feat: 插件改为浮窗模式
This commit is contained in:
R524809
@@ -10,6 +10,7 @@ import asyncio
|
||||
import base64
|
||||
import logging
|
||||
import mimetypes
|
||||
import re
|
||||
from uuid import UUID
|
||||
|
||||
import httpx
|
||||
@@ -19,21 +20,38 @@ from config import get_settings
|
||||
from db import get_session_factory
|
||||
from models import Product, ProductAsset, Suite, SuiteImage, SUITE_RUNNING, SUITE_DONE, SUITE_PARTIAL, SUITE_FAILED, STATUS_OK, STATUS_FAILED
|
||||
from services import storage
|
||||
from services.prompt import build_prompt, build_context, type_name
|
||||
from services.prompt import build_prompt, build_context, type_name, wrap_prompt_for_gpt_edits
|
||||
|
||||
log = logging.getLogger("suite.generator")
|
||||
|
||||
|
||||
class ApiError(RuntimeError):
|
||||
"""带 HTTP 状态码的 API 错误（用于区分可重试的网关/限流错误）。"""
|
||||
|
||||
def __init__(self, message: str, status: int = 0):
|
||||
super().__init__(message)
|
||||
self.status = status
|
||||
|
||||
|
||||
_HTML_TITLE_RE = re.compile(r"<title[^>]*>(.*?)</title>", re.IGNORECASE | re.DOTALL)
|
||||
|
||||
|
||||
def _raise_api_error(resp, provider: str):
|
||||
"""HTTP 错误时抛出带 API 错误码/信息的异常（响应体里有真正的失败原因）。"""
|
||||
if resp.is_success:
|
||||
return
|
||||
text = resp.text or ""
|
||||
if "<html" in text[:300].lower() or text.lstrip()[:15].lower().startswith("<!doctype"):
|
||||
# Cloudflare/网关错误页：取 <title> 作摘要，避免整段 HTML 进错误信息
|
||||
m = _HTML_TITLE_RE.search(text)
|
||||
detail = (re.sub(r"\s+", " ", m.group(1)).strip() if m else "") or "网关返回 HTML 错误页（上游/CDN 故障）"
|
||||
raise ApiError(f"{provider} API HTTP {resp.status_code} — {detail}", resp.status_code)
|
||||
try:
|
||||
body = resp.json()
|
||||
detail = f"{body.get('code', '')}: {body.get('message', '')}".strip(': ')
|
||||
except Exception: # noqa: BLE001
|
||||
detail = resp.text[:200]
|
||||
raise RuntimeError(f"{provider} API HTTP {resp.status_code} — {detail or '无错误详情'}")
|
||||
detail = text[:200]
|
||||
raise ApiError(f"{provider} API HTTP {resp.status_code} — {detail or '无错误详情'}", resp.status_code)
|
||||
|
||||
# 参考图选择：material 用第 2 张（背面/细节），其余用第 1 张（正面）
|
||||
TYPE_REF_INDEX = {
|
||||
@@ -42,7 +60,7 @@ TYPE_REF_INDEX = {
|
||||
DEFAULT_REF_COUNT = 2 # 每次生图最多带的参考图数（正面 1 张 + 背面/细节 1 张）
|
||||
|
||||
|
||||
def _image_size(provider: str, ratio: str, is_wan: bool = True) -> str:
|
||||
def _image_size(provider: str, ratio: str, is_wan: bool = True, model: str = "") -> 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"
|
||||
@@ -50,6 +68,9 @@ def _image_size(provider: str, ratio: str, is_wan: bool = True) -> str:
|
||||
# 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
|
||||
# wan2.6 系列总像素限制在 [1280², 1440²]，wan2.7 的 1536*2048/2048*2048 会超限
|
||||
if model.startswith("wan2.6"):
|
||||
return "1152*1536" if ratio == "3:4" else "1440*1440"
|
||||
if ratio == "3:4":
|
||||
return "1536*2048" if is_wan else "768*1024"
|
||||
return "2048*2048" if is_wan else "1024*1024"
|
||||
@@ -136,6 +157,11 @@ def _is_wan_model(model: str) -> bool:
|
||||
return model.lower().startswith("wan")
|
||||
|
||||
|
||||
def _is_t2i_model(model: str) -> bool:
|
||||
"""纯文生图模型（如 wan2.6-t2i）：不接受参考图，商品一致性只能靠文案描述。"""
|
||||
return "t2i" in model.lower()
|
||||
|
||||
|
||||
async def _tongyi_poll_task(client: httpx.AsyncClient, key: str, task_id: str, max_wait: int) -> str:
|
||||
poll_url = "https://dashscope.aliyuncs.com/api/v1/tasks/" + task_id
|
||||
elapsed, interval = 0, 3
|
||||
@@ -174,7 +200,10 @@ async def generate_tongyi(prompt: str, ref_images: list[str], size: str = "2048*
|
||||
else "https://dashscope.aliyuncs.com/api/v1/services/aigc/multimodal-generation/generation"
|
||||
)
|
||||
|
||||
content: list[dict] = [{"image": await _resolve_ref(u)} for u in ref_images]
|
||||
# t2i 模型不接受参考图：content 只有文本，商品一致性依赖 prompt 里的标题/卖点描述
|
||||
content: list[dict] = []
|
||||
if not _is_t2i_model(model):
|
||||
content = [{"image": await _resolve_ref(u)} for u in ref_images]
|
||||
content.append({"text": prompt})
|
||||
|
||||
params = {"size": size, "n": 1, "watermark": False}
|
||||
@@ -210,18 +239,25 @@ async def generate_tongyi(prompt: str, ref_images: list[str], size: str = "2048*
|
||||
|
||||
# ── 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:
|
||||
# 可重试的状态码：中转限流/网关抖动（该中转限流时返回 Cloudflare 502 而非 429）
|
||||
RETRYABLE_STATUS = {429, 500, 502, 503, 504}
|
||||
|
||||
# 中转对 input_fidelity 参数的支持探测：None=未探测，True=支持，False=不支持（已降级）
|
||||
_rightapi_fidelity_supported: bool | None = None
|
||||
|
||||
|
||||
async def _rightapi_request(s, prompt: str, ref_images: list[str], size: str, model: str) -> 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。
|
||||
input_fidelity=high 强制高保真保留输入图细节（商品一致性关键参数，仅 edits 端点）；
|
||||
中转若不认该参数（400），自动去掉重试并记住，后续请求不再带。
|
||||
"""
|
||||
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
|
||||
global _rightapi_fidelity_supported
|
||||
base = s.rightapi_base_url.rstrip("/")
|
||||
headers = {"Authorization": f"Bearer {s.rightapi_api_key}"}
|
||||
use_fidelity = bool(ref_images) and s.rightapi_input_fidelity and _rightapi_fidelity_supported is not False
|
||||
|
||||
async with httpx.AsyncClient(timeout=max(s.request_timeout, 600), verify=False) as client:
|
||||
common = {
|
||||
@@ -232,12 +268,22 @@ async def generate_rightapi(prompt: str, ref_images: list[str], size: str = "204
|
||||
"output_format": "jpeg", # 与落盘 .jpg 后缀一致
|
||||
"n": 1,
|
||||
}
|
||||
if use_fidelity:
|
||||
common["input_fidelity"] = s.rightapi_input_fidelity
|
||||
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)
|
||||
# 中转不认 input_fidelity：去掉参数重试一次（仅一次探测）
|
||||
if resp.status_code == 400 and use_fidelity and "input_fidelity" in resp.text:
|
||||
_rightapi_fidelity_supported = False
|
||||
log.warning("RightAPI 不支持 input_fidelity 参数，已自动去掉并降级（后续请求不再带）")
|
||||
common.pop("input_fidelity", None)
|
||||
resp = await client.post(f"{base}/v1/images/edits", headers=headers, files=files, data=common)
|
||||
elif resp.is_success and use_fidelity:
|
||||
_rightapi_fidelity_supported = True
|
||||
else:
|
||||
resp = await client.post(
|
||||
f"{base}/v1/images/generations",
|
||||
@@ -257,6 +303,36 @@ async def generate_rightapi(prompt: str, ref_images: list[str], size: str = "204
|
||||
return dl.content
|
||||
|
||||
|
||||
async def generate_rightapi(prompt: str, ref_images: list[str], size: str = "2048x2048", model: str | None = None) -> bytes:
|
||||
"""带重试的 RightAPI 入口：429/5xx/超时按递增间隔重试。
|
||||
|
||||
实测该中转对同 key 连续请求有分钟级冷却（成功一张后紧接着的请求会被网关秒拒 502），
|
||||
60s → 120s → 240s 的退避基本能等到窗口放开。
|
||||
"""
|
||||
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
|
||||
attempts = max(1, s.rightapi_max_retries)
|
||||
|
||||
last_exc: Exception | None = None
|
||||
for i in range(attempts):
|
||||
try:
|
||||
return await _rightapi_request(s, prompt, ref_images, size, model)
|
||||
except ApiError as exc:
|
||||
last_exc = exc
|
||||
if exc.status not in RETRYABLE_STATUS:
|
||||
raise # 参数错误等不可重试，立即失败
|
||||
except (httpx.TimeoutException, httpx.TransportError) as exc:
|
||||
last_exc = exc # 网络抖动/超时可重试
|
||||
if i == attempts - 1:
|
||||
break
|
||||
wait = s.rightapi_retry_wait * (2 ** i)
|
||||
log.warning("RightAPI 第 %d/%d 次请求失败（%s），%ds 后重试", i + 1, attempts, last_exc, wait)
|
||||
await asyncio.sleep(wait)
|
||||
raise last_exc # type: ignore[misc]
|
||||
|
||||
|
||||
GENERATORS = {"doubao": generate_doubao, "tongyi": generate_tongyi, "rightapi": generate_rightapi}
|
||||
|
||||
|
||||
@@ -344,7 +420,7 @@ async def run_suite(suite_id: str) -> None:
|
||||
"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)
|
||||
size = _image_size(provider_name, suite.ratio, is_wan=is_wan, model=model)
|
||||
|
||||
# 任务列表：方案（逐张）优先，旧路径按 types
|
||||
if suite.plan:
|
||||
@@ -372,6 +448,9 @@ async def run_suite(suite_id: str) -> None:
|
||||
type_id, ctx, suite.style_set, suite.lang,
|
||||
extra=job, style_prompt=suite.style_prompt, requirements=suite.requirements,
|
||||
)
|
||||
# gpt-image edits 语义：商品冻结契约前置，防止风格词改商品
|
||||
if provider_name == "rightapi":
|
||||
prompt = wrap_prompt_for_gpt_edits(prompt)
|
||||
if product:
|
||||
refs = await _select_ref_images(db, product.id, type_id)
|
||||
else:
|
||||
|
||||
@@ -116,6 +116,39 @@ DEFAULT_NEGATIVE_INTENT = (
|
||||
"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。
|
||||
|
||||
GPT_EDITS_CONTRACT = (
|
||||
"INPUT IMAGES: Image 1 (and Image 2 if present) are reference photos of ONE product "
|
||||
"from different angles. Use them ONLY as the source of the product's true appearance.\n"
|
||||
"PRESERVE (frozen, never change): the product itself — silhouette, proportions, colors, "
|
||||
"print/pattern (keep stripes / logos / labels exactly), materials, texture, stitching, "
|
||||
"hardware and every design detail. The product in the output must be the same physical "
|
||||
"item as in the input images, merely photographed in a new setting.\n"
|
||||
"MAY CHANGE: background, scene, props, camera angle, lighting, composition "
|
||||
"and in-image marketing typography.\n"
|
||||
"STYLE SCOPE: all style, mood, color-palette and decoration instructions below describe "
|
||||
"the SCENE AND BACKGROUND ONLY — never apply them to the product itself. "
|
||||
"Do not restyle, recolor, re-pattern or redecorate the product. "
|
||||
"You may relight the product so it sits naturally in the new scene "
|
||||
"(matched shadows and color temperature), but never change its design, colors or pattern."
|
||||
)
|
||||
|
||||
GPT_EDITS_FINAL_CHECK = (
|
||||
"FINAL CHECK before output: if the product in your result differs from the input product 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) -> 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}\n\n{prompt}\n\n{GPT_EDITS_FINAL_CHECK}"
|
||||
|
||||
# ── 商品上下文提炼 ────────────────────────────────────────────────────────
|
||||
|
||||
def _shorten(text: str, n: int) -> str:
|
||||
|
||||
Reference in New Issue
Block a user