window(drawing) if drawing else {"width_mm": 0, "height_mm": 0, "summary": ""} # 照片模式: 把框选洞口位置转成提示词方位描述, 引导 AI 装窗位置 rect_hint = "" if body.mode == "photo" and body.rect: try: r = body.rect # 由前端提供照片宽高与洞口坐标(像素) → 相对方位词 ph = float(r.get("photo_h") or 0) pw = float(r.get("photo_w") or 0) if pw > 0 and ph > 0: cx = (float(r["x"]) + float(r["w"]) / 2) / pw cy = (float(r["y"]) + float(r["h"]) / 2) / ph horiz = "left" if cx < 0.37 else ("right" if cx > 0.63 else "center") vert = "upper" if cy < 0.37 else ("lower" if cy > 0.63 else "middle") area_pct = round((float(r["w"]) * float(r["h"])) / (pw * ph) * 100) rect_hint = (f"Install the window into the existing window opening located in the " f"{vert}-{horiz} part of the image (opening covers about {area_pct}% of the frame), " f"replacing whatever old window or opening is there.") except Exception: rect_hint = "" if body.mode == "blueprint": prompt = build_blueprint_prompt(drawing) size = "1536x1024" # 大样图横版(量尺登记表整页比例), 对齐 autocad 项目口径 else: prompt = build_prompt(drawing, body.style, body.time, body.scene, rect_hint=rect_hint) size = pick_size(info["width_mm"], info["height_mm"]) except Exception: return fail("图纸数据解析失败") base = settings.ai_base_url.rstrip("/") headers = {"Authorization": f"Bearer {settings.ai_api_key}"} timeout = httpx.Timeout(settings.ai_timeout, connect=15) async def _post_image(client: httpx.AsyncClient) -> httpx.Response: if body.mode == "blueprint": img_bytes = base64.b64decode(body.sketch_b64.split(",")[-1]) mime = "image/jpeg" if img_bytes[:2] == b"\xff\xd8" else "image/png" ext = "jpg" if mime == "image/jpeg" else "png" return await client.post( f"{base}/v1/images/edits", headers=headers, data={"model": settings.ai_model, "prompt": prompt, "size": size, "n": "1"}, files={"image[]": (f"drawing.{ext}", img_bytes, mime)}) if body.mode == "photo": img_bytes = base64.b64decode(body.photo_b64.split(",")[-1]) mime = "image/jpeg" if img_bytes[:2] == b"\xff\xd8" else "image/png" ext = "jpg" if mime == "image/jpeg" else "png" return await client.post( f"{base}/v1/images/edits", headers=headers, data={"model": settings.ai_model, "prompt": prompt, "size": size, "n": "1"}, files={"image[]": (f"photo.{ext}", img_bytes, mime)}) return await client.post( f"{base}/v1/images/generations", headers=headers, json={"model": settings.ai_model, "prompt": prompt, "n": 1, "size": size}) b64 = None last_err = "" # 中转偶发超时/5xx 自动重试一次(生图耗时长, 失败重试显著提升成功率) for attempt in range(2): try: async with httpx.AsyncClient(timeout=timeout) as client: resp = await _post_image(client) resp.raise_for_status() data = resp.json() item = (data.get("data") or [{}])[0] b64 = item.get("b64_json") if b64: break last_err = f"生图服务返回异常: {str(data)[:200]}" except httpx.TimeoutException: last_err = "AI 生图超时，请稍后重试" except httpx.HTTPStatusError as e: detail = e.response.text[:200] if e.response is not None else "" last_err = f"生图服务错误({e.response.status_code if e.response else '?'}): {detail}" # 4xx(如鉴权/参数)重试无意义 if e.response is not None and e.response.status_code < 500: break except Exception as e: # noqa: BLE001 last_err = f"生图失败: {str(e)[:150]}" if attempt == 0 and b64 is None: stage_wait = 3 # 重试前短暂等待 import asyncio await asyncio.sleep(stage_wait) if not b64: return fail(last_err or "生成失败") return ok(data={"image_b64": b64, "prompt": prompt, "size": size, "features": {"width_mm": info["width_mm"], "height_mm": info["height_mm"], "summary": info["summary"]}}, message="生成成功")