Joey
430 lines
20 KiB
Python
430 lines
20 KiB
Python
"""套图生成服务：图像 provider（豆包 Seedream / 通义万相）+ 任务执行器。
|
||
|
||
Provider 调用方式移植自 ecommerce-image-suite/scripts/generate.py：
|
||
- doubao：火山方舟 images/generations，同步返回 URL；参考图走 image 字段（data URI）
|
||
- tongyi：wan* 万象模型走异步任务轮询；qwen* 走同步 multimodal-generation
|
||
"""
|
||
from __future__ import annotations
|
||
|
||
import asyncio
|
||
import base64
|
||
import logging
|
||
import mimetypes
|
||
import re
|
||
|
||
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
|
||
|
||
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 = 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 = {
|
||
"material": 1,
|
||
}
|
||
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, 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"
|
||
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
|
||
# 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"
|
||
|
||
_DOUBAO_ANTI_AI = (
|
||
"authentic real-world photography, natural imperfections, genuine texture, "
|
||
"no synthetic look, no CGI quality, no heavy post-processing"
|
||
)
|
||
|
||
DEFAULT_NEGATIVE_PROMPT = (
|
||
"AI-generated look, artificial, CGI quality, 3D render, synthetic texture, "
|
||
"plastic skin, mannequin-like, too perfect, oversaturated, HDR, heavy vignette, "
|
||
"low resolution, blurry, deformed, bad anatomy, overexposed, underexposed, grainy, "
|
||
"watermark, text distortion, bad typography, overlapping text, cheap look, cartoon"
|
||
)
|
||
|
||
|
||
# ── 参考图解析 ────────────────────────────────────────────────────────────
|
||
|
||
def _bytes_to_data_uri(data: bytes, mime: str) -> str:
|
||
return f"data:{mime};base64,{base64.b64encode(data).decode()}"
|
||
|
||
|
||
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原始字节再进请求体（data URI 或 multipart）。
|
||
"""
|
||
if url.startswith("data:"):
|
||
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 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 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:
|
||
s = get_settings()
|
||
if not s.ark_api_key:
|
||
raise RuntimeError("未配置 ARK_API_KEY（.env）")
|
||
body = {
|
||
"model": model or s.ark_image_model,
|
||
"prompt": prompt.rstrip(". ") + ". " + _DOUBAO_ANTI_AI,
|
||
"size": size,
|
||
"response_format": "url",
|
||
"watermark": False,
|
||
"n": 1,
|
||
}
|
||
if ref_images:
|
||
body["image"] = [await _resolve_ref(u) for u in ref_images]
|
||
async with httpx.AsyncClient(timeout=s.request_timeout, verify=False) as client:
|
||
resp = await client.post(
|
||
s.ark_base_url,
|
||
headers={"Authorization": f"Bearer {s.ark_api_key}", "Content-Type": "application/json"},
|
||
json=body,
|
||
)
|
||
_raise_api_error(resp, "豆包")
|
||
img_url = resp.json()["data"][0]["url"]
|
||
dl = await client.get(img_url, timeout=s.request_timeout)
|
||
dl.raise_for_status()
|
||
return dl.content
|
||
|
||
|
||
# ── Provider：通义万相 / 千问（DashScope）────────────────────────────────
|
||
|
||
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
|
||
while elapsed < max_wait:
|
||
resp = await client.get(poll_url, headers={"Authorization": f"Bearer {key}"}, timeout=30)
|
||
resp.raise_for_status()
|
||
result = resp.json()
|
||
status = result.get("output", {}).get("task_status", "")
|
||
if status == "SUCCEEDED":
|
||
choices = result["output"].get("choices", [])
|
||
if choices:
|
||
content = choices[0].get("message", {}).get("content", [])
|
||
if content:
|
||
return content[0].get("image", "")
|
||
results = result["output"].get("results", [])
|
||
if results:
|
||
return results[0].get("url") or results[0].get("b64_image", "")
|
||
raise RuntimeError(f"通义任务成功但无结果: {result}")
|
||
if status in ("FAILED", "UNKNOWN"):
|
||
raise RuntimeError(f"通义任务失败: {result}")
|
||
await asyncio.sleep(interval)
|
||
elapsed += interval
|
||
interval = min(interval + 2, 10)
|
||
raise TimeoutError(f"通义异步任务超时 ({max_wait}s): task_id={task_id}")
|
||
|
||
|
||
async def generate_tongyi(prompt: str, ref_images: list[str], size: str = "2048*2048", model: str | None = None) -> bytes:
|
||
s = get_settings()
|
||
if not s.dashscope_api_key:
|
||
raise RuntimeError("未配置 DASHSCOPE_API_KEY（.env）")
|
||
model = model or s.dashscope_model
|
||
is_wan = _is_wan_model(model)
|
||
url = s.dashscope_base_url or (
|
||
"https://dashscope.aliyuncs.com/api/v1/services/aigc/image-generation/generation"
|
||
if is_wan
|
||
else "https://dashscope.aliyuncs.com/api/v1/services/aigc/multimodal-generation/generation"
|
||
)
|
||
|
||
# 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}
|
||
if not is_wan:
|
||
params["prompt_extend"] = False
|
||
params["negative_prompt"] = DEFAULT_NEGATIVE_PROMPT[:500]
|
||
|
||
headers = {"Authorization": f"Bearer {s.dashscope_api_key}", "Content-Type": "application/json"}
|
||
if is_wan:
|
||
headers["X-DashScope-Async"] = "enable"
|
||
|
||
body = {"model": model, "input": {"messages": [{"role": "user", "content": content}]}, "parameters": params}
|
||
|
||
async with httpx.AsyncClient(timeout=s.request_timeout, verify=False) as client:
|
||
resp = await client.post(url, headers=headers, json=body)
|
||
_raise_api_error(resp, "通义")
|
||
data = resp.json()
|
||
if is_wan:
|
||
task_id = data.get("output", {}).get("task_id", "")
|
||
if not task_id:
|
||
raise RuntimeError(f"通义万象未返回 task_id: {data}")
|
||
img_url = await _tongyi_poll_task(client, s.dashscope_api_key, task_id, s.poll_max_wait)
|
||
if img_url.startswith("data:") or len(img_url) > 500:
|
||
return base64.b64decode(img_url.split(",", 1)[-1] if "," in img_url else img_url)
|
||
dl = await client.get(img_url, timeout=s.request_timeout)
|
||
dl.raise_for_status()
|
||
return dl.content
|
||
img_url = data["output"]["choices"][0]["message"]["content"][0]["image"]
|
||
dl = await client.get(img_url, timeout=s.request_timeout)
|
||
dl.raise_for_status()
|
||
return dl.content
|
||
|
||
|
||
# ── Provider：RightAPI（gpt-image，OpenAI 兼容中转）──────────────────────
|
||
|
||
# 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：按模型记忆不支持该参数的模型（gpt-image 系列支持，
|
||
# nano-banana 系列可能不认；降级只影响触发过的模型，不牵连其他模型）
|
||
_rightapi_fidelity_unsupported: set[str] = set()
|
||
|
||
|
||
async def _rightapi_request(s, prompt: str, ref_images: list[str], size: str, model: str) -> bytes:
|
||
"""RightAPI 各模型：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 是 gpt-image-1 的 edits 保真参数（gpt-image-2 官方已移除、默认高保真，
|
||
官逆通道更是不识别）；带上是为了兼容按 gpt-image-1 语义实现的中转，中转不认（400）则按模型
|
||
自动去掉重试并记住，该模型后续请求不再带。
|
||
"""
|
||
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 model not in _rightapi_fidelity_unsupported
|
||
)
|
||
|
||
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 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），降级只记到当前模型
|
||
if resp.status_code == 400 and use_fidelity and "input_fidelity" in resp.text:
|
||
_rightapi_fidelity_unsupported.add(model)
|
||
log.warning("RightAPI 模型 %s 不支持 input_fidelity 参数，已自动去掉并降级（该模型后续请求不再带）", model)
|
||
common.pop("input_fidelity", None)
|
||
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
|
||
|
||
|
||
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}
|
||
|
||
|
||
# ── 任务执行器 ────────────────────────────────────────────────────────────
|
||
|
||
def _order_refs(refs: list[str], type_id: str) -> list[str]:
|
||
"""参考图槽位选择 + 截断：material 偏好第 2 张，其余用第 1 张。"""
|
||
preferred = TYPE_REF_INDEX.get(type_id)
|
||
if preferred is not None and len(refs) > preferred:
|
||
refs = [refs[preferred]] + [r for i, r in enumerate(refs) if i != preferred]
|
||
return refs[:DEFAULT_REF_COUNT]
|
||
|
||
|
||
def _refs_for_job(images: list[dict], job: dict) -> list[str]:
|
||
"""无状态路径：按方案项选参考图。
|
||
|
||
优先 variant_name 精确匹配（「主图·粉色」用粉色那张 SKU 图）；
|
||
匹配不到则回退 main 组第一张（再退到任意第一张）。
|
||
"""
|
||
variant = job.get("variant_name")
|
||
if variant:
|
||
matched = [i["url"] for i in images if i.get("variant_name") == variant]
|
||
if matched:
|
||
return matched[:DEFAULT_REF_COUNT]
|
||
mains = [i["url"] for i in images if i.get("group_key") == "main"]
|
||
others = [i["url"] for i in images if i.get("group_key") != "main"]
|
||
pool = mains or others or [i["url"] for i in images]
|
||
if not pool:
|
||
raise RuntimeError("任务没有参考图")
|
||
return _order_refs(pool, job.get("kind", ""))
|
||
|
||
|
||
# 串行生成队列：所有用户共享同一批 API key，并发生成会触发中转限流
|
||
# （rightapi 同 key 分钟级冷却）；同一时间只跑一个任务，其余保持 pending 排队。
|
||
_GEN_LOCK = asyncio.Lock()
|
||
|
||
|
||
async def run_suite(task: Task) -> None:
|
||
"""后台执行套图任务：排队 → 逐张生成 → 落盘 → 更新内存状态；单张失败不中断。"""
|
||
settings = get_settings()
|
||
provider_name = task.provider or settings.image_provider
|
||
generator = GENERATORS.get(provider_name)
|
||
if generator is None:
|
||
task.status = TASK_FAILED
|
||
task.error = f"未知 provider: {provider_name}"
|
||
return
|
||
|
||
ctx = build_context(task.context or {}, fallback_name="")
|
||
model = task.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, task.ratio, is_wan=is_wan, model=model)
|
||
jobs = [dict(j) for j in task.plan]
|
||
|
||
async with _GEN_LOCK:
|
||
task.status = TASK_RUNNING
|
||
ok, failed = 0, 0
|
||
failures: list[str] = []
|
||
for job in jobs:
|
||
type_id = job["kind"]
|
||
image = TaskImage(type_id=type_id, name=job.get("title") or type_name(type_id))
|
||
task.images.append(image)
|
||
try:
|
||
prompt = build_prompt(
|
||
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
|
||
ext = ".png" if data[:8] == b"\x89PNG\r\n\x1a\n" else ".jpg"
|
||
key = storage.write_bytes(data, key_prefix=f"suites/{task.id}", ext=ext)
|
||
image.url = storage.public_url(key)
|
||
image.status = IMG_OK
|
||
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.error = str(exc)[:500]
|
||
failures.append(f"{job.get('title') or type_name(type_id)}：{str(exc)[:200]}")
|
||
failed += 1
|
||
|
||
task.status = TASK_DONE if failed == 0 else (TASK_PARTIAL if ok > 0 else TASK_FAILED)
|
||
if failed:
|
||
uniq = list(dict.fromkeys(failures)) # 去重保序
|
||
detail = "；".join(uniq[:6])
|
||
if len(uniq) > 6:
|
||
detail += f"；…等共 {failed} 张失败"
|
||
if ok == 0:
|
||
task.error = f"全部生成失败。{detail}"
|
||
else:
|
||
task.error = f"部分生成失败（{failed} 张）。{detail}"
|