feat: 修改模型提示词
This commit is contained in:
R524809
@@ -45,8 +45,12 @@ Chrome 插件（WXT + React + antd） Python 后端（FastAPI，无数
|
||||
|
||||
### 套图生成（server/services）
|
||||
|
||||
- `prompt.py`：7 种图类型 × 5 套风格模板，公共组件 QUALITY / PRODUCT_REF_LOCK（商品一致性锁）/ TEXT_RENDER
|
||||
- 图类型：白底主图 / 核心卖点图 / 卖点图 / 材质图 / 场景展示图 / 多场景拼图 / 电商详情图
|
||||
- `services/prompts/`：提示词按模型家族独立封装，`__init__.py` 按 (provider, model) 路由分发
|
||||
- `common.py`：商品上下文提炼、5 套风格模板、图内文案规范 TEXT_RENDER（家族共用）
|
||||
- `alibaba.py`：通义 wan*/qwen*（主体参考语义）；`doubao.py`：豆包（同语义，复用阿里装配）
|
||||
- `gpt.py`：gpt-image-2/-vip（`/v1/images/edits` 编辑语义，保真优先：商品只由 Image 1 定义，文字锚定仅作识别）
|
||||
- `google.py`：nano-banana 系列（原生主体保持语义）
|
||||
- 图类型：白底主图 / 核心卖点图 / 卖点图 / 材质图 / 场景展示图 / 多场景拼图 / 电商详情图 / 尺寸标注图 / SKU合集 / 创意图
|
||||
- 风格：北欧极简 / 清新明亮 / 高级感深色 / 暖调生活 / 纯净棚拍
|
||||
- 卖点从采集的参数表/卖点文本自动提炼
|
||||
- `generator.py`：图像 provider（图生图，参考图 = 采集主图）
|
||||
|
||||
@@ -0,0 +1,146 @@
|
||||
# RightAPI（gpt-image / nano-banana）调用方式排查报告
|
||||
|
||||
> 2026-08-20 · 状态：**待确认**（确认后再改代码）
|
||||
> 结论先行：**是调用方式不对**。现行代码把参考图用 multipart 传给未在文档中的
|
||||
> `/v1/images/edits` 端点；该中转已于 2026-07-14 全面切换"统一异步模式"，文档中的
|
||||
> 正确用法是 `/v1/images/generations` + JSON `image`（data-URI 数组）+ `async: true`
|
||||
> 提交任务，再轮询 `/v1/tasks/{task_id}` 取图。实测：**文档路径下 gpt-image-2、
|
||||
> gpt-image-2-vip、nano-banana-2-lite 全部逐像素保真**；现行 edits 路径要么 502、
|
||||
> 要么出图但参考图未生效（商品按文字重造）。
|
||||
|
||||
---
|
||||
|
||||
## 1. 现行代码怎么调的（generator.py `_rightapi_request`）
|
||||
|
||||
```python
|
||||
# 有参考图（套图流程必然有）→ multipart POST /v1/images/edits
|
||||
files = [("image[]", ("ref-1.png", data, mime)), ...]
|
||||
data = {"model", "prompt", "size": "2048x2048", "quality": "high",
|
||||
"output_format": "jpeg", "n": 1, "input_fidelity": "high"}
|
||||
resp = client.post(f"{base}/v1/images/edits", files=files, data=data)
|
||||
# 期望同步响应 data[0].b64_json / data[0].url，无任务轮询
|
||||
```
|
||||
|
||||
问题：
|
||||
- **`/v1/images/edits` 不在文档接口列表里**（文档只有：图片生成、Gemini 生成、任务查询）；
|
||||
- 参考图用 multipart `image[]` 传输——新管道只认 JSON body 里的 `image`（data-URI 数组）；
|
||||
- 未带 `"async": true`，也没有任务轮询逻辑；
|
||||
- `quality` / `output_format` / `input_fidelity` 均不在文档参数表中。
|
||||
|
||||
## 2. 文档的正确用法（docs.rightapi.ai，2026-07-14 更新）
|
||||
|
||||
### 2.1 提交：POST `/v1/images/generations`（OpenAI Images 兼容，异步）
|
||||
|
||||
```json
|
||||
{
|
||||
"model": "gpt-image-2", // 或 nano-banana 系列等
|
||||
"prompt": "...",
|
||||
"n": 1,
|
||||
"size": "1:1", // 比例 1:1 / 16:9 / 9:16 / 4:3，或像素串 "1024x1024"
|
||||
"async": true, // 固定带
|
||||
"image": ["data:image/png;base64,..."] // 参考图：data-URI 数组（保真关键）
|
||||
}
|
||||
```
|
||||
|
||||
响应（立即返回）：
|
||||
|
||||
```json
|
||||
{"task_id": "task_xxx", "status": "processing", "progress": 0, "message": "..."}
|
||||
```
|
||||
|
||||
### 2.2 轮询：GET `/v1/tasks/{task_id}`（站点级，**不带 /draw 前缀**）
|
||||
|
||||
- 进行中：`{"id","task_id","object","model","status":"in_progress","progress":0~2,"created_at"}`
|
||||
- **完成：`{"created": ..., "data": [{"url": "https://...jpeg"}]}`**
|
||||
——实测完成响应**没有 `status: "completed"` 字段**（与文档描述不符），
|
||||
完成判定 = 响应里出现 `data`；结果只有 `url`（未见 b64_json）。
|
||||
- `progress` 基本不动（一直 0~2），只能当装饰，不能当进度条依据。
|
||||
|
||||
### 2.3 其他要点
|
||||
|
||||
- Gemini 原生端点 `/v1beta/models/{model}:generateContent`（contents/parts + inline_data，
|
||||
generationConfig.imageConfig 支持 aspectRatio / imageSize）——nano-banana 系列可走，
|
||||
但非必需（generations 端点同样支持传参考图），本期可不做；
|
||||
- `imageSize`："1K"/"2K"/"4K"，**仅 nano-banana / gpt-image vip 模型可用**；
|
||||
- 文档域名示例为 `www.right.codes/draw`，实测现有配置 `rightapi.ai/draw` 仍通
|
||||
（提交与任务查询都可用，`rightapi.ai/v1/tasks/...` 实测正常）。
|
||||
|
||||
## 3. 实测证据（2026-08-20，受控对照实验）
|
||||
|
||||
测试图：程序生成的特征图形——白底 + 青色杯身 + 红色横条纹 + 三颗黄色五角星 + 右侧把手。
|
||||
提示词："把背景替换成纯绿色，保持图中那个青色杯子完全不变……"。
|
||||
保真判定 = 逐项核对杯身/条纹/星星/把手是否原样（我人工查看生成图）。
|
||||
|
||||
| # | 路径 | 模型 | 结果 |
|
||||
|---|------|------|------|
|
||||
| A | **文档路径** generations + image[] + async | nano-banana-2-lite | ✅ **保真完美**，仅背景变绿 |
|
||||
| C | **文档路径** generations + image[] + async | gpt-image-2 | ✅ **保真完美**，仅背景变绿 |
|
||||
| D | **文档路径** generations + image[] + async | gpt-image-2-vip（官逆） | ✅ **保真完美**，仅背景变绿 |
|
||||
| B | **现行代码** edits + multipart image[] | nano-banana-2-lite | ❌ **502 Bad Gateway**（间隔 90s 重试仍 502；同期 generations 路径正常） |
|
||||
|
||||
用户今日实测（11:08–11:16，本地任务表，同一鲨鱼玩偶参考图）：
|
||||
|
||||
| 套图 | 模型（路径） | 结果 |
|
||||
|------|--------------|------|
|
||||
| 7bd57ffa | gpt-image-2-vip（现行 edits） | ⚠️ 出图，但鲨鱼被**重新设计**（眼睛/鱼鳍/比例全变） |
|
||||
| ee36e92f | nano-banana-2（现行 edits） | ⚠️ 同上，商品被重造 |
|
||||
| 42c37fa2 | wan2.6-image（DashScope，正常链路） | ⚠️ 鲨鱼同样有漂移（**另一层问题**，见 §5） |
|
||||
|
||||
探针产物（供复核）：`/tmp/rightapi-probe/`（ref.png / gen-async-lite.png / gen-async-0.png）。
|
||||
|
||||
## 4. 根因分析
|
||||
|
||||
1. **参考图从未真正送达模型**：edits + multipart 是旧同步模式的调用方式；中转 7-14
|
||||
切到统一异步管道后，multipart 参考图不被解析 → 模型只收到 prompt 文字 → 按文字
|
||||
（含标题/风格词）重新合成商品 → **"不是原商品"必现**。gpt 与 google 全中，因为
|
||||
它们共用这一条错误链路。
|
||||
2. **端点本身进入半废弃状态**：今天 edits 已对 lite 模型直接 502（两次、间隔 90s），
|
||||
对 gpt-image-2-vip / nano-banana-2 尚能返回（用户 11 点实测出图）——属于残留兼容，
|
||||
随时可能全断。之前代码里"同 key 分钟级冷却 502"的注释，与该端点的不稳定状态吻合。
|
||||
3. 提示词层面的修复（上一轮 gpt/google 家族重写）方向正确但**没治病根**：参考图没到
|
||||
模型，提示词写得再保真也没用。证据：同一套提示词组件，走文档路径（探测 A/C/D）
|
||||
保真完美。
|
||||
|
||||
## 5. 顺带观察：通义今日也有漂移（不在本次修复范围）
|
||||
|
||||
wan2.6-image 走 DashScope 正常链路（参考图确实送达）仍重造了鲨鱼——这是
|
||||
主体参考模型能力/提示词层面的问题（wan2.6-image 是参考遵循较弱的一档），
|
||||
与本次 RightAPI 调用方式无关，建议后续单独评估（比如套餐默认模型换成
|
||||
wan2.7-image-pro 或 qwen-image-3.0-pro，两者参考遵循更强）。
|
||||
|
||||
## 6. 修复方案（确认后实施）
|
||||
|
||||
只改 `server/services/generator.py` 的 RightAPI provider，提示词层不动：
|
||||
|
||||
1. **统一走 `/v1/images/generations`**（有无参考图都走它；无参考图就不带 `image` 字段）：
|
||||
```python
|
||||
body = {"model": model, "prompt": prompt, "n": 1,
|
||||
"size": size, "async": True}
|
||||
if refs:
|
||||
body["image"] = [data_uri, ...] # data-URI 数组（≤2 张，沿用现选图逻辑）
|
||||
resp = post(f"{base}/v1/images/generations", json=body)
|
||||
task_id = resp.json()["task_id"]
|
||||
```
|
||||
2. **新增任务轮询**：`GET {origin}/v1/tasks/{task_id}`（origin = base 去掉 `/draw`）；
|
||||
3s 起步、逐步加到 10s，上限沿用 `poll_max_wait`（600s，gpt 高质量单张 1–5 分钟）；
|
||||
完成判定 = `data` 出现（不能依赖 `status == "completed"`）；失败态 = `status` 为
|
||||
failed/error/cancelled；然后下载 `data[0].url`。
|
||||
3. **参数清理**：删 `quality` / `output_format` / `input_fidelity`（均非文档参数；
|
||||
`input_fidelity` 的探测-降级机制整体移除）。`size` 改传像素串
|
||||
`"1536x2048"`（3:4）/ `"2048x2048"`（1:1）——比例枚举里没有 3:4，像素串是文档允许的写法。
|
||||
4. **重试保留**：提交/轮询遇到 429/5xx/超时，沿用 60→120→240s 退避（`rightapi_max_retries`）。
|
||||
5. **配置**：`RIGHTAPI_BASE_URL` 保持 `https://rightapi.ai/draw` 不变；`rightapi_image_quality`
|
||||
配置项删除（或停用）。
|
||||
|
||||
预计工作量：`_rightapi_request` 重写约 60 行 + 轮询函数 30 行，其余层（提示词分发、
|
||||
任务执行器、前端）零改动。
|
||||
|
||||
## 7. 上线前待确认项
|
||||
|
||||
1. **3:4 像素串 `1536x2048` 是否被接受**——探测只验证了 `size: "1:1"`（文档说像素串
|
||||
合法，但建议改完后先出 1 张 Ozon 规格图验证）；
|
||||
2. nano-banana / nano-banana-2 / nano-banana-pro 三个型号未逐一实测（同族接口一致，
|
||||
lite / gpt 系已验证通路，风险低）；
|
||||
3. 是否启用 `imageSize`（2K/4K，仅 nano-banana 与 gpt-image vip 支持）——默认不传，
|
||||
需要高清再说；
|
||||
4. Gemini 原生端点（`:generateContent`）本期不接，留作后续选项。
|
||||
@@ -894,7 +894,8 @@ const App: React.FC = () => {
|
||||
{' '}· 风格「{STYLE_SET_OPTIONS.find(s => s.value === suite.style_set)?.label ?? '自定义'}」
|
||||
</div>
|
||||
<div className="result-grid">
|
||||
{suite.images.map(img => (
|
||||
{/* 只渲染终态（ok/failed）格子：生成中的 pending 占位不出现，避免"…/✗"占位格被误点 */}
|
||||
{suite.images.filter(img => img.status === 'ok' || img.status === 'failed').map(img => (
|
||||
<div
|
||||
key={img.type_id + img.name}
|
||||
className={`result-cell ${img.status !== 'ok' ? 'fail' : ''}`}
|
||||
@@ -907,7 +908,7 @@ const App: React.FC = () => {
|
||||
) : (
|
||||
<div style={{ aspectRatio: suite.ratio === '3:4' ? '3/4' : '1', background: 'var(--card-soft)' }} />
|
||||
)}
|
||||
{img.status !== 'ok' && <span className="fail-tag">{img.status === 'failed' ? '✗' : '…'}</span>}
|
||||
{img.status !== 'ok' && <span className="fail-tag">✗</span>}
|
||||
<div className="cap" title={img.error || img.name}>
|
||||
{img.status === 'failed' && img.error ? img.error : img.name}
|
||||
</div>
|
||||
|
||||
@@ -4,7 +4,7 @@
|
||||
*/
|
||||
import type { ImageMaterial } from '../collector/scan';
|
||||
|
||||
/** 服务端支持的套图类型（与 server/services/prompt.py 保持一致） */
|
||||
/** 服务端支持的套图类型（与 server/services/prompts/common.py 保持一致） */
|
||||
export const SUITE_TYPE_OPTIONS = [
|
||||
{ value: 'white_bg', label: '白底主图' },
|
||||
{ value: 'key_features', label: '核心卖点图' },
|
||||
|
||||
@@ -11,7 +11,7 @@ from schemas import (
|
||||
)
|
||||
from services.generator import run_suite
|
||||
from services.planner import generate_plan
|
||||
from services.prompt import type_name
|
||||
from services.prompts import type_name
|
||||
from services.tasks import create_task
|
||||
|
||||
router = APIRouter(prefix="/api", tags=["generate"])
|
||||
|
||||
@@ -16,8 +16,8 @@ import httpx
|
||||
|
||||
from config import get_settings
|
||||
from services import storage
|
||||
from services.prompt import build_prompt, build_context, type_name, wrap_prompt_for_gpt_edits
|
||||
from services.tasks import Task, TaskImage, TASK_FAILED, TASK_RUNNING, TASK_DONE, TASK_PARTIAL, IMG_OK
|
||||
from services.prompts import build_prompt, build_context, type_name
|
||||
from services.tasks import Task, TaskImage, TASK_FAILED, TASK_RUNNING, TASK_DONE, TASK_PARTIAL, IMG_FAILED, IMG_OK
|
||||
|
||||
log = logging.getLogger("suite.generator")
|
||||
|
||||
@@ -396,13 +396,11 @@ async def run_suite(task: Task) -> None:
|
||||
image = TaskImage(type_id=type_id, name=job.get("title") or type_name(type_id))
|
||||
task.images.append(image)
|
||||
try:
|
||||
# 提示词按模型家族分发：国产主体参考 / gpt edits 保真 / google 主体保持
|
||||
prompt = build_prompt(
|
||||
type_id, ctx, task.style_set, task.lang,
|
||||
provider_name, model, type_id, ctx, task.style_set, task.lang,
|
||||
extra=job, style_prompt=task.style_prompt, requirements=task.requirements,
|
||||
)
|
||||
# gpt-image edits 语义：商品冻结契约前置（含商品文字锚定），防止风格词改商品
|
||||
if provider_name == "rightapi":
|
||||
prompt = wrap_prompt_for_gpt_edits(prompt, ctx)
|
||||
refs = _refs_for_job(list(task.ref_images or []), job)
|
||||
data = await generator(prompt, refs, size=size, model=model)
|
||||
# 部分中转不遵守 output_format（要 jpeg 回 PNG），按魔数定扩展名
|
||||
@@ -413,6 +411,7 @@ async def run_suite(task: Task) -> None:
|
||||
ok += 1
|
||||
except Exception as exc: # noqa: BLE001
|
||||
log.exception("套图 %s 类型 %s 生成失败", task.id, type_id)
|
||||
image.status = IMG_FAILED # 默认 pending，失败显式置 failed
|
||||
image.error = str(exc)[:500]
|
||||
failures.append(f"{job.get('title') or type_name(type_id)}：{str(exc)[:200]}")
|
||||
failed += 1
|
||||
|
||||
@@ -1,353 +0,0 @@
|
||||
"""套图 Prompt 引擎。
|
||||
|
||||
借鉴 ecommerce-image-suite 的动态 Prompt 架构，浓缩为：
|
||||
- 7 种图类型 × 5 套视觉风格模板
|
||||
- 公共组件：QUALITY（画质）/ PRODUCT_REF_LOCK（商品一致性锁）/ TEXT_RENDER（图内文案规范）
|
||||
- 卖点从采集的参数表/卖点文本自动提炼
|
||||
|
||||
核心原则：所有图严格保持商品一致性（same silhouette, same print, same color），
|
||||
只允许改变背景 / 角度 / 光线 / 排版。
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import re
|
||||
|
||||
# ── 风格模板（与插件端 STYLE_SET_OPTIONS 对应；提示词可被用户在插件里改写覆盖）───
|
||||
# 提示词用中文：生图 provider（通义万相/豆包）均为国产模型，中文理解一流，且便于用户自行改写。
|
||||
|
||||
STYLE_SETS: dict[int, dict] = {
|
||||
1: {
|
||||
"name": "北欧极简",
|
||||
"tone": "北欧极简风：浅灰或米白背景，柔和漫射光，低饱和色调，画面留白充足，构图克制干净",
|
||||
"bg": "",
|
||||
},
|
||||
2: {
|
||||
"name": "清新明亮",
|
||||
"tone": "清新明亮风：明亮的白色到浅蓝渐变背景，高调光线，色彩明快通透，整体轻盈干净",
|
||||
"bg": "",
|
||||
},
|
||||
3: {
|
||||
"name": "高级感深色",
|
||||
"tone": "高级质感风：深灰或炭黑背景，戏剧性侧光打光，突出商品材质与光泽，沉稳高级",
|
||||
"bg": "",
|
||||
},
|
||||
4: {
|
||||
"name": "暖调生活",
|
||||
"tone": "温暖生活风：暖米色背景，暖色灯光氛围，温馨的家居质感，亲和力强",
|
||||
"bg": "",
|
||||
},
|
||||
5: {
|
||||
"name": "纯净棚拍",
|
||||
"tone": "标准电商棚拍：纯色浅背景，均匀的正面柔光，无杂物干扰，商品居中突出",
|
||||
"bg": "",
|
||||
},
|
||||
}
|
||||
|
||||
# ── 图类型中文名（导出文件名用）───────────────────────────────────────────
|
||||
|
||||
TYPE_NAMES_ZH: dict[str, str] = {
|
||||
"white_bg": "白底主图",
|
||||
"key_features": "核心卖点图",
|
||||
"selling_pt": "卖点图",
|
||||
"material": "材质图",
|
||||
"lifestyle": "场景展示图",
|
||||
"multi_scene": "多场景拼图",
|
||||
"ecommerce_detail": "电商详情图",
|
||||
"size_chart": "尺寸标注图",
|
||||
"sku_collection": "SKU合集图",
|
||||
"custom": "创意图",
|
||||
}
|
||||
|
||||
# ── 公共组件 ──────────────────────────────────────────────────────────────
|
||||
|
||||
QUALITY = (
|
||||
"Shot on Sony A7R V with 85mm lens at f/2.0, ultra-detailed, photorealistic, "
|
||||
"8K commercial image quality, professional retouching."
|
||||
)
|
||||
|
||||
PRODUCT_REF_LOCK = (
|
||||
"CRITICAL: The product must look EXACTLY the same as in the reference image — "
|
||||
"identical silhouette, proportions, colors, print pattern, stitching and every design detail. "
|
||||
"Only the background, camera angle, lighting and styling may change. "
|
||||
"Do not redesign, add or remove any element of the product."
|
||||
)
|
||||
|
||||
TEXT_RENDER = {
|
||||
"zh": (
|
||||
"Render concise Chinese marketing text inside the image: main headline max 8 Chinese characters, "
|
||||
"sub-lines max 12 characters each, font is modern clean sans-serif (Source Han Sans style), "
|
||||
"high legibility, tasteful typography layout, colors harmonized with the composition. "
|
||||
"No spelling errors, no garbled characters."
|
||||
),
|
||||
"en": (
|
||||
"Render concise English marketing text inside the image: headline max 5 words, "
|
||||
"sub-lines max 8 words each, Helvetica Neue style sans-serif, high legibility, "
|
||||
"tasteful typography layout, colors harmonized with the composition. No spelling errors."
|
||||
),
|
||||
"ru": (
|
||||
"Render concise Russian marketing text inside the image: headline max 4 words, "
|
||||
"sub-lines max 6 words each, modern clean sans-serif (Inter / PT Sans style), "
|
||||
"proper Cyrillic typography, high legibility, tasteful layout, colors harmonized with the composition. "
|
||||
"No spelling errors, no mixed latin/cyrillic gibberish."
|
||||
),
|
||||
}
|
||||
|
||||
DEFAULT_NEGATIVE_INTENT = (
|
||||
"no AI-generated look, no CGI quality, no plastic appearance, no watermark, "
|
||||
"no distorted text, no deformed product, no extra limbs, no blurry areas"
|
||||
)
|
||||
|
||||
# ── gpt-image（/v1/images/edits 语义）专用包装 ─────────────────────────────
|
||||
# gpt-image 的 edits 端点把输入图当"被编辑的底图"、prompt 当"编辑指令"（豆包/通义则是
|
||||
# "主体参考"），风格词会被字面执行到商品上。按 OpenAI 官方提示词指南的编辑模式：
|
||||
# 按序号说明输入图、PRESERVE/MAY CHANGE 分列、首尾重申不变量、文案逐字渲染。
|
||||
|
||||
def gpt_edits_contract(ctx: dict) -> str:
|
||||
"""gpt-image edits 语义契约（放开头，指令权重最高处）。
|
||||
|
||||
除通用锁定条款外，注入商品文字锚定（标题 + 关键参数 + 描述）：
|
||||
官逆通道（gpt-image-2-vip 等）会把参考图当对话附件弱化处理，
|
||||
input_fidelity 类 API 参数不生效，此时商品文字描述是保真的唯一兜底。
|
||||
"""
|
||||
anchor = f" The product is: \"{ctx['title']}\""
|
||||
if ctx.get("params_line"):
|
||||
anchor += f" (key specs: {ctx['params_line']})"
|
||||
if ctx.get("desc"):
|
||||
anchor += f". {ctx['desc']}"
|
||||
return (
|
||||
"TASK: Edit the attached product photos — re-photograph THE SAME physical product "
|
||||
"in a new setting. This is an edit of the input images, NOT a new product design.\n"
|
||||
"INPUT IMAGES: Image 1 = product front view (PRIMARY source of truth for the product's "
|
||||
f"true appearance); Image 2 (if present) = product back / detail view.{anchor}\n"
|
||||
"PRODUCT LOCK (highest priority, overrides everything else in this prompt): exactly "
|
||||
"preserve the product's shape, silhouette, proportions, colors, label text, logos, "
|
||||
"print/pattern, materials and texture. Do not redesign, restyle, recolor, re-pattern "
|
||||
"or substitute the product with a similar one. The output must show the very product "
|
||||
"from the input images AND match the product description above; if the generated "
|
||||
"product differs from either in any design detail, the image is rejected.\n"
|
||||
"MAY CHANGE: background, scene, props, camera angle, lighting, composition "
|
||||
"and in-image marketing typography only. You may relight the product so it sits "
|
||||
"naturally in the new scene (matched shadows and color temperature).\n"
|
||||
"STYLE SCOPE: all style, mood, color-palette and decoration instructions below describe "
|
||||
"the SCENE AND BACKGROUND ONLY — never apply them to the product. "
|
||||
"When any style instruction conflicts with product fidelity, product fidelity always wins."
|
||||
)
|
||||
|
||||
|
||||
GPT_EDITS_FINAL_CHECK = (
|
||||
"FINAL CHECK before output: if the product in your result differs from the input product "
|
||||
"or the product description above in any design detail (shape, color, pattern, material, "
|
||||
"logo), the image is rejected. Render any listed marketing copy / headlines exactly as "
|
||||
"written (verbatim, no extra characters, no paraphrasing)."
|
||||
)
|
||||
|
||||
|
||||
def wrap_prompt_for_gpt_edits(prompt: str, ctx: dict) -> str:
|
||||
"""gpt-image edits 语义适配：契约放开头（指令权重最高处），终检放结尾。"""
|
||||
return f"{gpt_edits_contract(ctx)}\n\n{prompt}\n\n{GPT_EDITS_FINAL_CHECK}"
|
||||
|
||||
# ── 商品上下文提炼 ────────────────────────────────────────────────────────
|
||||
|
||||
def _shorten(text: str, n: int) -> str:
|
||||
text = re.sub(r"\s+", " ", (text or "")).strip()
|
||||
return text[:n]
|
||||
|
||||
def _clean_title(title: str) -> str:
|
||||
"""去掉常见堆砌词，让标题更可读。"""
|
||||
t = _shorten(title, 60)
|
||||
return re.sub(r"[【【】】\\[\\]|/]", " ", t).strip()
|
||||
|
||||
def build_context(raw: dict, fallback_name: str = "", fallback_desc: str = "") -> dict:
|
||||
"""从采集数据提炼生图上下文：标题、描述行、卖点列表、参数行。
|
||||
|
||||
raw: {title, desc, price, params: [{key, value}], sellingPoints}
|
||||
"""
|
||||
title = _clean_title(raw.get("title") or fallback_name or "product")
|
||||
desc = _shorten(raw.get("desc") or fallback_desc or "", 200)
|
||||
|
||||
# 卖点：优先显式卖点文本；否则从参数表里挑短而有信息量的键值对
|
||||
selling_points: list[dict] = []
|
||||
sp_text = raw.get("sellingPoints") or ""
|
||||
if sp_text:
|
||||
for chunk in re.split(r"[；;\n·]+|(?<!\d)\.(?!\d)", sp_text):
|
||||
c = _shorten(chunk, 20)
|
||||
if c and len(selling_points) < 5:
|
||||
selling_points.append({"zh": c, "en": c})
|
||||
if not selling_points:
|
||||
for p in (raw.get("params") or [])[:12]:
|
||||
k, v = _shorten(p.get("key", ""), 10), _shorten(str(p.get("value", "")), 16)
|
||||
if k and v and k.lower() not in {"货号", "sku", "isbn", "上架时间"}:
|
||||
selling_points.append({"zh": f"{k} {v}", "en": f"{k} {v}"})
|
||||
if len(selling_points) >= 5:
|
||||
break
|
||||
|
||||
params_line = "; ".join(
|
||||
f"{p.get('key')}: {p.get('value')}" for p in (raw.get("params") or [])[:8]
|
||||
)
|
||||
return {
|
||||
"title": title,
|
||||
"title_en": title, # 采集源多为中文标题，英文场景直接用原词避免乱翻译
|
||||
"desc": desc,
|
||||
"selling_points": selling_points[:3],
|
||||
"params_line": params_line,
|
||||
"price": raw.get("price") or "",
|
||||
}
|
||||
|
||||
def _sp_lines(ctx: dict, lang: str, max_n: int = 3) -> str:
|
||||
sps = ctx["selling_points"][:max_n]
|
||||
if not sps:
|
||||
return ""
|
||||
key = "zh" if lang == "zh" else "en"
|
||||
return "; ".join(s[key] for s in sps if s.get(key))
|
||||
|
||||
# ── 各图类型 Prompt ───────────────────────────────────────────────────────
|
||||
|
||||
def _prompt_white_bg(ctx: dict, style: dict, lang: str) -> str:
|
||||
return (
|
||||
f"E-commerce main product image on pure white background (RGB 255,255,255), "
|
||||
f"product \"{ctx['title']}\" centered and filling about 85% of the frame, "
|
||||
f"front view, even shadowless studio lighting with a faint natural contact shadow, "
|
||||
f"{style['tone']}. No text, no props, no background elements. {QUALITY} {PRODUCT_REF_LOCK}"
|
||||
)
|
||||
|
||||
def _prompt_key_features(ctx: dict, style: dict, lang: str) -> str:
|
||||
sp = _sp_lines(ctx, lang) or ctx["title"]
|
||||
return (
|
||||
f"E-commerce key-features infographic for product \"{ctx['title']}\", square layout: "
|
||||
f"product on the left two-thirds ({style['bg']}), right column lists 3 feature callouts "
|
||||
f"with minimal line icons, thin leader lines pointing to product details. "
|
||||
f"Feature callouts: {sp}. {style['tone']}. {TEXT_RENDER[lang]} {QUALITY} {PRODUCT_REF_LOCK}"
|
||||
)
|
||||
|
||||
def _prompt_selling_pt(ctx: dict, style: dict, lang: str) -> str:
|
||||
sp = _sp_lines(ctx, lang, 1) or ctx["title"]
|
||||
return (
|
||||
f"Single-selling-point e-commerce poster for product \"{ctx['title']}\": "
|
||||
f"hero product close-up at dynamic angle ({style['bg']}), one large bold headline "
|
||||
f"about \"{sp}\", generous negative space, one small magnified detail circle "
|
||||
f"highlighting material or craft. {style['tone']}. {TEXT_RENDER[lang]} {QUALITY} {PRODUCT_REF_LOCK}"
|
||||
)
|
||||
|
||||
def _prompt_material(ctx: dict, style: dict, lang: str) -> str:
|
||||
return (
|
||||
f"Macro material close-up of product \"{ctx['title']}\": extreme detail shot revealing "
|
||||
f"fabric weave / surface texture / stitching / finish, shallow depth of field, "
|
||||
f"raking light across the surface, {style['tone']}. Small caption label in corner. "
|
||||
f"{TEXT_RENDER[lang]} {QUALITY} {PRODUCT_REF_LOCK}"
|
||||
)
|
||||
|
||||
def _prompt_lifestyle(ctx: dict, style: dict, lang: str) -> str:
|
||||
bg = f" ({style['bg']})" if style.get("bg") else ""
|
||||
return (
|
||||
f"Lifestyle in-context scene for product \"{ctx['title']}\": the product is naturally "
|
||||
f"used / placed in a real environment{bg}, realistic human-scale surroundings, "
|
||||
f"soft daylight, authentic candid mood, product remains the clear visual focus. "
|
||||
f"{style['tone']}. {TEXT_RENDER[lang]} {QUALITY} {PRODUCT_REF_LOCK}"
|
||||
)
|
||||
|
||||
def _prompt_multi_scene(ctx: dict, style: dict, lang: str) -> str:
|
||||
sp = _sp_lines(ctx, lang)
|
||||
return (
|
||||
f"Triptych multi-scene e-commerce image for product \"{ctx['title']}\": three vertical panels "
|
||||
f"separated by thin gutters, each panel shows the SAME product in a different usage scene "
|
||||
f"(e.g. home interior / outdoor street / office desk), consistent color grading across panels. "
|
||||
f"Panel captions: {sp}. {style['tone']}. {TEXT_RENDER[lang]} {QUALITY} {PRODUCT_REF_LOCK}"
|
||||
)
|
||||
|
||||
def _prompt_ecommerce_detail(ctx: dict, style: dict, lang: str) -> str:
|
||||
sp = _sp_lines(ctx, lang) or ctx["title"]
|
||||
params = ctx["params_line"]
|
||||
return (
|
||||
f"E-commerce detail-page hero section for product \"{ctx['title']}\", square layout: "
|
||||
f"top half is a hero banner with the product at a 3/4 angle ({style['bg']}); "
|
||||
f"bottom half is a clean spec card listing 3 feature rows with line icons"
|
||||
+ (f" (specs: {params})" if params else "")
|
||||
+ f" and one highlighted row: {sp}. {style['tone']}. {TEXT_RENDER[lang]} {QUALITY} {PRODUCT_REF_LOCK}"
|
||||
)
|
||||
|
||||
def _prompt_size_chart(ctx: dict, style: dict, lang: str) -> str:
|
||||
dims = ctx["params_line"]
|
||||
return (
|
||||
f"Product size chart infographic for \"{ctx['title']}\": product shown in clean front and side views "
|
||||
f"on light background, with thin measurement annotation lines (arrows) marking length, width and height, "
|
||||
f"measurement values rendered next to each line"
|
||||
+ (f" (known specs: {dims})" if dims else "")
|
||||
+ f", small caption row, precise technical drawing aesthetic. {style['tone']}. "
|
||||
f"{TEXT_RENDER[lang]} {QUALITY} {PRODUCT_REF_LOCK}"
|
||||
)
|
||||
|
||||
def _prompt_sku_collection(ctx: dict, style: dict, lang: str) -> str:
|
||||
return (
|
||||
f"Colorway collection image for product \"{ctx['title']}\": the SAME product in all its color/variant "
|
||||
f"options arranged in a neat equal grid (2-4 items per row), each colorway with a small label chip below it, "
|
||||
f"consistent lighting and scale across all items, clean e-commerce presentation. "
|
||||
f"{style['tone']}. {TEXT_RENDER[lang]} {QUALITY} {PRODUCT_REF_LOCK}"
|
||||
)
|
||||
|
||||
def _prompt_custom(ctx: dict, style: dict, lang: str, extra: dict) -> str:
|
||||
hint = (extra.get("prompt_hint") or "").strip()
|
||||
purpose = extra.get("title") or ""
|
||||
detail = extra.get("detail") or ""
|
||||
bg = f" {style['bg']} as environment." if style.get("bg") else ""
|
||||
composed = (
|
||||
f"E-commerce marketing image for product \"{ctx['title']}\""
|
||||
+ (f" — {purpose}" if purpose else "")
|
||||
+ (f": {detail}" if detail else "")
|
||||
+ "."
|
||||
)
|
||||
if hint:
|
||||
composed += f" Composition: {hint}."
|
||||
return f"{composed}{bg} {style['tone']}. {TEXT_RENDER[lang]} {QUALITY} {PRODUCT_REF_LOCK}"
|
||||
|
||||
_PROMPT_BUILDERS = {
|
||||
"white_bg": _prompt_white_bg,
|
||||
"key_features": _prompt_key_features,
|
||||
"selling_pt": _prompt_selling_pt,
|
||||
"material": _prompt_material,
|
||||
"lifestyle": _prompt_lifestyle,
|
||||
"multi_scene": _prompt_multi_scene,
|
||||
"ecommerce_detail": _prompt_ecommerce_detail,
|
||||
"size_chart": _prompt_size_chart,
|
||||
"sku_collection": _prompt_sku_collection,
|
||||
}
|
||||
|
||||
|
||||
def build_prompt(type_id: str, ctx: dict, style_set: int, lang: str, extra: dict | None = None,
|
||||
style_prompt: str | None = None, requirements: str | None = None) -> str:
|
||||
"""构造指定图类型的完整生图 prompt。
|
||||
|
||||
extra: 方案项信息 {title, detail, prompt_hint}——custom 类型必需，
|
||||
预设类型也会把 prompt_hint 作为构图补充注入。
|
||||
style_prompt: 用户改写的风格提示词，覆盖 style_set 内置模板（tone/bg 整体替换）。
|
||||
requirements: 生图要求（最高优先级，强制约束），置于 prompt 最前面，
|
||||
声明覆盖一切冲突指令，用户可在此输入强制要求。
|
||||
"""
|
||||
if style_prompt and style_prompt.strip():
|
||||
style = {"name": "custom", "tone": style_prompt.strip(), "bg": ""}
|
||||
else:
|
||||
style = STYLE_SETS.get(style_set, STYLE_SETS[1])
|
||||
extra = extra or {}
|
||||
if type_id == "custom":
|
||||
prompt = _prompt_custom(ctx, style, lang, extra)
|
||||
else:
|
||||
builder = _PROMPT_BUILDERS.get(type_id)
|
||||
if builder is None:
|
||||
raise ValueError(f"未知图类型: {type_id}")
|
||||
prompt = builder(ctx, style, lang)
|
||||
hint = (extra.get("prompt_hint") or "").strip()
|
||||
if hint:
|
||||
prompt = prompt.rstrip(".") + f". Additional composition guidance: {hint}."
|
||||
# 生图要求：最高优先级，置于最前并声明覆盖冲突指令（用户输入原样保留，不翻译）
|
||||
if requirements and requirements.strip():
|
||||
prompt = (
|
||||
"STRICT REQUIREMENTS (highest priority, must be followed exactly, "
|
||||
"override any conflicting instruction): "
|
||||
+ requirements.strip().rstrip(".")
|
||||
+ ". "
|
||||
+ prompt
|
||||
)
|
||||
return prompt + ". " + DEFAULT_NEGATIVE_INTENT
|
||||
|
||||
|
||||
def type_name(type_id: str) -> str:
|
||||
return TYPE_NAMES_ZH.get(type_id, type_id)
|
||||
@@ -0,0 +1,45 @@
|
||||
"""套图提示词引擎：按模型家族分发，各家族独立封装。
|
||||
|
||||
不同家族的生图语义差异极大，共用一套提示词会导致语义错配
|
||||
（gpt-image-2 按文字重造商品即由此而来），故按家族各自成册：
|
||||
|
||||
alibaba 通义 wan*/qwen*（DashScope）—— 主体参考语义
|
||||
doubao 豆包 Seedream（火山方舟）—— 主体参考语义，与通义共用装配
|
||||
gpt gpt-image-2 / gpt-image-2-vip（RightAPI）—— /v1/images/edits 编辑语义
|
||||
google nano-banana 系列（RightAPI）—— 原生主体保持语义
|
||||
|
||||
路由规则：provider 为主；rightapi 内再按模型名细分 gpt / google。
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
from . import alibaba, doubao, google, gpt
|
||||
from .common import build_context, type_name
|
||||
|
||||
_MODULE_BY_FAMILY = {
|
||||
"alibaba": alibaba,
|
||||
"doubao": doubao,
|
||||
"gpt": gpt,
|
||||
"google": google,
|
||||
}
|
||||
|
||||
|
||||
def prompt_family(provider: str, model: str | None) -> str:
|
||||
"""(provider, model) → 提示词家族名。"""
|
||||
if provider == "rightapi":
|
||||
if (model or "").lower().startswith("nano-banana"):
|
||||
return "google"
|
||||
return "gpt" # gpt-image-* 及未知中转模型默认按 edits 语义处理
|
||||
if provider == "tongyi":
|
||||
return "alibaba"
|
||||
return "doubao" # doubao 及默认 provider
|
||||
|
||||
|
||||
def build_prompt(provider: str, model: str | None, type_id: str, ctx: dict, style_set: int,
|
||||
lang: str, extra: dict | None = None, style_prompt: str | None = None,
|
||||
requirements: str | None = None) -> str:
|
||||
"""按模型家族构造指定图类型的完整生图 prompt。参数含义见各家族 build_prompt。"""
|
||||
module = _MODULE_BY_FAMILY[prompt_family(provider, model)]
|
||||
return module.build_prompt(
|
||||
type_id, ctx, style_set, lang,
|
||||
extra=extra, style_prompt=style_prompt, requirements=requirements,
|
||||
)
|
||||
@@ -0,0 +1,169 @@
|
||||
"""阿里通义（wan* 万相 / qwen* 千问）提示词：国产"主体参考"语义。
|
||||
|
||||
生图 API 把参考图当商品锚（subject reference）、prompt 当场景描述，
|
||||
风格词/文字商品描述不会反噬商品本体，负面清单也可以安全写入 prompt。
|
||||
豆包（doubao.py）与此语义一致，直接复用本模块装配。
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
from .common import (
|
||||
STYLE_SETS, TEXT_RENDER, requirements_block, resolve_style, selling_point_lines,
|
||||
)
|
||||
|
||||
# ── 公共组件（主体参考语义专用）────────────────────────────────────────────
|
||||
|
||||
QUALITY = (
|
||||
"Shot on Sony A7R V with 85mm lens at f/2.0, ultra-detailed, photorealistic, "
|
||||
"8K commercial image quality, professional retouching."
|
||||
)
|
||||
|
||||
PRODUCT_REF_LOCK = (
|
||||
"CRITICAL: The product must look EXACTLY the same as in the reference image — "
|
||||
"identical silhouette, proportions, colors, print pattern, stitching and every design detail. "
|
||||
"Only the background, camera angle, lighting and styling may change. "
|
||||
"Do not redesign, add or remove any element of the product."
|
||||
)
|
||||
|
||||
DEFAULT_NEGATIVE_INTENT = (
|
||||
"no AI-generated look, no CGI quality, no plastic appearance, no watermark, "
|
||||
"no distorted text, no deformed product, no extra limbs, no blurry areas"
|
||||
)
|
||||
|
||||
# ── 各图类型 Prompt ───────────────────────────────────────────────────────
|
||||
|
||||
def _prompt_white_bg(ctx: dict, style: dict, lang: str) -> str:
|
||||
return (
|
||||
f"E-commerce main product image on pure white background (RGB 255,255,255), "
|
||||
f"product \"{ctx['title']}\" centered and filling about 85% of the frame, "
|
||||
f"front view, even shadowless studio lighting with a faint natural contact shadow, "
|
||||
f"{style['tone']}. No text, no props, no background elements. {QUALITY} {PRODUCT_REF_LOCK}"
|
||||
)
|
||||
|
||||
def _prompt_key_features(ctx: dict, style: dict, lang: str) -> str:
|
||||
sp = selling_point_lines(ctx, lang) or ctx["title"]
|
||||
return (
|
||||
f"E-commerce key-features infographic for product \"{ctx['title']}\", square layout: "
|
||||
f"product on the left two-thirds ({style['bg']}), right column lists 3 feature callouts "
|
||||
f"with minimal line icons, thin leader lines pointing to product details. "
|
||||
f"Feature callouts: {sp}. {style['tone']}. {TEXT_RENDER[lang]} {QUALITY} {PRODUCT_REF_LOCK}"
|
||||
)
|
||||
|
||||
def _prompt_selling_pt(ctx: dict, style: dict, lang: str) -> str:
|
||||
sp = selling_point_lines(ctx, lang, 1) or ctx["title"]
|
||||
return (
|
||||
f"Single-selling-point e-commerce poster for product \"{ctx['title']}\": "
|
||||
f"hero product close-up at dynamic angle ({style['bg']}), one large bold headline "
|
||||
f"about \"{sp}\", generous negative space, one small magnified detail circle "
|
||||
f"highlighting material or craft. {style['tone']}. {TEXT_RENDER[lang]} {QUALITY} {PRODUCT_REF_LOCK}"
|
||||
)
|
||||
|
||||
def _prompt_material(ctx: dict, style: dict, lang: str) -> str:
|
||||
return (
|
||||
f"Macro material close-up of product \"{ctx['title']}\": extreme detail shot revealing "
|
||||
f"fabric weave / surface texture / stitching / finish, shallow depth of field, "
|
||||
f"raking light across the surface, {style['tone']}. Small caption label in corner. "
|
||||
f"{TEXT_RENDER[lang]} {QUALITY} {PRODUCT_REF_LOCK}"
|
||||
)
|
||||
|
||||
def _prompt_lifestyle(ctx: dict, style: dict, lang: str) -> str:
|
||||
bg = f" ({style['bg']})" if style.get("bg") else ""
|
||||
return (
|
||||
f"Lifestyle in-context scene for product \"{ctx['title']}\": the product is naturally "
|
||||
f"used / placed in a real environment{bg}, realistic human-scale surroundings, "
|
||||
f"soft daylight, authentic candid mood, product remains the clear visual focus. "
|
||||
f"{style['tone']}. {TEXT_RENDER[lang]} {QUALITY} {PRODUCT_REF_LOCK}"
|
||||
)
|
||||
|
||||
def _prompt_multi_scene(ctx: dict, style: dict, lang: str) -> str:
|
||||
sp = selling_point_lines(ctx, lang)
|
||||
return (
|
||||
f"Triptych multi-scene e-commerce image for product \"{ctx['title']}\": three vertical panels "
|
||||
f"separated by thin gutters, each panel shows the SAME product in a different usage scene "
|
||||
f"(e.g. home interior / outdoor street / office desk), consistent color grading across panels. "
|
||||
f"Panel captions: {sp}. {style['tone']}. {TEXT_RENDER[lang]} {QUALITY} {PRODUCT_REF_LOCK}"
|
||||
)
|
||||
|
||||
def _prompt_ecommerce_detail(ctx: dict, style: dict, lang: str) -> str:
|
||||
sp = selling_point_lines(ctx, lang) or ctx["title"]
|
||||
params = ctx["params_line"]
|
||||
return (
|
||||
f"E-commerce detail-page hero section for product \"{ctx['title']}\", square layout: "
|
||||
f"top half is a hero banner with the product at a 3/4 angle ({style['bg']}); "
|
||||
f"bottom half is a clean spec card listing 3 feature rows with line icons"
|
||||
+ (f" (specs: {params})" if params else "")
|
||||
+ f" and one highlighted row: {sp}. {style['tone']}. {TEXT_RENDER[lang]} {QUALITY} {PRODUCT_REF_LOCK}"
|
||||
)
|
||||
|
||||
def _prompt_size_chart(ctx: dict, style: dict, lang: str) -> str:
|
||||
dims = ctx["params_line"]
|
||||
return (
|
||||
f"Product size chart infographic for \"{ctx['title']}\": product shown in clean front and side views "
|
||||
f"on light background, with thin measurement annotation lines (arrows) marking length, width and height, "
|
||||
f"measurement values rendered next to each line"
|
||||
+ (f" (known specs: {dims})" if dims else "")
|
||||
+ f", small caption row, precise technical drawing aesthetic. {style['tone']}. "
|
||||
f"{TEXT_RENDER[lang]} {QUALITY} {PRODUCT_REF_LOCK}"
|
||||
)
|
||||
|
||||
def _prompt_sku_collection(ctx: dict, style: dict, lang: str) -> str:
|
||||
return (
|
||||
f"Colorway collection image for product \"{ctx['title']}\": the SAME product in all its color/variant "
|
||||
f"options arranged in a neat equal grid (2-4 items per row), each colorway with a small label chip below it, "
|
||||
f"consistent lighting and scale across all items, clean e-commerce presentation. "
|
||||
f"{style['tone']}. {TEXT_RENDER[lang]} {QUALITY} {PRODUCT_REF_LOCK}"
|
||||
)
|
||||
|
||||
def _prompt_custom(ctx: dict, style: dict, lang: str, extra: dict) -> str:
|
||||
hint = (extra.get("prompt_hint") or "").strip()
|
||||
purpose = extra.get("title") or ""
|
||||
detail = extra.get("detail") or ""
|
||||
bg = f" {style['bg']} as environment." if style.get("bg") else ""
|
||||
composed = (
|
||||
f"E-commerce marketing image for product \"{ctx['title']}\""
|
||||
+ (f" — {purpose}" if purpose else "")
|
||||
+ (f": {detail}" if detail else "")
|
||||
+ "."
|
||||
)
|
||||
if hint:
|
||||
composed += f" Composition: {hint}."
|
||||
return f"{composed}{bg} {style['tone']}. {TEXT_RENDER[lang]} {QUALITY} {PRODUCT_REF_LOCK}"
|
||||
|
||||
_PROMPT_BUILDERS = {
|
||||
"white_bg": _prompt_white_bg,
|
||||
"key_features": _prompt_key_features,
|
||||
"selling_pt": _prompt_selling_pt,
|
||||
"material": _prompt_material,
|
||||
"lifestyle": _prompt_lifestyle,
|
||||
"multi_scene": _prompt_multi_scene,
|
||||
"ecommerce_detail": _prompt_ecommerce_detail,
|
||||
"size_chart": _prompt_size_chart,
|
||||
"sku_collection": _prompt_sku_collection,
|
||||
}
|
||||
|
||||
|
||||
def build_prompt(type_id: str, ctx: dict, style_set: int, lang: str, extra: dict | None = None,
|
||||
style_prompt: str | None = None, requirements: str | None = None) -> str:
|
||||
"""构造指定图类型的完整生图 prompt（主体参考语义）。
|
||||
|
||||
extra: 方案项信息 {title, detail, prompt_hint}——custom 类型必需，
|
||||
预设类型也会把 prompt_hint 作为构图补充注入。
|
||||
style_prompt: 用户改写的风格提示词，覆盖 style_set 内置模板（tone/bg 整体替换）。
|
||||
requirements: 生图要求（最高优先级，强制约束），置于 prompt 最前面，
|
||||
声明覆盖一切冲突指令，用户可在此输入强制要求。
|
||||
"""
|
||||
style = resolve_style(style_set, style_prompt)
|
||||
extra = extra or {}
|
||||
if type_id == "custom":
|
||||
prompt = _prompt_custom(ctx, style, lang, extra)
|
||||
else:
|
||||
builder = _PROMPT_BUILDERS.get(type_id)
|
||||
if builder is None:
|
||||
raise ValueError(f"未知图类型: {type_id}")
|
||||
prompt = builder(ctx, style, lang)
|
||||
hint = (extra.get("prompt_hint") or "").strip()
|
||||
if hint:
|
||||
prompt = prompt.rstrip(".") + f". Additional composition guidance: {hint}."
|
||||
req = requirements_block(requirements)
|
||||
if req:
|
||||
prompt = f"{req} {prompt}"
|
||||
return prompt + ". " + DEFAULT_NEGATIVE_INTENT
|
||||
@@ -0,0 +1,154 @@
|
||||
"""提示词公共层：与模型家族无关的商品上下文、风格模板、图类型名与文案组件。
|
||||
|
||||
各家族模块（alibaba / doubao / gpt / google）只负责"如何对模型说话"，
|
||||
商品信息提炼与风格体系统一在这里维护，避免多处漂移。
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import re
|
||||
|
||||
# ── 风格模板（与插件端 STYLE_SET_OPTIONS 对应；提示词可被用户在插件里改写覆盖）───
|
||||
|
||||
STYLE_SETS: dict[int, dict] = {
|
||||
1: {
|
||||
"name": "北欧极简",
|
||||
"tone": "北欧极简风：浅灰或米白背景，柔和漫射光，低饱和色调，画面留白充足，构图克制干净",
|
||||
"bg": "",
|
||||
},
|
||||
2: {
|
||||
"name": "清新明亮",
|
||||
"tone": "清新明亮风：明亮的白色到浅蓝渐变背景，高调光线，色彩明快通透，整体轻盈干净",
|
||||
"bg": "",
|
||||
},
|
||||
3: {
|
||||
"name": "高级感深色",
|
||||
"tone": "高级质感风：深灰或炭黑背景，戏剧性侧光打光，突出商品材质与光泽，沉稳高级",
|
||||
"bg": "",
|
||||
},
|
||||
4: {
|
||||
"name": "暖调生活",
|
||||
"tone": "温暖生活风：暖米色背景，暖色灯光氛围，温馨的家居质感，亲和力强",
|
||||
"bg": "",
|
||||
},
|
||||
5: {
|
||||
"name": "纯净棚拍",
|
||||
"tone": "标准电商棚拍：纯色浅背景，均匀的正面柔光，无杂物干扰，商品居中突出",
|
||||
"bg": "",
|
||||
},
|
||||
}
|
||||
|
||||
# ── 图类型中文名（导出文件名用）───────────────────────────────────────────
|
||||
|
||||
TYPE_NAMES_ZH: dict[str, str] = {
|
||||
"white_bg": "白底主图",
|
||||
"key_features": "核心卖点图",
|
||||
"selling_pt": "卖点图",
|
||||
"material": "材质图",
|
||||
"lifestyle": "场景展示图",
|
||||
"multi_scene": "多场景拼图",
|
||||
"ecommerce_detail": "电商详情图",
|
||||
"size_chart": "尺寸标注图",
|
||||
"sku_collection": "SKU合集图",
|
||||
"custom": "创意图",
|
||||
}
|
||||
|
||||
# ── 图内营销文案渲染规范（各家族共用；语言由平台决定）──────────────────────
|
||||
|
||||
TEXT_RENDER = {
|
||||
"zh": (
|
||||
"Render concise Chinese marketing text inside the image: main headline max 8 Chinese characters, "
|
||||
"sub-lines max 12 characters each, font is modern clean sans-serif (Source Han Sans style), "
|
||||
"high legibility, tasteful typography layout, colors harmonized with the composition. "
|
||||
"No spelling errors, no garbled characters."
|
||||
),
|
||||
"en": (
|
||||
"Render concise English marketing text inside the image: headline max 5 words, "
|
||||
"sub-lines max 8 words each, Helvetica Neue style sans-serif, high legibility, "
|
||||
"tasteful typography layout, colors harmonized with the composition. No spelling errors."
|
||||
),
|
||||
"ru": (
|
||||
"Render concise Russian marketing text inside the image: headline max 4 words, "
|
||||
"sub-lines max 6 words each, modern clean sans-serif (Inter / PT Sans style), "
|
||||
"proper Cyrillic typography, high legibility, tasteful layout, colors harmonized with the composition. "
|
||||
"No spelling errors, no mixed latin/cyrillic gibberish."
|
||||
),
|
||||
}
|
||||
|
||||
|
||||
def resolve_style(style_set: int, style_prompt: str | None = None) -> dict:
|
||||
"""用户改写的风格提示词整体覆盖内置模板（tone/bg 整体替换）。"""
|
||||
if style_prompt and style_prompt.strip():
|
||||
return {"name": "custom", "tone": style_prompt.strip(), "bg": ""}
|
||||
return STYLE_SETS.get(style_set, STYLE_SETS[1])
|
||||
|
||||
|
||||
def requirements_block(requirements: str | None) -> str:
|
||||
"""用户强制要求块：最高优先级、置于提示词最前、覆盖冲突指令（原文保留不翻译）。"""
|
||||
if requirements and requirements.strip():
|
||||
return (
|
||||
"STRICT REQUIREMENTS (highest priority, must be followed exactly, "
|
||||
"override any conflicting instruction): "
|
||||
+ requirements.strip().rstrip(".")
|
||||
+ "."
|
||||
)
|
||||
return ""
|
||||
|
||||
|
||||
# ── 商品上下文提炼 ────────────────────────────────────────────────────────
|
||||
|
||||
def _shorten(text: str, n: int) -> str:
|
||||
text = re.sub(r"\s+", " ", (text or "")).strip()
|
||||
return text[:n]
|
||||
|
||||
def _clean_title(title: str) -> str:
|
||||
"""去掉常见堆砌词，让标题更可读。"""
|
||||
t = _shorten(title, 60)
|
||||
return re.sub(r"[【【】】\\[\\]|/]", " ", t).strip()
|
||||
|
||||
def build_context(raw: dict, fallback_name: str = "", fallback_desc: str = "") -> dict:
|
||||
"""从采集数据提炼生图上下文：标题、描述行、卖点列表、参数行。
|
||||
|
||||
raw: {title, desc, price, params: [{key, value}], sellingPoints}
|
||||
"""
|
||||
title = _clean_title(raw.get("title") or fallback_name or "product")
|
||||
desc = _shorten(raw.get("desc") or fallback_desc or "", 200)
|
||||
|
||||
# 卖点：优先显式卖点文本；否则从参数表里挑短而有信息量的键值对
|
||||
selling_points: list[dict] = []
|
||||
sp_text = raw.get("sellingPoints") or ""
|
||||
if sp_text:
|
||||
for chunk in re.split(r"[；;\n·]+|(?<!\d)\.(?!\d)", sp_text):
|
||||
c = _shorten(chunk, 20)
|
||||
if c and len(selling_points) < 5:
|
||||
selling_points.append({"zh": c, "en": c})
|
||||
if not selling_points:
|
||||
for p in (raw.get("params") or [])[:12]:
|
||||
k, v = _shorten(p.get("key", ""), 10), _shorten(str(p.get("value", "")), 16)
|
||||
if k and v and k.lower() not in {"货号", "sku", "isbn", "上架时间"}:
|
||||
selling_points.append({"zh": f"{k} {v}", "en": f"{k} {v}"})
|
||||
if len(selling_points) >= 5:
|
||||
break
|
||||
|
||||
params_line = "; ".join(
|
||||
f"{p.get('key')}: {p.get('value')}" for p in (raw.get("params") or [])[:8]
|
||||
)
|
||||
return {
|
||||
"title": title,
|
||||
"title_en": title, # 采集源多为中文标题，英文场景直接用原词避免乱翻译
|
||||
"desc": desc,
|
||||
"selling_points": selling_points[:3],
|
||||
"params_line": params_line,
|
||||
"price": raw.get("price") or "",
|
||||
}
|
||||
|
||||
def selling_point_lines(ctx: dict, lang: str, max_n: int = 3) -> str:
|
||||
"""卖点列表 → 单行文案（图内 callout/标题用），无卖点返回空串。"""
|
||||
sps = ctx["selling_points"][:max_n]
|
||||
if not sps:
|
||||
return ""
|
||||
key = "zh" if lang == "zh" else "en"
|
||||
return "; ".join(s[key] for s in sps if s.get(key))
|
||||
|
||||
|
||||
def type_name(type_id: str) -> str:
|
||||
return TYPE_NAMES_ZH.get(type_id, type_id)
|
||||
@@ -0,0 +1,9 @@
|
||||
"""豆包（火山方舟 Seedream）提示词。
|
||||
|
||||
豆包与通义同为国产"主体参考"生图模型：参考图即商品锚、prompt 为场景描述，
|
||||
提示词语义一致，直接复用阿里系装配；差异（去 AI 味后缀）在 generator 层追加。
|
||||
独立成文件便于后续按豆包特性分化。
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
from .alibaba import build_prompt as build_prompt # noqa: F401 主体参考语义与通义共用
|
||||
@@ -0,0 +1,163 @@
|
||||
"""Google 图像模型（nano-banana / nano-banana-2 / nano-banana-2-lite / nano-banana-pro）提示词。
|
||||
|
||||
语义：Gemini 图像编辑 —— 原生主体保持能力强，输入图即"主体 + 底图"，
|
||||
对自然语言指令遵循好。不套用 GPT 的编辑契约（冗长的拒绝条款反而稀释指令），
|
||||
也不用负面清单（无 negative_prompt 参数）。要点：
|
||||
- 开头一句话钉死"主体 = 第一张图里的商品，逐像素保持"；
|
||||
- 指令自然语言描述目标画面（场景/排版/文案），不重述商品外观；
|
||||
- 标题/参数仅作识别背景并声明以图为准。
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
from .common import TEXT_RENDER, requirements_block, resolve_style, selling_point_lines
|
||||
|
||||
_SUBJECT_LOCK = (
|
||||
"SUBJECT LOCK (highest priority): the product in the first image is the subject. "
|
||||
"Keep it exactly as photographed — same shape, proportions, colors, print/pattern, "
|
||||
"logo, label and every detail — and place that very product into the result. "
|
||||
"A second image, when present, is another view of the same product for reference only."
|
||||
)
|
||||
|
||||
_QUALITY = (
|
||||
"OUTPUT: photorealistic commercial e-commerce photography, ultra-detailed, "
|
||||
"natural light and shadow, professional retouching."
|
||||
)
|
||||
|
||||
_REMINDER = (
|
||||
"Reminder: keep the product exactly as in the first image; change only its surroundings, "
|
||||
"composition, lighting and overlay graphics."
|
||||
)
|
||||
|
||||
|
||||
def _anchor(ctx: dict) -> str:
|
||||
"""商品文字锚定：仅供识别，明确以图为准（同 gpt 模块，避免文字反噬商品）。"""
|
||||
line = f"Context (identification only): the product is \"{ctx['title']}\""
|
||||
if ctx.get("params_line"):
|
||||
line += f" ({ctx['params_line']})"
|
||||
return line + ". The image, not this text, defines the product's appearance."
|
||||
|
||||
|
||||
# ── 各图类型指令（自然语言编辑口吻）────────────────────────────────────────
|
||||
|
||||
def _task_white_bg(ctx: dict, lang: str) -> str:
|
||||
return (
|
||||
"Replace the background of this product photo with seamless pure white (RGB 255,255,255): "
|
||||
"product centered in front view filling about 85% of the frame, even studio lighting with only "
|
||||
"a faint natural contact shadow. No props, no added text, no background elements."
|
||||
)
|
||||
|
||||
def _task_key_features(ctx: dict, lang: str) -> str:
|
||||
sp = selling_point_lines(ctx, lang) or ctx["title"]
|
||||
return (
|
||||
"Create a square key-features infographic: the unchanged product on the left two-thirds, "
|
||||
"a clean right-hand panel with 3 feature callouts using minimal line icons and thin leader "
|
||||
f"lines pointing at the product. Callout copy: {sp}."
|
||||
)
|
||||
|
||||
def _task_selling_pt(ctx: dict, lang: str) -> str:
|
||||
sp = selling_point_lines(ctx, lang, 1) or ctx["title"]
|
||||
return (
|
||||
"Turn the photo into a single-selling-point poster: hero close-up of the unchanged product at "
|
||||
f"a dynamic angle, one large bold headline about \"{sp}\", generous negative space, and a small "
|
||||
"magnified circle zooming into an existing detail of the product."
|
||||
)
|
||||
|
||||
def _task_material(ctx: dict, lang: str) -> str:
|
||||
return (
|
||||
"Create an extreme macro close-up of an existing area of the product's surface, showing its "
|
||||
"true fabric weave / texture / stitching exactly as in the photo; shallow depth of field, "
|
||||
"raking light, small caption in a corner."
|
||||
)
|
||||
|
||||
def _task_lifestyle(ctx: dict, lang: str) -> str:
|
||||
return (
|
||||
"Place the unchanged product into a realistic everyday scene where it would naturally be used: "
|
||||
"human-scale surroundings, soft daylight, authentic candid mood, the product as the clear visual focus."
|
||||
)
|
||||
|
||||
def _task_multi_scene(ctx: dict, lang: str) -> str:
|
||||
sp = selling_point_lines(ctx, lang)
|
||||
task = (
|
||||
"Build a triptych of three vertical panels separated by thin gutters: each panel shows an "
|
||||
"identical copy of the product in a different usage scene (home interior / outdoor street / "
|
||||
"office desk), with consistent color grading across panels."
|
||||
)
|
||||
if sp:
|
||||
task += f" Panel captions: {sp}."
|
||||
return task
|
||||
|
||||
def _task_ecommerce_detail(ctx: dict, lang: str) -> str:
|
||||
sp = selling_point_lines(ctx, lang) or ctx["title"]
|
||||
params = ctx["params_line"]
|
||||
return (
|
||||
"Compose a square detail-page hero section: top half a hero banner with the unchanged product "
|
||||
"at a 3/4 angle; bottom half a clean spec card with 3 feature rows and line icons"
|
||||
+ (f" (specs: {params})" if params else "")
|
||||
+ f", one highlighted row: {sp}."
|
||||
)
|
||||
|
||||
def _task_size_chart(ctx: dict, lang: str) -> str:
|
||||
dims = ctx["params_line"]
|
||||
return (
|
||||
"Create a size chart: the unchanged product in clean front and side views on a light background, "
|
||||
"thin measurement annotation lines (arrows) marking length, width and height with values beside "
|
||||
"each line"
|
||||
+ (f" (known specs: {dims})" if dims else "")
|
||||
+ ", small caption row, precise technical-drawing aesthetic."
|
||||
)
|
||||
|
||||
def _task_sku_collection(ctx: dict, lang: str) -> str:
|
||||
# 不展开"全部配色"：会凭空造出新商品；只排列同一件的多个副本
|
||||
return (
|
||||
"Arrange several identical copies of the product in a neat equal grid (2-4 per row) with a small "
|
||||
"label chip below each copy; identical lighting and scale across copies. Every copy shows this "
|
||||
"exact product — do not invent other colorways or variants."
|
||||
)
|
||||
|
||||
_TASK_BUILDERS = {
|
||||
"white_bg": (_task_white_bg, False),
|
||||
"key_features": (_task_key_features, True),
|
||||
"selling_pt": (_task_selling_pt, True),
|
||||
"material": (_task_material, True),
|
||||
"lifestyle": (_task_lifestyle, True),
|
||||
"multi_scene": (_task_multi_scene, True),
|
||||
"ecommerce_detail": (_task_ecommerce_detail, True),
|
||||
"size_chart": (_task_size_chart, True),
|
||||
"sku_collection": (_task_sku_collection, True),
|
||||
}
|
||||
|
||||
|
||||
def build_prompt(type_id: str, ctx: dict, style_set: int, lang: str, extra: dict | None = None,
|
||||
style_prompt: str | None = None, requirements: str | None = None) -> str:
|
||||
"""构造指定图类型的 prompt：要求块 → 指令 → 主体锁 → 锚定 → 风格 → 文案 → 画质 → 提醒。"""
|
||||
style = resolve_style(style_set, style_prompt)
|
||||
extra = extra or {}
|
||||
hint = (extra.get("prompt_hint") or "").strip()
|
||||
|
||||
if type_id == "custom":
|
||||
purpose = extra.get("title") or ""
|
||||
detail = extra.get("detail") or ""
|
||||
task = "Create an e-commerce marketing image featuring the product from the first image"
|
||||
task += f" — {purpose}" if purpose else ""
|
||||
task += f": {detail}" if detail else ""
|
||||
task += "."
|
||||
wants_text = True
|
||||
else:
|
||||
entry = _TASK_BUILDERS.get(type_id)
|
||||
if entry is None:
|
||||
raise ValueError(f"未知图类型: {type_id}")
|
||||
builder, wants_text = entry
|
||||
task = builder(ctx, lang)
|
||||
if hint:
|
||||
task += f" Composition guidance: {hint}."
|
||||
|
||||
parts = [p for p in (requirements_block(requirements),) if p]
|
||||
parts.append(task)
|
||||
parts.append(_SUBJECT_LOCK)
|
||||
parts.append(_anchor(ctx))
|
||||
parts.append(f"Scene style (scene and background only, never the product): {style['tone']}.")
|
||||
if wants_text:
|
||||
parts.append(f"Text overlay (a graphic layer, never printed on the product): {TEXT_RENDER[lang]}")
|
||||
parts.append(_QUALITY)
|
||||
parts.append(_REMINDER)
|
||||
return "\n\n".join(parts)
|
||||
@@ -0,0 +1,195 @@
|
||||
"""GPT 图像模型（gpt-image-2 / gpt-image-2-vip，RightAPI 中转）提示词。
|
||||
|
||||
语义：/v1/images/edits —— 输入图是"被编辑的照片"，prompt 是编辑指令；
|
||||
与通义/豆包的"主体参考"语义完全不同：参考图不是商品锚，模型会按文字指令
|
||||
重新渲染整张图。此前与国产模型共用场景提示词，再用文字锚定商品并要求输出
|
||||
"匹配商品描述"，导致模型把商品改造成营销关键词描述的样子（必现商品被改）。
|
||||
|
||||
本模块写法原则：
|
||||
1. 商品只由 Image 1 定义；标题/参数仅作识别背景并声明"以图为准"，
|
||||
绝不要求输出匹配文字描述（那等于授权模型改商品）；
|
||||
2. 指令只说"改什么"（背景/场景/排版/文案），不描述商品外观；
|
||||
3. 分节精简、首尾重申保真；不用负面清单（gpt 无 negative_prompt 参数，
|
||||
罗列畸形反而往上下文植入概念）；
|
||||
4. sku 合集 / 多拼图明确"复制同一件商品，禁止发明新配色或变体"。
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
from .common import TEXT_RENDER, requirements_block, resolve_style, selling_point_lines
|
||||
|
||||
# 保真锁：商品由 Image 1 唯一定义，其余指令一律不得触碰商品本体
|
||||
_PRESERVE = (
|
||||
"PRESERVE (absolute, overrides every other instruction below): the product shown in Image 1. "
|
||||
"Reuse the photographed product exactly as it is — identical shape, silhouette, proportions, "
|
||||
"colors, print/pattern, logo and label text, materials, stitching and surface details. "
|
||||
"Do not redesign, restyle, recolor, re-pattern, tidy up or substitute the product, "
|
||||
"and do not let any style or text instruction below alter it. Image 2 is a secondary "
|
||||
"view of the same product for reference only."
|
||||
)
|
||||
|
||||
_STYLE = (
|
||||
"SCENE STYLE (applies to background, scene, props and lighting only — never to the product): "
|
||||
)
|
||||
|
||||
_QUALITY = (
|
||||
"OUTPUT: photorealistic commercial e-commerce photography, ultra-detailed, "
|
||||
"natural light and shadow, professional retouching."
|
||||
)
|
||||
|
||||
_REMINDER = (
|
||||
"FINAL CHECK: the product itself must remain exactly as photographed in Image 1 — "
|
||||
"only its surroundings, composition, lighting and overlay graphics may differ."
|
||||
)
|
||||
|
||||
# 图内文案：明确是"排版图层"，不落在商品本体上
|
||||
_TEXT_SCOPE = (
|
||||
"TEXT OVERLAY (a graphic layer on the composition, never printed on the product): "
|
||||
)
|
||||
|
||||
|
||||
def _anchor(ctx: dict) -> str:
|
||||
"""商品文字锚定：仅供识别，明确声明以图为准。
|
||||
|
||||
只放标题 + 参数、不放营销描述——描述里的卖点词（"卡通""加固""防水"等）
|
||||
在 edits 语义下会被执行到商品上；官逆通道（-vip）参考图被弱化时，
|
||||
文字锚定用于帮模型认出"是哪件商品"，而不是"长什么样"。
|
||||
"""
|
||||
line = f"CONTEXT (identification only): the product is \"{ctx['title']}\""
|
||||
if ctx.get("params_line"):
|
||||
line += f" ({ctx['params_line']})"
|
||||
return (
|
||||
line
|
||||
+ ". Image 1 — not this text — defines the product's appearance; "
|
||||
"if they ever conflict, follow Image 1."
|
||||
)
|
||||
|
||||
|
||||
# ── 各图类型的编辑指令（只描述改动，不描述商品）────────────────────────────
|
||||
|
||||
def _task_white_bg(ctx: dict, lang: str) -> str:
|
||||
return (
|
||||
"TASK: Clean up this product photo for a marketplace listing. Replace the entire "
|
||||
"background with seamless pure white (RGB 255,255,255); recompose with the product "
|
||||
"centered in front view filling about 85% of the frame; keep only a faint natural "
|
||||
"contact shadow. No props, no text, no background elements."
|
||||
)
|
||||
|
||||
def _task_key_features(ctx: dict, lang: str) -> str:
|
||||
sp = selling_point_lines(ctx, lang) or ctx["title"]
|
||||
return (
|
||||
"TASK: Feature infographic on a square canvas. Keep the product unchanged on the left "
|
||||
"two-thirds; build the right third as a clean info panel listing 3 feature callouts with "
|
||||
f"minimal line icons and thin leader lines pointing at parts of the product. Callout copy: {sp}."
|
||||
)
|
||||
|
||||
def _task_selling_pt(ctx: dict, lang: str) -> str:
|
||||
sp = selling_point_lines(ctx, lang, 1) or ctx["title"]
|
||||
return (
|
||||
"TASK: Single-selling-point poster. Hero close-up of the unchanged product at a dynamic "
|
||||
f"angle, generous negative space, one large bold headline about \"{sp}\", plus one small "
|
||||
"magnified circle zooming into an existing detail of the product (zoom only — do not "
|
||||
"invent details that are not in the photo)."
|
||||
)
|
||||
|
||||
def _task_material(ctx: dict, lang: str) -> str:
|
||||
return (
|
||||
"TASK: Material close-up. Zoom tightly into an existing area of the product's surface and "
|
||||
"show its true texture — fabric weave, surface finish, stitching — exactly as it appears in "
|
||||
"Image 1; shallow depth of field, raking light across the surface; small caption label in a corner."
|
||||
)
|
||||
|
||||
def _task_lifestyle(ctx: dict, lang: str) -> str:
|
||||
return (
|
||||
"TASK: Lifestyle scene. Place the unchanged product into a realistic everyday environment "
|
||||
"where it would naturally be used: human-scale surroundings, soft daylight, authentic candid "
|
||||
"mood, matched shadows and color temperature, the product remaining the clear visual focus."
|
||||
)
|
||||
|
||||
def _task_multi_scene(ctx: dict, lang: str) -> str:
|
||||
sp = selling_point_lines(ctx, lang)
|
||||
task = (
|
||||
"TASK: Triptych showcase. Build three vertical panels separated by thin gutters; every panel "
|
||||
"contains an IDENTICAL copy of the product from Image 1 (do not re-render it differently per "
|
||||
"panel) placed in a different usage scene (e.g. home interior / outdoor street / office desk), "
|
||||
"with consistent color grading across panels."
|
||||
)
|
||||
if sp:
|
||||
task += f" Panel captions: {sp}."
|
||||
return task
|
||||
|
||||
def _task_ecommerce_detail(ctx: dict, lang: str) -> str:
|
||||
sp = selling_point_lines(ctx, lang) or ctx["title"]
|
||||
params = ctx["params_line"]
|
||||
return (
|
||||
"TASK: Detail-page hero section on a square canvas. Top half: hero banner with the unchanged "
|
||||
"product at a 3/4 angle. Bottom half: clean spec card with 3 feature rows and line icons"
|
||||
+ (f" (specs: {params})" if params else "")
|
||||
+ f", one highlighted row: {sp}."
|
||||
)
|
||||
|
||||
def _task_size_chart(ctx: dict, lang: str) -> str:
|
||||
dims = ctx["params_line"]
|
||||
return (
|
||||
"TASK: Measurement chart. Show the unchanged product in clean front and side views on a light "
|
||||
"background; overlay thin technical annotation lines (arrows) marking length, width and height "
|
||||
"with measurement values rendered beside each line"
|
||||
+ (f" (known specs: {dims})" if dims else "")
|
||||
+ "; precise technical-drawing aesthetic, small caption row."
|
||||
)
|
||||
|
||||
def _task_sku_collection(ctx: dict, lang: str) -> str:
|
||||
# 关键差异：不允许像国产模型那样展开"全部配色"——edits 语义下那会凭空造出新商品
|
||||
return (
|
||||
"TASK: Product lineup. Arrange several IDENTICAL copies of the product from Image 1 in a neat "
|
||||
"equal grid (2-4 per row) with a small label chip below each copy; identical lighting and scale "
|
||||
"across copies. Every copy must show this exact product — do NOT invent other colorways, "
|
||||
"variants or versions."
|
||||
)
|
||||
|
||||
_TASK_BUILDERS = {
|
||||
"white_bg": (_task_white_bg, False),
|
||||
"key_features": (_task_key_features, True),
|
||||
"selling_pt": (_task_selling_pt, True),
|
||||
"material": (_task_material, True),
|
||||
"lifestyle": (_task_lifestyle, True),
|
||||
"multi_scene": (_task_multi_scene, True),
|
||||
"ecommerce_detail": (_task_ecommerce_detail, True),
|
||||
"size_chart": (_task_size_chart, True),
|
||||
"sku_collection": (_task_sku_collection, True),
|
||||
}
|
||||
|
||||
|
||||
def build_prompt(type_id: str, ctx: dict, style_set: int, lang: str, extra: dict | None = None,
|
||||
style_prompt: str | None = None, requirements: str | None = None) -> str:
|
||||
"""构造指定图类型的 edits 语义 prompt：要求块 → 编辑指令 → 保真锁 → 锚定 → 风格 → 文案 → 画质 → 终检。"""
|
||||
style = resolve_style(style_set, style_prompt)
|
||||
extra = extra or {}
|
||||
hint = (extra.get("prompt_hint") or "").strip()
|
||||
|
||||
if type_id == "custom":
|
||||
purpose = extra.get("title") or ""
|
||||
detail = extra.get("detail") or ""
|
||||
task = "TASK: Create an e-commerce marketing image featuring the product from Image 1"
|
||||
task += f" — {purpose}" if purpose else ""
|
||||
task += f": {detail}" if detail else ""
|
||||
task += "."
|
||||
wants_text = True
|
||||
else:
|
||||
entry = _TASK_BUILDERS.get(type_id)
|
||||
if entry is None:
|
||||
raise ValueError(f"未知图类型: {type_id}")
|
||||
builder, wants_text = entry
|
||||
task = builder(ctx, lang)
|
||||
if hint:
|
||||
task += f" Composition guidance: {hint}."
|
||||
|
||||
parts = [p for p in (requirements_block(requirements),) if p]
|
||||
parts.append(task)
|
||||
parts.append(_PRESERVE)
|
||||
parts.append(_anchor(ctx))
|
||||
parts.append(f"{_STYLE}{style['tone']}.")
|
||||
if wants_text:
|
||||
parts.append(f"{_TEXT_SCOPE}{TEXT_RENDER[lang]}")
|
||||
parts.append(_QUALITY)
|
||||
parts.append(_REMINDER)
|
||||
return "\n\n".join(parts)
|
||||
@@ -16,6 +16,7 @@ TASK_PARTIAL = "partial"
|
||||
TASK_FAILED = "failed"
|
||||
|
||||
# 任务内单张图状态
|
||||
IMG_PENDING = "pending" # 生成中（前端据此隐藏占位格，只渲染 ok/failed 终态）
|
||||
IMG_OK = "ok"
|
||||
IMG_FAILED = "failed"
|
||||
|
||||
@@ -26,7 +27,7 @@ class TaskImage:
|
||||
|
||||
type_id: str
|
||||
name: str
|
||||
status: str = IMG_FAILED # 循环里先建后跑，成功后改为 ok
|
||||
status: str = IMG_PENDING # 循环里先建后跑，成功改 ok、失败显式改 failed
|
||||
url: str = ""
|
||||
error: str | None = None
|
||||
|
||||
|
||||
Reference in New Issue
Block a user