zycandClaude Opus 4.8
77759678a7
真因:商品库点「AI生成三视图」走 reference_product 路径,产出资产 name=「图片创作」、 metadata 只有 mode/product_id —— 没有 view=three_view,也非 tri_view 类。导致: ①商品库详情三视图面板刷新后认不出(显示「尚未生成」、采用也丢); ②视频项目第二步基础资产读不到「商品已有三视图」。 修:reference_product 资产与视频项目基础资产阶段生成的商品三视图同款标记 —— category=product_image + 名字含「商品三视图」+ metadata.view=three_view + product_id。 这样 products is_triview(排除 product_image 不误入角色区)、商品库前端回填(view===three_view)、 视频项目商品三视图判定都能认出。需部署生效;存量旧资产无标记不追溯。 Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2561 lines
139 KiB
Python
2561 lines
139 KiB
Python
import json
|
||
import math
|
||
import re
|
||
import subprocess
|
||
import tempfile
|
||
import time
|
||
import uuid
|
||
from datetime import timedelta
|
||
from decimal import Decimal
|
||
from io import BytesIO
|
||
from pathlib import Path
|
||
|
||
import requests
|
||
from django.conf import settings
|
||
from django.core.exceptions import ObjectDoesNotExist
|
||
from django.db import transaction
|
||
from django.utils import timezone
|
||
|
||
from apps.ai.models import AITask, ModelConfig
|
||
from apps.ai.providers import (
|
||
OpenAICompatibleProvider,
|
||
TtsNotConfigured,
|
||
VolcanoArkProvider,
|
||
VolcanoTtsProvider,
|
||
)
|
||
from apps.assets.models import Asset, AssetFile
|
||
from apps.assets.storage import TosStorage
|
||
from apps.billing.services.ledger import charge_reserved_credit, release_credit, reserve_credit
|
||
from apps.projects.models import (
|
||
BaseAssetGroup,
|
||
ExportJob,
|
||
ProjectStage,
|
||
ScriptSegment,
|
||
ScriptVersion,
|
||
StoryboardFrame,
|
||
StoryboardShot,
|
||
StoryboardShotVersion,
|
||
StoryboardVersion,
|
||
Timeline,
|
||
VideoSegment,
|
||
VideoSegmentVersion,
|
||
)
|
||
|
||
|
||
def get_default_model(capability: str) -> ModelConfig:
|
||
qs = (
|
||
ModelConfig.objects.select_related("provider")
|
||
.filter(capability=capability, status=ModelConfig.Status.ACTIVE, provider__status="active")
|
||
)
|
||
# 优先平台超管钦定的默认模型;未钦定则回落「最早创建的 active」(原行为,零回归)
|
||
return qs.filter(is_default=True).order_by("created_at").first() or qs.order_by("created_at").first()
|
||
|
||
|
||
def resolve_image_model(key: str | None) -> "ModelConfig | None":
|
||
"""前端「生图模型选择」→ ModelConfig。用户显式选的可以是 disabled 模型(故不按 status 过滤)。
|
||
· "volcano" → 火山官方 Seedream(取最新一版)
|
||
· "gpt-image"→ gpt-image-2(优先 active provider 的那个)
|
||
· "provider:name" 或裸 name → 精确匹配
|
||
解析不到返回 None,调用方回落 get_default_model。"""
|
||
if not key:
|
||
return None
|
||
qs = ModelConfig.objects.select_related("provider").filter(capability=ModelConfig.Capability.IMAGE)
|
||
if key == "volcano":
|
||
# 火山 Seedream:优先版本号最高的(seedream-5 > seedream-4),按 name 倒序
|
||
vqs = qs.filter(provider__name__in=OFFICIAL_DIRECT_PROVIDERS)
|
||
return vqs.filter(name__icontains="seedream").order_by("-name").first() or vqs.order_by("-name").first()
|
||
if key in ("gpt-image", "gpt-image-2"):
|
||
return (qs.filter(name__icontains="gpt-image", provider__status="active").first()
|
||
or qs.filter(name__icontains="gpt-image").first())
|
||
if ":" in key:
|
||
pname, mname = key.split(":", 1)
|
||
return qs.filter(provider__name=pname, name=mname).first()
|
||
return qs.filter(name=key).first()
|
||
|
||
|
||
# 火山官方直连(SeeDream 生图 / Seedance 视频 / 豆包文本)走 ARK SDK;其余 provider 一律
|
||
# 视为「OpenAI 兼容中转站」走通用适配器。加/换中转站 = DB 加一行 ModelProvider,零改代码。
|
||
# 注意:DB 里火山 provider 实际命名为 "volcengine"(豆包),必须包含,否则会被错路由到中转站。
|
||
OFFICIAL_DIRECT_PROVIDERS = {"volcengine", "volcano", "ark", "volcano_ark"}
|
||
|
||
|
||
def resolve_provider_credentials(provider) -> tuple[str | None, str | None]:
|
||
"""解析中转站凭证。可插拔顺序:DB(ModelProvider.base_url/api_key)优先 → settings(.env)回退。
|
||
两者都不写死;换站只改 DB 这一行,或改 .env 对应项。"""
|
||
base_url = (provider.base_url or "").strip() or settings.PROVIDER_BASE_URLS.get(provider.name)
|
||
api_key = (getattr(provider, "api_key", "") or "").strip() or settings.PROVIDER_KEYS.get(provider.name)
|
||
return (base_url or None), (api_key or None)
|
||
|
||
|
||
def build_provider(model_config: ModelConfig):
|
||
"""按 provider.name 分流:火山官方直连 → VolcanoArkProvider;其余 → 通用 OpenAICompatibleProvider。
|
||
两条路都走 resolve_provider_credentials,统一 DB→.env 优先级(官方直连的 None 再由 __post_init__ 回退 settings.VOLCANO)。"""
|
||
provider = model_config.provider
|
||
base_url, api_key = resolve_provider_credentials(provider)
|
||
if provider.name in OFFICIAL_DIRECT_PROVIDERS:
|
||
# 临时:视频(Seedance)借 AirDrama 账号的 ARK key,使真人素材库 asset:// 引用与 Seedance 同账号、可解析。
|
||
# 仅覆盖 VIDEO,图像/文本仍用自有 key。VIDEO_ARK_API_KEY 留空则不覆盖(回落默认)。待自有账号开通素材库后清掉。
|
||
video_key = (getattr(settings, "VIDEO_ARK_API_KEY", "") or "").strip()
|
||
if video_key and model_config.capability == ModelConfig.Capability.VIDEO:
|
||
api_key = video_key
|
||
return VolcanoArkProvider(base_url=base_url, api_key=api_key)
|
||
# api_version:某些中转站(yunqi)的 images/edits 需 Azure 风格 ?api-version=...。
|
||
# DB(ModelProvider.metadata.api_version)优先 → settings(.env)回退,与凭证同款可插拔。
|
||
api_version = (provider.metadata or {}).get("api_version") or settings.PROVIDER_API_VERSIONS.get(provider.name)
|
||
return OpenAICompatibleProvider(base_url=base_url, api_key=api_key, api_version=api_version or None)
|
||
|
||
|
||
def get_image_provider(model_config: ModelConfig):
|
||
return build_provider(model_config)
|
||
|
||
|
||
def get_text_provider(model_config: ModelConfig):
|
||
return build_provider(model_config)
|
||
|
||
|
||
def get_video_provider(model_config: ModelConfig):
|
||
return build_provider(model_config)
|
||
|
||
|
||
def estimate_cost(model_config: ModelConfig) -> Decimal:
|
||
return model_config.unit_price if model_config.unit_price > 0 else Decimal("1.0000")
|
||
|
||
|
||
def parse_segment_fields(block: str) -> tuple[str, str]:
|
||
"""从一镜文本里拆出(旁白, 画面)。
|
||
|
||
模型按 build_script_prompt 的格式输出「旁白：…/画面：…」标签行时精确拆分;
|
||
自带脚本/旧格式没有标签则两个字段都用整段(保持旧行为),字幕/故事板各自兜底。
|
||
"""
|
||
narration_lines: list[str] = []
|
||
visual_lines: list[str] = []
|
||
current: list[str] | None = None
|
||
for raw in (block or "").splitlines():
|
||
line = raw.strip()
|
||
if not line:
|
||
continue
|
||
matched = re.match(r"^(旁白|口播|台词|文案)\s*[:：]\s*(.*)$", line)
|
||
if matched:
|
||
current = narration_lines
|
||
if matched.group(2):
|
||
current.append(matched.group(2))
|
||
continue
|
||
matched = re.match(r"^(画面|镜头描述|视觉|画面描述)\s*[:：]\s*(.*)$", line)
|
||
if matched:
|
||
current = visual_lines
|
||
if matched.group(2):
|
||
current.append(matched.group(2))
|
||
continue
|
||
if re.match(r"^(镜头|分镜|场)\s*\d+", line):
|
||
continue # 「镜头N」标题行不计入任何字段
|
||
if current is not None:
|
||
current.append(line)
|
||
narration = " ".join(narration_lines).strip()
|
||
visual = " ".join(visual_lines).strip()
|
||
if not narration and not visual:
|
||
return block.strip(), block.strip()
|
||
return narration or visual, visual or narration
|
||
|
||
|
||
def build_cast_scene_extract_prompt(content: str) -> list[dict[str, str]]:
|
||
"""轻量抽取提示词:从镜头脚本里提炼人物 / 场景标签,并给每个标签一句可直接生图的画面提示词。"""
|
||
system = (
|
||
"你是短视频脚本分析助手。请从给定的镜头脚本中提取出现的『人物』和『场景』,"
|
||
"并为每个人物 / 场景写一句可直接用于文生图的画面提示词(中文,30 字内,描述外形 / 着装 / 环境 / 光线)。"
|
||
"人物提示词用 9:16 竖屏(出镜角色全身);场景提示词用 16:9 横屏(环境 / 背景空镜)。"
|
||
"人物指出镜的角色(例:女主、同事、闺蜜);场景指画面发生的地点或环境(例:卫生间、地铁、办公室)。"
|
||
"去重,人物与场景各最多 6 个。只输出一个 JSON 对象,不要 markdown 代码块,不要任何额外文字,格式如下：\n"
|
||
'{"cast":[{"name":"女主","prompt":"26岁都市女性,自然妆容,米色针织衫,柔和室内光,9:16竖屏"}],'
|
||
'"scenes":[{"name":"卫生间","prompt":"现代简约浴室,暖色灯光,干净台面,16:9横屏"}]}'
|
||
)
|
||
return [{"role": "system", "content": system}, {"role": "user", "content": f"镜头脚本如下：\n{content}".strip()}]
|
||
|
||
|
||
def _coerce_tag_entries(items: object, limit: int = 6) -> tuple[list[str], dict[str, str]]:
|
||
"""把模型回的 [{"name","prompt"}] 列表整理成 (标签列表, {标签: 提示词}),去重保序、容错。"""
|
||
names: list[str] = []
|
||
prompts: dict[str, str] = {}
|
||
if not isinstance(items, list):
|
||
return names, prompts
|
||
for item in items:
|
||
if isinstance(item, dict):
|
||
name = str(item.get("name") or "").strip()
|
||
prompt = str(item.get("prompt") or "").strip()
|
||
else:
|
||
name, prompt = str(item or "").strip(), ""
|
||
if not name or name in ("无", "暂无", "未提及") or name in names:
|
||
continue
|
||
names.append(name)
|
||
if prompt:
|
||
prompts[name] = prompt
|
||
if len(names) >= limit:
|
||
break
|
||
return names, prompts
|
||
|
||
|
||
def extract_cast_and_scenes(*, project, user, content: str) -> dict:
|
||
"""轻量调一次文本模型,从脚本里抽取人物 / 场景标签及每个标签的建议生图提示词。
|
||
|
||
出稿后的增益步骤(对齐流程文档「自动从脚本提取信息」),走完整的 AITask + 计费闭环
|
||
(reserve→charge/release,任务类型记为 script_optimization)。但全程 best-effort:
|
||
无可用模型 / 预扣失败 / 调用失败 / 解析失败都吞掉返回空,绝不阻断脚本生成主流程。
|
||
返回 {cast, scenes, cast_prompts, scene_prompts}。
|
||
"""
|
||
empty = {"cast": [], "scenes": [], "cast_prompts": {}, "scene_prompts": {}}
|
||
model_config = get_default_model(ModelConfig.Capability.TEXT)
|
||
if model_config is None or not (content or "").strip():
|
||
return empty
|
||
|
||
messages = build_cast_scene_extract_prompt(content)
|
||
try:
|
||
task = create_ai_task(
|
||
project=project,
|
||
user=user,
|
||
task_type=AITask.Type.SCRIPT_OPTIMIZATION,
|
||
model_config=model_config,
|
||
request_payload={"model": model_config.name, "endpoint": model_config.endpoint, "messages": messages},
|
||
)
|
||
except Exception:
|
||
return empty # 余额不足等预扣失败:跳过提取,不挡出稿
|
||
reservation = task.credit_reservation
|
||
|
||
try:
|
||
task.status = AITask.Status.SUBMITTED
|
||
task.submitted_at = timezone.now()
|
||
task.save(update_fields=["status", "submitted_at", "updated_at"])
|
||
|
||
provider = build_provider(model_config)
|
||
response = provider.chat_completion(model=model_config.name, endpoint=model_config.endpoint, messages=messages)
|
||
text = provider.extract_text(response)
|
||
|
||
# LLM 调用已真实消耗 token → 按成功计费(无论后续能否解析出标签)
|
||
with transaction.atomic():
|
||
task.status = AITask.Status.SUCCEEDED
|
||
task.response_payload = response
|
||
task.actual_cost = task.estimated_cost
|
||
task.completed_at = timezone.now()
|
||
task.save(update_fields=["status", "response_payload", "actual_cost", "completed_at", "updated_at"])
|
||
charge_reserved_credit(reservation=reservation, actual_amount=task.actual_cost)
|
||
|
||
match = re.search(r"\{.*\}", text, re.DOTALL) # 容忍模型多裹了 markdown / 解释文字
|
||
if not match:
|
||
return empty
|
||
data = json.loads(match.group(0))
|
||
cast, cast_prompts = _coerce_tag_entries(data.get("cast"))
|
||
scenes, scene_prompts = _coerce_tag_entries(data.get("scenes"))
|
||
return {"cast": cast, "scenes": scenes, "cast_prompts": cast_prompts, "scene_prompts": scene_prompts}
|
||
except Exception as exc:
|
||
with transaction.atomic():
|
||
task.status = AITask.Status.FAILED
|
||
task.error_message = str(exc)
|
||
task.completed_at = timezone.now()
|
||
task.save(update_fields=["status", "error_message", "completed_at", "updated_at"])
|
||
release_credit(reservation=reservation, reason=str(exc))
|
||
return empty
|
||
|
||
|
||
def _skills_root() -> "Path":
|
||
"""skills 根目录。优先 BASE_DIR/skills(= core/backend/skills,随后端打进 Docker 镜像);
|
||
回落仓库根 BASE_DIR.parent.parent/skills(本地/旧布局)。
|
||
|
||
⚠️ 历史坑:镜像由 `./core/backend` 构建,skills 旧在仓库根 → 不在构建上下文 → 镜像里没有 →
|
||
`_load_skill_system_prompt` 返回空串 → 提取拿不到「只输出 JSON」铁律 → 模型吐散文 → 解析失败。
|
||
现 skills 已挪进 core/backend 随镜像打包;此处仍保留双路径兜底。"""
|
||
from pathlib import Path
|
||
|
||
from django.conf import settings
|
||
|
||
base = Path(settings.BASE_DIR)
|
||
for cand in (base / "skills", base.parent.parent / "skills"):
|
||
if cand.is_dir():
|
||
return cand
|
||
return base / "skills"
|
||
|
||
|
||
def _load_skill_system_prompt(name: str) -> str:
|
||
"""读取 skills/<name>/SKILL.md + references/*.md 拼成系统提示词(领域知识)。缺文件返回空串,不致命。"""
|
||
skill_dir = _skills_root() / name
|
||
parts: list[str] = []
|
||
main = skill_dir / "SKILL.md"
|
||
if main.exists():
|
||
parts.append(main.read_text(encoding="utf-8"))
|
||
ref_dir = skill_dir / "references"
|
||
if ref_dir.exists():
|
||
for ref in sorted(ref_dir.glob("*.md")):
|
||
parts.append(f"\n\n===== references/{ref.name} =====\n\n{ref.read_text(encoding='utf-8')}")
|
||
return "\n".join(parts)
|
||
|
||
|
||
def _normalize_extracted_entities(items: object) -> list[dict]:
|
||
"""整理提取出的 entities:只留 character/scene(绝不收 product),保留模型给的 id(供 segment refs 对应),
|
||
按 id 去重,角色/场景各最多 6 个,补 ref_index。"""
|
||
out: list[dict] = []
|
||
if not isinstance(items, list):
|
||
return out
|
||
seen_ids: set[str] = set()
|
||
n_char = n_scene = 0
|
||
for it in items:
|
||
if not isinstance(it, dict):
|
||
continue
|
||
typ = str(it.get("type") or "").strip()
|
||
if typ not in ("character", "scene"): # 丢弃 product / 未知类型
|
||
continue
|
||
name = str(it.get("name") or "").strip()
|
||
eid = str(it.get("id") or "").strip()
|
||
if not name or not eid or eid in seen_ids:
|
||
continue
|
||
if typ == "character":
|
||
n_char += 1
|
||
if n_char > 6:
|
||
continue
|
||
else:
|
||
n_scene += 1
|
||
if n_scene > 6:
|
||
continue
|
||
seen_ids.add(eid)
|
||
out.append(
|
||
{
|
||
"id": eid,
|
||
"type": typ,
|
||
"name": name,
|
||
"visual_prompt": str(it.get("visual_prompt") or "").strip(),
|
||
"ref_index": len(out) + 1,
|
||
}
|
||
)
|
||
return out
|
||
|
||
|
||
def _normalize_extracted_segment_refs(items: object, valid_ids: set[str]) -> list[dict]:
|
||
"""整理每镜 refs:只留指向存活实体(角色/场景)的 id,丢掉被剔除的(如商品/无效 id)。"""
|
||
out: list[dict] = []
|
||
if not isinstance(items, list):
|
||
return out
|
||
for it in items:
|
||
if not isinstance(it, dict):
|
||
continue
|
||
idx = it.get("index")
|
||
if not isinstance(idx, int):
|
||
continue
|
||
raw = it.get("entity_refs") or it.get("refs") or []
|
||
refs = [r for r in raw if isinstance(r, str) and r in valid_ids]
|
||
out.append({"index": idx, "entity_refs": refs})
|
||
return out
|
||
|
||
|
||
# 提取步固定锁定豆包 2.0 Pro(与脚本生成同款),不靠 get_default_model 的「最早创建」排序——
|
||
# 避免不同环境 DB 创建序漂移把提取路由到别的(中转站)推理模型。取不到再回落默认文本模型。
|
||
EXTRACT_TEXT_MODEL_NAME = "doubao-seed-2-0-pro-260215"
|
||
|
||
|
||
def _resolve_extract_model_config():
|
||
"""提取实体用的文本模型:优先豆包 2.0 Pro(active 且 provider active),否则回落默认文本模型。"""
|
||
pinned = (
|
||
ModelConfig.objects.select_related("provider")
|
||
.filter(
|
||
name=EXTRACT_TEXT_MODEL_NAME,
|
||
capability=ModelConfig.Capability.TEXT,
|
||
status=ModelConfig.Status.ACTIVE,
|
||
provider__status="active",
|
||
)
|
||
.order_by("created_at")
|
||
.first()
|
||
)
|
||
return pinned or get_default_model(ModelConfig.Capability.TEXT)
|
||
|
||
|
||
def _collect_extract_text(provider, model_config, messages) -> tuple[str, dict]:
|
||
"""走与「脚本生成」同一条已在生产验证稳定的流式通道把模型输出收全。
|
||
|
||
豆包 seed-pro / GPT / Gemini 等思考模型,思考期只发 reasoning_content、正文期才发 content,
|
||
流式分支(chat_completion_stream)已分别转发为 `reasoning` / `delta` 事件。这里**只收正文 delta**;
|
||
仅当正文为空(整轮被思考占满的极端兜底)才回退已收到的 reasoning。如此从根上避免非流式只读
|
||
content 拿到空串 → 解析不到 JSON 的老问题。返回 (text, 存档用精简 payload)。
|
||
"""
|
||
content_parts: list[str] = []
|
||
reasoning_parts: list[str] = []
|
||
for ev in provider.chat_completion_stream(
|
||
model=model_config.name,
|
||
endpoint=model_config.endpoint,
|
||
messages=messages,
|
||
temperature=0.3, # 结构化抽取要稳:低温降低 JSON 漂移
|
||
):
|
||
etype = ev.get("type")
|
||
if etype == "delta":
|
||
content_parts.append(ev.get("text") or "")
|
||
elif etype == "reasoning":
|
||
reasoning_parts.append(ev.get("text") or "")
|
||
content = "".join(content_parts).strip()
|
||
reasoning = "".join(reasoning_parts).strip()
|
||
text = content or reasoning
|
||
payload = {
|
||
"streamed": True,
|
||
"model": model_config.name,
|
||
"content": content,
|
||
"reasoning_chars": len(reasoning),
|
||
"reasoning_preview": reasoning[:500], # 失败时看一眼模型在想啥/有没有跑题
|
||
"system_chars": sum(len(m.get("content") or "") for m in messages if m.get("role") == "system"),
|
||
}
|
||
return text, payload
|
||
|
||
|
||
# 在途状态:据此判「已有提取在跑」(提交侧防重复扣费 + 前端刷新后重建 loading)
|
||
_EXTRACT_INFLIGHT = (
|
||
AITask.Status.CREATED,
|
||
AITask.Status.RESERVED,
|
||
AITask.Status.SUBMITTED,
|
||
AITask.Status.POLLING,
|
||
)
|
||
|
||
|
||
# 写死的提取输出契约兜底(对齐脚本 agent 的 _OUTPUT_PROTOCOL):随代码进镜像、永远在。
|
||
# **只管"格式骨架"**(JSON 形状 / id 规则 / 不提商品 / 只输出 JSON)—— 这是流水线硬底线,skill 丢了也不会塌。
|
||
# **不碰任何"创作细则"**(穿搭/空手/字数/visual_prompt 怎么写),那些归 skill 正文独占,避免两处重复→漂移→稀释质量。
|
||
_EXTRACT_OUTPUT_CONTRACT = """
|
||
|
||
---
|
||
|
||
## 输出契约(硬性 · 仅格式底线)
|
||
**只输出且仅输出一个 JSON 对象**(UTF-8,无注释,无 ```json 代码块外的任何散文/解释/markdown),形状如下:
|
||
{
|
||
"entities": [
|
||
{"id": "c1", "type": "character", "name": "角色称呼", "visual_prompt": "一句生图描述"},
|
||
{"id": "s1", "type": "scene", "name": "地点名", "visual_prompt": "一句生图描述"}
|
||
],
|
||
"segments": [{"index": 0, "entity_refs": ["c1", "s1"]}]
|
||
}
|
||
硬性规则(**仅格式;角色/场景怎么写、穿搭/visual_prompt 细则一律以上文技能正文为准**):
|
||
① 只认 character / scene 两类,**绝不输出 product 商品实体**;② 角色 id 用 c1/c2…、场景 id 用 s1/s2…,严格去重;
|
||
③ segments 必须覆盖输入每一镜(index 从 0 开始),列出该镜出现的角色/场景 id;④ 整个回复就是那个 JSON,无前后缀。
|
||
"""
|
||
|
||
|
||
def get_inflight_extraction(project):
|
||
"""本项目正在跑的实体提取任务(最近一条优先);无则 None。"""
|
||
return (
|
||
AITask.objects.filter(
|
||
project=project, task_type=AITask.Type.ENTITY_EXTRACTION, status__in=_EXTRACT_INFLIGHT
|
||
)
|
||
.order_by("-created_at")
|
||
.first()
|
||
)
|
||
|
||
|
||
def submit_extract_entities(*, project, user) -> AITask:
|
||
"""提交独立实体提取步(**异步**)。读已定稿(或最新)脚本 → 校验 + 建 RESERVED 任务 + 预留额度(秒级),
|
||
慢活(豆包思考模型流式抽取,可达数十秒)交给 Celery worker(run_extract_entities_task)跑。
|
||
|
||
这样 Web 层(gunicorn/nginx)不被数十秒的模型请求占住 → 不再 502。
|
||
**防重复扣费**:本项目已有在途提取任务时直接复用它,不新建/不二次预扣
|
||
(用户刷新后手痒重点、或并发点击都安全 —— 提取是实体唯一权威来源,本就只需跑一次)。
|
||
校验失败(无脚本/无分镜/无模型/额度不足)抛 ValueError 由端点转 400;运行期失败由 worker 记进
|
||
task.error_message,前端轮询 extract-status 读取。返回 AITask。"""
|
||
from apps.projects.models import ScriptVersion
|
||
from apps.ai.tasks import extract_entities_task
|
||
|
||
script = (
|
||
ScriptVersion.objects.filter(project=project, is_adopted=True).order_by("-created_at").first()
|
||
or ScriptVersion.objects.filter(project=project).order_by("-created_at").first()
|
||
)
|
||
if script is None:
|
||
raise ValueError("请先生成并定稿脚本,再提取角色 / 场景")
|
||
segments = list(script.segments.order_by("sort_order"))
|
||
if not segments:
|
||
raise ValueError("脚本没有分镜,无法提取")
|
||
|
||
inflight = get_inflight_extraction(project)
|
||
if inflight is not None:
|
||
return inflight # 已有提取在跑:复用,绝不二次预扣 / 重复出活
|
||
|
||
model_config = _resolve_extract_model_config()
|
||
if model_config is None:
|
||
raise ValueError("没有可用的文本模型")
|
||
|
||
product = project.product
|
||
if product is not None:
|
||
sp = "、".join([p.title for p in product.selling_points.all()[:5]])
|
||
prod_line = f"名称：{product.title}\n品类：{product.category or '未填'}\n卖点：{sp or '未填'}"
|
||
else:
|
||
prod_line = "（无商品信息）"
|
||
seg_lines = [
|
||
f"镜{i} role={s.role or ''} narration={(s.narration or '').strip()} visual={(s.visual_prompt or '').strip()}"
|
||
for i, s in enumerate(segments)
|
||
]
|
||
user_msg = (
|
||
f"商品信息：\n{prod_line}\n\n分镜脚本（共 {len(segments)} 镜，index 从 0 开始）：\n" + "\n".join(seg_lines)
|
||
)
|
||
# skill 正文(领域功力)+ 写死的输出契约兜底(保证「只输出 JSON」永远在,即便 skill 丢失)
|
||
system = _load_skill_system_prompt("ecommerce-entity-extract") + _EXTRACT_OUTPUT_CONTRACT
|
||
messages = [{"role": "system", "content": system}, {"role": "user", "content": user_msg}]
|
||
|
||
try:
|
||
task = create_ai_task(
|
||
project=project,
|
||
user=user,
|
||
task_type=AITask.Type.ENTITY_EXTRACTION,
|
||
model_config=model_config,
|
||
request_payload={
|
||
"model": model_config.name,
|
||
"endpoint": model_config.endpoint,
|
||
"messages": messages,
|
||
"script_id": str(script.id),
|
||
},
|
||
)
|
||
except Exception as exc: # 余额不足等预扣失败
|
||
raise ValueError("额度不足,无法提取(请先充值)") from exc
|
||
|
||
extract_entities_task.delay(str(task.id))
|
||
return task
|
||
|
||
|
||
def run_extract_entities_task(*, task_id: str) -> None:
|
||
"""Celery worker 内执行实体提取慢活:流式调模型 → 解析 JSON → 落库(覆盖 project.metadata 的
|
||
cast/scenes/*_prompts/script_entities + 回填每镜 entity_refs + entities_extracted 标记)并扣费;
|
||
失败退费并把可读错误记进 task.error_message(前端轮询 extract-status 读取)。
|
||
幂等:只处理 RESERVED 任务,重复投递不会二次出活、二次扣费。"""
|
||
from apps.projects.models import ScriptVersion
|
||
from apps.ai.script_agent import _map_entities_to_project_metadata
|
||
|
||
task = AITask.objects.select_related("team", "created_by", "project", "model_config").filter(id=task_id).first()
|
||
if task is None or task.status != AITask.Status.RESERVED:
|
||
return
|
||
project = task.project
|
||
payload = task.request_payload or {}
|
||
messages = payload.get("messages") or []
|
||
model_config = task.model_config
|
||
reservation = task.credit_reservation
|
||
response: dict = {} # 模型返回的精简存档;失败时也落库(便于事后定位"模型到底吐了啥")
|
||
try:
|
||
task.status = AITask.Status.SUBMITTED
|
||
task.submitted_at = timezone.now()
|
||
task.save(update_fields=["status", "submitted_at", "updated_at"])
|
||
|
||
provider = build_provider(model_config)
|
||
text, response = _collect_extract_text(provider, model_config, messages)
|
||
|
||
match = re.search(r"\{.*\}", text, re.DOTALL)
|
||
if not match:
|
||
raise ValueError("提取结果解析失败(模型未返回有效 JSON),请重试")
|
||
try:
|
||
data = json.loads(match.group(0))
|
||
except json.JSONDecodeError as exc: # 抓出来给用户可读话术,别把裸解析报错塞进 error_message
|
||
raise ValueError("提取结果解析失败,请重试") from exc
|
||
entities = _normalize_extracted_entities(data.get("entities"))
|
||
if not entities:
|
||
raise ValueError("没有从脚本里识别到角色 / 场景,可调整脚本后重试")
|
||
seg_refs = _normalize_extracted_segment_refs(data.get("segments"), {e["id"] for e in entities})
|
||
|
||
script_id = payload.get("script_id")
|
||
script = ScriptVersion.objects.filter(id=script_id).first() if script_id else None
|
||
if script is None: # 极端兜底:payload 里 script 没了就回落项目当前定稿稿
|
||
script = (
|
||
ScriptVersion.objects.filter(project=project, is_adopted=True).order_by("-created_at").first()
|
||
or ScriptVersion.objects.filter(project=project).order_by("-created_at").first()
|
||
)
|
||
segments = list(script.segments.order_by("sort_order")) if script is not None else []
|
||
|
||
with transaction.atomic():
|
||
task.status = AITask.Status.SUCCEEDED
|
||
task.response_payload = response
|
||
task.actual_cost = task.estimated_cost
|
||
task.completed_at = timezone.now()
|
||
task.save(update_fields=["status", "response_payload", "actual_cost", "completed_at", "updated_at"])
|
||
charge_reserved_credit(reservation=reservation, actual_amount=task.actual_cost)
|
||
|
||
# 落库:覆盖 project.metadata + 回填每镜 entity_refs(提取为实体唯一权威来源)。
|
||
# entities_extracted 标记:前端资产页提取闸门据此显隐(没走过 → 盖蒙版露按钮;走过 → 露卡片)。
|
||
_map_entities_to_project_metadata(project, entities)
|
||
md_flag = dict(project.metadata or {})
|
||
md_flag["entities_extracted"] = True
|
||
project.metadata = md_flag
|
||
project.save(update_fields=["metadata", "updated_at"])
|
||
refs_by_index = {r["index"]: r["entity_refs"] for r in seg_refs}
|
||
for i, seg in enumerate(segments):
|
||
new_refs = refs_by_index.get(i, [])
|
||
if (seg.entity_refs or []) != new_refs:
|
||
seg.entity_refs = new_refs
|
||
seg.save(update_fields=["entity_refs", "updated_at"])
|
||
except Exception as exc: # noqa: BLE001 — 失败退费并把错误记进 AITask 供前端轮询;不向上抛(避免 celery 重试二次扣费)
|
||
# ValueError 是我们给用户写好的可读话术(解析失败 / 没识别到角色等);其余(网络/模型异常)给通用话术。
|
||
msg = str(exc) if isinstance(exc, ValueError) else "提取调用失败,请重试"
|
||
with transaction.atomic():
|
||
task.status = AITask.Status.FAILED
|
||
task.error_message = msg
|
||
task.response_payload = response # 即便失败也存下模型输出片段(content/reasoning 长度等),供事后定位
|
||
task.completed_at = timezone.now()
|
||
task.save(update_fields=["status", "response_payload", "error_message", "completed_at", "updated_at"])
|
||
release_credit(reservation=reservation, reason=msg)
|
||
|
||
|
||
def split_script_into_segments(content: str, count: int = 4) -> list[str]:
|
||
"""把一段脚本稳健地拆成 `count` 个分镜文本,保证每镜都非空、且所有内容都被分配到某一镜。
|
||
|
||
原实现按行 `[:4]`,ARK 返回整段散文时常变成「第1镜有词、2/3/4镜全空」,
|
||
导致后续故事板帧 / 视频段拿到空提示词,前后内容断裂。这里改为:
|
||
优先按空行/标号块切,块数够就把全部块均匀分桶;块不够再按句子切;仍不够则补齐。
|
||
"""
|
||
|
||
def _bucketize(items: list[str], joiner: str) -> list[str]:
|
||
buckets: list[list[str]] = [[] for _ in range(count)]
|
||
per = len(items) / count
|
||
for index, item in enumerate(items):
|
||
buckets[min(count - 1, int(index / per))].append(item)
|
||
return [joiner.join(bucket).strip() for bucket in buckets]
|
||
|
||
text = (content or "").strip()
|
||
if not text:
|
||
return [""] * count
|
||
|
||
# 1) 优先按空行分段;只有一段时退回按行分
|
||
blocks = [block.strip() for block in re.split(r"\n\s*\n", text) if block.strip()]
|
||
if len(blocks) < 2:
|
||
blocks = [line.strip() for line in text.splitlines() if line.strip()]
|
||
if len(blocks) >= count:
|
||
return _bucketize(blocks, "\n")
|
||
|
||
# 2) 段落不足:按中英文句末标点切句,再均匀分桶
|
||
sentences = [s.strip() for s in re.split(r"(?<=[。！？!?.;；\n])", text) if s.strip()]
|
||
if len(sentences) >= count:
|
||
return _bucketize(sentences, " ")
|
||
|
||
# 3) 仍不足:用已有块/句补齐到 count,绝不留空镜
|
||
base = blocks or sentences or [text]
|
||
filled = list(base)
|
||
while len(filled) < count:
|
||
filled.append(base[-1])
|
||
return filled[:count]
|
||
|
||
|
||
@transaction.atomic
|
||
def create_ai_task(*, project, user, task_type: str, model_config: ModelConfig, request_payload: dict) -> AITask:
|
||
cost = estimate_cost(model_config)
|
||
task = AITask.objects.create(
|
||
team=project.team,
|
||
created_by=user,
|
||
project=project,
|
||
task_type=task_type,
|
||
status=AITask.Status.CREATED,
|
||
model_config=model_config,
|
||
idempotency_key=f"{task_type}:{project.id}:{uuid.uuid4()}",
|
||
request_payload=request_payload,
|
||
estimated_cost=cost,
|
||
)
|
||
reserve_credit(team=project.team, user=user, task=task, amount=cost)
|
||
task.status = AITask.Status.RESERVED
|
||
task.save(update_fields=["status", "updated_at"])
|
||
return task
|
||
|
||
|
||
def regenerate_script_segment(*, project, user, segment, instruction: str = "") -> ScriptVersion:
|
||
"""单镜重跑(「场次刷新」按钮):复用脚本 agent 的精准改一镜——读全脚本上下文、只动该镜、保留其余镜,落新 ScriptVersion。
|
||
旧的「散文 prompt + 正则解析」整条已废,统一走 agent(结构化 + entity_refs 不丢)。"""
|
||
from apps.ai.script_agent import regenerate_segment_via_agent
|
||
|
||
model_config = get_default_model(ModelConfig.Capability.TEXT)
|
||
if model_config is None:
|
||
raise ValueError("no active text model configured")
|
||
return regenerate_segment_via_agent(
|
||
project=project, user=user, model_config=model_config, segment=segment, instruction=instruction
|
||
)
|
||
|
||
|
||
def _generate_video_poster(*, video_bytes: bytes, team, project, asset_id) -> "StoredObject | None":
|
||
"""用 ffmpeg 抽视频首帧作为封面(poster)并上传 TOS。best-effort:任何失败都返回 None,不影响视频资产落地。"""
|
||
if not video_bytes:
|
||
return None
|
||
try:
|
||
with tempfile.TemporaryDirectory(prefix="airshelf-poster-") as tmp:
|
||
tmp_dir = Path(tmp)
|
||
video_path = tmp_dir / "in.mp4"
|
||
poster_path = tmp_dir / "poster.jpg"
|
||
video_path.write_bytes(video_bytes)
|
||
proc = subprocess.run(
|
||
["ffmpeg", "-y", "-ss", "0", "-i", str(video_path), "-frames:v", "1", "-q:v", "3", str(poster_path)],
|
||
capture_output=True,
|
||
timeout=60,
|
||
)
|
||
if proc.returncode != 0 or not poster_path.exists():
|
||
return None
|
||
poster_bytes = poster_path.read_bytes()
|
||
if not poster_bytes:
|
||
return None
|
||
object_key = f"teams/{team.id}/projects/{project.id}/generated/{asset_id}-poster.jpg"
|
||
return TosStorage().upload_fileobj(
|
||
fileobj=BytesIO(poster_bytes), object_key=object_key, content_type="image/jpeg"
|
||
)
|
||
except Exception: # noqa: BLE001 — poster 仅用于展示,失败不阻断
|
||
return None
|
||
|
||
|
||
def _store_generated_media(*, team, user, project, task, media: str, name: str, category: str, asset_type: str) -> Asset:
|
||
fileobj, content_type = VolcanoArkProvider.media_to_bytes(media)
|
||
suffix = ".png"
|
||
if "video" in content_type:
|
||
suffix = ".mp4"
|
||
elif "jpeg" in content_type:
|
||
suffix = ".jpg"
|
||
elif "webp" in content_type:
|
||
suffix = ".webp"
|
||
asset_id = uuid.uuid4()
|
||
object_key = f"teams/{team.id}/projects/{project.id}/generated/{asset_id}{suffix}"
|
||
stored = TosStorage().upload_fileobj(fileobj=fileobj, object_key=object_key, content_type=content_type)
|
||
asset = Asset.objects.create(
|
||
id=asset_id,
|
||
team=team,
|
||
created_by=user,
|
||
name=name,
|
||
asset_type=asset_type,
|
||
source=Asset.Source.AI_GENERATED,
|
||
category=category,
|
||
origin_task=task,
|
||
)
|
||
AssetFile.objects.create(
|
||
asset=asset,
|
||
object_key=stored.object_key,
|
||
bucket=stored.bucket,
|
||
content_type=stored.content_type,
|
||
size_bytes=stored.size_bytes,
|
||
is_primary=True,
|
||
)
|
||
# 视频资产:额外抽首帧作为封面图,挂成同一 Asset 下的 image 文件,供任务中心/列表显示缩略图
|
||
if "video" in content_type:
|
||
try:
|
||
video_bytes = fileobj.getvalue() if isinstance(fileobj, BytesIO) else b""
|
||
except Exception: # noqa: BLE001
|
||
video_bytes = b""
|
||
poster = _generate_video_poster(video_bytes=video_bytes, team=team, project=project, asset_id=asset_id)
|
||
if poster:
|
||
AssetFile.objects.create(
|
||
asset=asset,
|
||
object_key=poster.object_key,
|
||
bucket=poster.bucket,
|
||
content_type=poster.content_type,
|
||
size_bytes=poster.size_bytes,
|
||
is_primary=False,
|
||
)
|
||
return asset
|
||
|
||
|
||
def _find_entity_group(project, kind: str, label: str, group_id: str | None):
|
||
"""复用同一实体的基础资产组(版本=candidate_assets,采用=adopted_asset)。
|
||
优先 group_id;商品=该项目唯一商品组;人物/场景按 label 命中非三视图组;都没有则 None(新建)。"""
|
||
if group_id:
|
||
return project.base_asset_groups.filter(id=group_id).first()
|
||
candidates = list(project.base_asset_groups.filter(kind=kind).order_by("created_at"))
|
||
candidates = [g for g in candidates if not (g.metadata or {}).get("triview_of")] # 三视图组不算实体本体
|
||
if kind == BaseAssetGroup.Kind.PRODUCT:
|
||
return candidates[0] if candidates else None
|
||
lbl = (label or "").strip()
|
||
if lbl:
|
||
return next((g for g in candidates if (g.metadata or {}).get("label") == lbl), None)
|
||
return None
|
||
|
||
|
||
def _ratio_to_image_size(ratio: str) -> str:
|
||
"""前端比例 → gpt-image 支持的尺寸(宽高均需被 16 整除)。
|
||
⚠️ 该网关只接受 1024x1024 / 1024x1536 / 1536x1024 三种固定 size,无法直接产出精确 3:4 / 9:16 / 4:5,
|
||
这里只能取最接近的「竖/横/方」近似;要拿到精确像素比例需在出图后做主体保护裁切/补边(见优化文档 §7.1
|
||
normalize_output_image,属待部署的后处理项)。修复点:4:5 原会 fallback 成 1024x1024(方图)→ 比例错误,
|
||
现归到竖图 1024x1536。"""
|
||
return {
|
||
"1:1": "1024x1024",
|
||
"3:4": "1024x1536", # 近似竖图(网关无精确 3:4)
|
||
"4:5": "1024x1536", # 近似竖图(原 fallback 成方图是比例 bug)
|
||
"9:16": "1024x1536", # 近似竖图(网关无精确 9:16)
|
||
"4:3": "1536x1024",
|
||
"16:9": "1536x864", # 真 16:9(原来误用 1536x1024 = 3:2)
|
||
}.get((ratio or "").strip(), "1024x1024")
|
||
|
||
|
||
def _ratio_to_volcano_size(ratio: str) -> str:
|
||
"""前端比例 → 火山 Seedream 尺寸(~2K 面积,各边夹在 [1024,4096] 且取 16 的倍数)。
|
||
预设比例直接给好尺寸;自定义 W:H 按 2K 面积换算;解析不到回落 '2K'。"""
|
||
presets = {
|
||
"1:1": "2048x2048",
|
||
"3:4": "1728x2304",
|
||
"4:3": "2304x1728",
|
||
"9:16": "1440x2560",
|
||
"16:9": "2560x1440",
|
||
# 4:5 必须 ≥ Seedream 5.0 的最小面积 3,686,400 px:旧值 1664x2080=3.46M 会被 ARK 退 400
|
||
# (InvalidParameter: image size must be at least 3686400 pixels)。1728x2160 精确 4:5 且 3.73M 达标。
|
||
"4:5": "1728x2160",
|
||
}
|
||
r = (ratio or "").strip()
|
||
if r in presets:
|
||
return presets[r]
|
||
if ":" in r:
|
||
try:
|
||
w_str, h_str = r.split(":", 1)
|
||
w, h = float(w_str), float(h_str)
|
||
if w > 0 and h > 0:
|
||
scale = math.sqrt((2048 * 2048) / (w * h))
|
||
side = lambda v: min(max(int(round(v * scale / 16) * 16), 1024), 4096) # noqa: E731
|
||
return f"{side(w)}x{side(h)}"
|
||
except (ValueError, ZeroDivisionError):
|
||
pass
|
||
return "2K"
|
||
|
||
|
||
def _product_cover_url(product) -> str:
|
||
"""商品主图 URL:优先 cover_asset,其次标记为主图的商品图,再次首张商品图。无图返回 ''。"""
|
||
if product is None:
|
||
return ""
|
||
if product.cover_asset_id:
|
||
url = _asset_preview_url(product.cover_asset)
|
||
if url:
|
||
return url
|
||
image = product.images.filter(is_primary=True).first() or product.images.order_by("sort_order", "created_at").first()
|
||
if image is not None:
|
||
return _asset_preview_url(image.asset)
|
||
return ""
|
||
|
||
|
||
def _product_reference_urls(product, limit: int = 3) -> list[str]:
|
||
"""模特上身图的商品参考图(可多张):**真实上传图优先,排除 AI 生成图**——避免拿生成图当真相
|
||
再喂回模型造成误差累积。按主图/排序取前 limit 张真实上传图;一张都没有时回落 cover(即便是 AI 图,
|
||
至少保证能走 image_edit 而不是纯文生图)。"""
|
||
if product is None:
|
||
return []
|
||
from apps.assets.models import Asset
|
||
|
||
urls: list[str] = []
|
||
seen: set[str] = set()
|
||
rels = list(product.images.select_related("asset").all())
|
||
rels.sort(key=lambda im: (not im.is_primary, im.sort_order))
|
||
for im in rels:
|
||
a = im.asset
|
||
if a is None or getattr(a, "source", "") == Asset.Source.AI_GENERATED:
|
||
continue
|
||
u = _asset_preview_url(a)
|
||
if u and u not in seen:
|
||
seen.add(u)
|
||
urls.append(u)
|
||
if len(urls) >= limit:
|
||
return urls
|
||
if not urls: # 无任何真实上传图 → 回落 cover(可能是 AI 图,但好过纯文生图)
|
||
cover = _product_cover_url(product)
|
||
if cover:
|
||
urls.append(cover)
|
||
return urls
|
||
|
||
|
||
def quality_words(stage: str, slot: str = "quality") -> list[str]:
|
||
"""平台单层质量词配置(QualityWord)。无配置/表不存在 → 返回 [],调用方回落写死值,保证零回归。"""
|
||
try:
|
||
from apps.ai.models import QualityWord
|
||
|
||
return list(
|
||
QualityWord.objects.filter(stage=stage, slot=slot, enabled=True)
|
||
.order_by("sort", "created_at")
|
||
.values_list("text", flat=True)
|
||
)
|
||
except Exception: # noqa: BLE001 — 配置读取失败绝不阻断生成
|
||
return []
|
||
|
||
|
||
def quality_suffix(stage: str, fallback: str, slot: str = "quality", sep: str = "，") -> str:
|
||
"""取该阶段配置的质量词拼成尾串;无配置回落到 fallback(各 builder 的原写死值)。"""
|
||
words = quality_words(stage, slot)
|
||
return sep.join(words) if words else fallback
|
||
|
||
|
||
_PROMPT_PLACEHOLDER_RE = re.compile(r"\{(\w+)\}", re.UNICODE)
|
||
# gpt-image 要求宽高都能被 16 整除;1536x864 = 真 16:9(长边封顶 1536),1536x1024 = 3:2
|
||
_RATIO_TO_SIZE = {"1:1": "1024x1024", "portrait": "1024x1536", "landscape": "1536x1024", "16:9": "1536x864"}
|
||
|
||
|
||
def _prompt_template_row(key: str):
|
||
"""读 admin 可编辑的提示词模板行(启用的);缺表/缺行/异常一律返回 None(回落写死默认,零回归)。"""
|
||
try:
|
||
from apps.ai.models import PromptTemplate
|
||
|
||
return PromptTemplate.objects.filter(key=key, enabled=True).first()
|
||
except Exception: # noqa: BLE001 — 表还没迁移/查询异常都不该挡生成
|
||
return None
|
||
|
||
|
||
def render_prompt(key: str, default_text: str, **fields) -> str:
|
||
"""按 admin 可编辑模板渲染提示词。无模板/停用/正文空 → 回落 default_text(各 builder 的写死值)。
|
||
安全替换 {占位符}:已知字段换值,未知占位符**原样保留**(用户改错也不会让生成崩)。占位符名支持中文。"""
|
||
row = _prompt_template_row(key)
|
||
tpl = row.template if (row and (row.template or "").strip()) else default_text
|
||
return _PROMPT_PLACEHOLDER_RE.sub(lambda m: str(fields.get(m.group(1), m.group(0))), tpl)
|
||
|
||
|
||
def prompt_ratio_size(key: str, default_size: str) -> str:
|
||
"""模板配的比例 → gpt-image 尺寸(1024x1024 / 1024x1536 / 1536x1024);未配则用 default_size。"""
|
||
row = _prompt_template_row(key)
|
||
if row and row.ratio:
|
||
return _RATIO_TO_SIZE.get(row.ratio, default_size)
|
||
return default_size
|
||
|
||
|
||
def build_product_triview_prompt_refs(product, base_prompt: str = "") -> str:
|
||
"""商品三视图 image_edit 提示词(refs 版):参考图1=商品真实主图。正文可在 admin「提示词」页改;
|
||
{商品}=主图商品名(写死字段,不可删),{补充}=调用方附加文本。"""
|
||
name = (getattr(product, "title", "") or "商品").strip()
|
||
default = (
|
||
"参考图1是「{商品}」的真实商品主图。请严格参照该图的包装外形、品牌文字、配色、Logo 与材质，"
|
||
"生成同一件商品的三视图：从左到右依次为正面、侧面、背面，统一光照，纯白背景，16:9 构图。"
|
||
"三个视图必须是同一件商品，品牌字样/配色/外形高度一致，不要改动或重新设计包装。{补充}"
|
||
)
|
||
return render_prompt("product_triview", default, 商品=name, 补充=(base_prompt or "").strip())
|
||
|
||
|
||
# 模特上身图:按品类分「穿戴 / 非穿戴」注入不同的商品-模特关系;每张按序号变化动作/场景/镜头。
|
||
_TRYON_WEARABLE_HINTS = ("服", "衣", "裤", "裙", "鞋", "帽", "袜", "围巾", "外套", "卫衣", "内衣", "文胸", "胸罩", "泳", "bra")
|
||
_TRYON_VARIATIONS = [
|
||
{"action": "模特正面自然展示该商品", "scene": "干净的室内空间", "shot": "半身近景,商品清晰可见"},
|
||
{"action": "模特正在穿着 / 使用该商品", "scene": "生活化的真实居家场景", "shot": "侧面角度,突出穿着或使用方式"},
|
||
{"action": "模特手持或局部展示商品细节", "scene": "明亮的时尚生活场景", "shot": "中近景,商品占比较高"},
|
||
{"action": "模特与商品自然互动", "scene": "温暖时尚的生活场景", "shot": "半身,强调使用情境"},
|
||
]
|
||
_TRYON_NEGATIVE = (
|
||
"不要换人,不要改商品设计,不要改商品颜色与结构,不要生成错误或乱码的 Logo 与文字,"
|
||
"不要把商品改成相似款,不要多余的商品堆叠,不要多余文字,不要水印,不要边框,"
|
||
"不要低清模糊,不要过度磨皮,不要畸变,不要扭曲身体,不要夸张滤镜"
|
||
)
|
||
|
||
|
||
def _is_wearable_product(product) -> bool:
|
||
"""据类目/标题判断是否「穿戴类」(服饰鞋帽内衣) → 真实穿到身上;否则非穿戴 → 手持/佩戴/使用。"""
|
||
blob = f"{getattr(product, 'category', '') or ''} {getattr(product, 'title', '') or ''}".lower()
|
||
return any(h in blob for h in _TRYON_WEARABLE_HINTS)
|
||
|
||
|
||
def build_model_tryon_prompt_refs(product, has_model: bool, base_prompt: str = "", index: int = 0, n_product: int = 1) -> str:
|
||
"""模特上身图 image_edit 提示词(refs 版):
|
||
参考图1~N=商品真实图(多角度,锁外形/品牌/配色),参考图N+1=选中模特(锁人脸/身形/气质)。
|
||
按品类分穿戴/非穿戴(穿戴=真实穿身上、替换原衣;非穿戴=手持/佩戴/使用,不动原衣);
|
||
`index` 让每张图动作/场景/镜头不同;`n_product` 让参考图序号自适应。"""
|
||
name = (getattr(product, "title", "") or "商品").strip()
|
||
n_product = max(1, int(n_product or 1))
|
||
# 参考图序号自适应:N 张商品图 → 参考图1~N=商品, 参考图N+1=模特
|
||
if n_product <= 1:
|
||
intro = f"参考图1是「{name}」的真实商品图,是该商品外观的唯一依据。"
|
||
prod_ref = "参考图1"
|
||
model_idx = 2
|
||
else:
|
||
rng = f"1-{n_product}" if n_product > 2 else "1、2"
|
||
intro = f"参考图{rng}是「{name}」同一件真实商品的不同角度图,是该商品外观的唯一依据,请综合这些角度还原商品。"
|
||
prod_ref = f"参考图{rng}"
|
||
model_idx = n_product + 1
|
||
|
||
if _is_wearable_product(product):
|
||
relation = (
|
||
f"真实穿着{prod_ref}中的这件商品,替换掉模特原本的衣服,让商品自然合身地穿在身上,"
|
||
"而不是放在一旁展示,保持商品的版型、领口、袖型、长度、纹样不变"
|
||
)
|
||
else:
|
||
relation = (
|
||
f"自然地手持 / 在合适位置佩戴 / 正在使用{prod_ref}中的这件商品(如为耳机则佩戴在耳朵上),"
|
||
"不要改动模特原本的服装,不要把商品强行穿到身上"
|
||
)
|
||
|
||
var = _TRYON_VARIATIONS[index % len(_TRYON_VARIATIONS)]
|
||
lines = [intro]
|
||
if has_model:
|
||
lines.append(f"参考图{model_idx}是出镜模特。请生成参考图{model_idx}中这位模特{relation}的电商详情页效果图。")
|
||
lines.append(f"模特的五官、发型、肤色、身形、年龄与气质必须与参考图{model_idx}(模特图)高度一致,不要换人,不要自行生成另一位模特。")
|
||
else:
|
||
lines.append(f"请生成一位真人模特{relation}的电商详情页效果图。")
|
||
lines.append(f"商品的外形、配色、材质、品牌文字与 Logo、图案必须与{prod_ref}(商品图)严格一致,不要重新设计、不要改样、不要生成相似款。")
|
||
lines.append(f"本张画面:{var['action']};场景:{var['scene']};镜头:{var['shot']}。")
|
||
lines.append(quality_suffix("model_tryon", "自然光、真实质感、干净背景、电商主图构图,人物与商品比例真实协调。"))
|
||
if base_prompt and base_prompt.strip():
|
||
lines.append(base_prompt.strip())
|
||
lines.append("请规避:" + _TRYON_NEGATIVE)
|
||
return " ".join(lines)
|
||
|
||
|
||
# 平台套图同批多张要「同款商品、不同版式」,否则 N 张文案/构图雷同(PMC#24)。
|
||
# 锁死商品一致性,只让排版/构图/视角/配色基调按张变化。
|
||
# 注:仅纯文生图回落路径(无参考图)仍用这个简表;refs 版改用下面的 slot 体系。
|
||
_COVER_VARIATIONS = [
|
||
"本张:正面居中主图版式,商品占画面主体,纯净背景,经典电商主图构图",
|
||
"本张:换一种排版——商品偏置 + 卖点文案分区,场景化背景,杂志感构图",
|
||
"本张:特写细节版式,放大商品材质/做工,近景视角,突出质感",
|
||
"本张:生活场景套图版式,商品置于真实使用情境,环境光,氛围感构图",
|
||
"本张:多角度组合版式,商品换一个朝向/视角,几何分区背景,现代简约风",
|
||
"本张:促销封面版式,留出标题/价签区,高对比配色,强视觉冲击构图",
|
||
]
|
||
|
||
# 平台套图优化版(对照「平台套图线上提示词优化版.md」):平台差异 + slot 版式 + 低信息密度上限。
|
||
# 平台 id 用规范化 key(前端 dy/tb… 已在 PLATFORM_ID_MAP 里映射成这些 canonical key)。
|
||
_PLATFORM_NAMES = {
|
||
"taobao": "淘宝", "tmall": "天猫", "jd": "京东", "pdd": "拼多多", "douyin": "抖音电商",
|
||
"xhs": "小红书", "kuaishou": "快手", "wechat": "视频号", "amazon": "亚马逊", "1688": "1688",
|
||
}
|
||
|
||
# §4 平台块:每块只描述平台调性 / 版式倾向,统一服从「头图低信息密度」,不写比例(比例由 size 控制)。
|
||
_PLATFORM_COVER_BLOCKS = {
|
||
"taobao": (
|
||
"平台:淘宝。画面像淘宝搜索货架和商品主图,移动端缩略图下商品一眼可识别;"
|
||
"常见左上小品牌区、中央或偏右商品主体、底部克制活动条 / 轻卖点条;"
|
||
"背景干净但不单调,可用浅灰棚拍 / 浅色墙面 / 试衣间 / 窗边挂拍 / 干净台面;"
|
||
"禁止详情页 / 长图 / 平台 UI / 二维码 / 直播间界面 / 虚假价格与未提供折扣。"
|
||
),
|
||
"tmall": (
|
||
"平台:天猫。画面像品牌旗舰店商品首图,品牌感 / 质感 / 留白 / 材质光影更重要;"
|
||
"促销感弱于淘宝,不要大红大黄低价感,不要拼多多式强利益区。"
|
||
),
|
||
"jd": (
|
||
"平台:京东。画面清晰 / 可信 / 理性 / 标准,像京东商品主图;白底或浅灰商品图、克制品质短标题、包装 / 配件 / 材质细节;"
|
||
"信息区极少,只允许 1 个短标题或 1-2 个短标签;不要参数表 / 功能说明长区 / 详情页式模块,不要小红书滤镜,不要强生活方式过度氛围。"
|
||
),
|
||
"pdd": (
|
||
"平台:拼多多。主体大 / 信息直接 / 利益区明显,像商品主图或活动头图;"
|
||
"可有高对比活动区和大标题但只讲一个利益点,不能编造价格 / 优惠 / 折扣 / 销量 / 平台补贴;"
|
||
"背景明亮,商品边缘清楚,不要复杂品牌大片感。"
|
||
),
|
||
"douyin": (
|
||
"平台:抖音电商。像信息流商品封面 / 短视频电商封面,近景 / 强裁切 / 真实使用瞬间 / 动作感强;"
|
||
"常见背景:衣橱 / 梳妆台 / 卧室 / 开箱 / 手部整理 / 拿取 / 穿搭准备;文案更短、商品主体更大;"
|
||
"不要抖音 UI / 播放按钮 / 直播间贴片 / 字幕条,也不要套淘宝底部活动条。"
|
||
),
|
||
"xhs": (
|
||
"平台:小红书。像生活方式笔记封面,自然光 / 真实体验 / 种草氛围;"
|
||
"常见背景:卧室 / 衣橱 / 梳妆台 / 桌面 / 浴室 / 旅行收纳 / 通勤准备;标题像用户体验表达,少硬广 / 少参数表 / 少促销条。"
|
||
),
|
||
"kuaishou": (
|
||
"平台:快手。像快手小店商品图 / 直播前置封面素材,真实 / 直接 / 可信;"
|
||
"可有桌面展示 / 手持展示 / 打包台 / 家中真实环境 / 开箱场景;不要直播 UI / 主播贴片 / 平台水印。"
|
||
),
|
||
"wechat": (
|
||
"平台:视频号 / 微信小店。画面克制可信,适合社交分享和商品卡;"
|
||
"背景可用干净家居 / 礼赠 / 办公 / 生活空间 / 柔和自然光;不要微信聊天界面 / 二维码 / 公众号 UI。"
|
||
),
|
||
"amazon": (
|
||
"平台:亚马逊。第 1 张为 MAIN 合规图:纯白背景、只展示售卖商品本体、无文字 / 无道具 / 无边框 / 无水印,商品占画面主要区域;"
|
||
"第 2 张以后可为 Lifestyle / Feature / Detail / Package 辅图;不要 Amazon 徽章 / 评分 / 排名 / 优惠 / 平台 UI。"
|
||
),
|
||
"1688": (
|
||
"平台:1688。像批发采购图,商品清楚,偏规格 / 材质 / 工艺 / 包装 / 供货信息;"
|
||
"背景可用白底 / 浅灰 / 工厂台面 / 仓储 / 包装台 / 材料工艺背景;头图仍保持低信息密度,只允许 1 个短标题或 1-2 个短标签;"
|
||
"不要做成规格表详情页,不要虚构工厂资质 / 库存 / 起订量 / 认证与价格。"
|
||
),
|
||
}
|
||
|
||
# §3 / §3.1 slot 版式骨架(低信息密度版):默认 4 张按 hero→scene→selling→detail 取,8/12 张再补 multi/promo。
|
||
_COVER_SLOTS = {
|
||
"hero": "正面或四分之三角度的商品 / 上身主视觉,主体占画面 60%-80%,纯净干净背景,经典电商主图构图,无文案",
|
||
"scene": "场景主视觉:挂拍 / 衣架 / 生活场景 / 手部整理 / 使用情境,环境自然光,氛围感,无文案或仅 1 个极短标题",
|
||
"selling": "轻卖点封面:商品主体居中或偏置,最多 1 个短标题加 1-2 个极短标签,文字区克制,不堆参数",
|
||
"detail": "质感特写:放大材质 / 做工 / 关键结构(扣位 / 肩带 / 边缘走线),近景视角,突出质感,无文案或 1 个短标签",
|
||
"multi": "多角度组合:商品换一个朝向 / 视角,几何分区干净背景,现代简约风,无文案",
|
||
"promo": "促销封面:主体大、利益点单一,底部或角落留 1 条活动短语,高对比配色,不铺满文字、不编造价格",
|
||
}
|
||
_COVER_SLOT_ORDER = ["hero", "scene", "selling", "detail", "multi", "promo"]
|
||
|
||
# §3.1 / §8 头图低信息密度上限:所有平台 / 所有模型都必须遵守,防止漂成详情页。
|
||
_COVER_LOW_DENSITY = (
|
||
"这是一张商品头图,不是详情页:画面第一优先级是商品主体 / 上身效果 / 使用场景,"
|
||
"商品或模特主体占画面 60%-80%,文字与装饰不得抢主体;如需文案最多 1 个短标题(≤8 字)加 1-2 个极短标签;"
|
||
"严禁参数表 / 规格表 / 长句卖点 / 三段式说明 / 密集图标 / 大量箭头 / 2x3 或 3x3 信息宫格 / 左右对比详情页 / 多屏排版。"
|
||
)
|
||
|
||
# §5 背景反差:避免商品与背景同色相融(无法逐图取色时给通用避让规则)。
|
||
_COVER_BG_CONTRAST = (
|
||
"背景必须与商品主色形成清楚反差,不能让商品与背景同色相融;"
|
||
"浅色商品避免奶白 / 浅米 / 浅粉等近似大面积底,优先冷灰 / 蓝灰 / 浅绿 / 自然木色或明确投影、边缘光;"
|
||
"深色商品避免黑 / 深灰 / 深棕大底,优先暖白 / 浅灰 / 浅木色 / 日光窗边;"
|
||
"除亚马逊 MAIN 或明确白底图外,不要整组都做成纯白 / 浅灰底。"
|
||
)
|
||
|
||
# §6.1 内衣真人上身强约束:防止被第二件衣服遮挡、防性感化漂移。
|
||
_UNDERWEAR_ON_MODEL = (
|
||
"这是真人试穿商品头图:成人模特直接穿着参考商品中的内衣,内衣是唯一服饰重点,"
|
||
"罩杯 / 肩带 / 下围 / 杯型 / 面料 / 颜色与关键版型须完整可见且与参考商品一致;"
|
||
"不要出现 T 恤 / 衬衫 / 吊带背心 / 运动上衣 / 连衣裙 / 制服 / 外套等遮挡内衣的第二件衣服,"
|
||
"不要把内衣穿在其他衣物外面或在其下面加衣;最多允许轻薄开衫 / 薄纱松搭在肩臂外侧但不得盖住罩杯 / 肩带 / 下围;"
|
||
"正规电商服饰目录拍摄,成人、克制、商品展示向,非性感化、非挑逗姿势,关键结构不被裁切。"
|
||
)
|
||
|
||
# §9 通用负面提示词(平台套图专用,压缩版)。
|
||
_COVER_NEGATIVE = (
|
||
"请规避:详情页 / 详情长图 / 多屏排版 / 店铺页 / 平台 UI / 直播间界面 / 二维码 / 联系方式 / 水印 / 平台 Logo / 虚假角标;"
|
||
"不要重新设计商品或改其包装 / Logo / 配色 / 材质 / 文字 / 图案 / 比例,不要生成相似款或改品类,不要擅自加套装 / 配件 / 赠品;"
|
||
"不要出现商品数据外的品牌名 / 英文 Logo / 系列名 / 角落签名 / 模型名 / 伪水印;"
|
||
"不要编造价格 / 优惠券 / 满减 / 折扣 / 限时活动 / 销量 / 排名 / 认证 / 功效承诺 / 绝对化用语;"
|
||
"不要错别字 / 乱码 / 小字堆叠 / 大段文案;不要让文字遮挡主体;不要让商品与背景同色相融。"
|
||
)
|
||
|
||
# 内衣 / 贴身衣物品类提示(比通用穿戴更窄,用于触发内衣强约束分支)。
|
||
_UNDERWEAR_HINTS = ("内衣", "文胸", "胸罩", "bra", "内裤", "泳", "比基尼", "bikini", "睡衣", "塑身", "束身")
|
||
|
||
|
||
def _is_underwear_product(product) -> bool:
|
||
blob = f"{getattr(product, 'category', '') or ''} {getattr(product, 'title', '') or ''}".lower()
|
||
return any(h in blob for h in _UNDERWEAR_HINTS)
|
||
|
||
|
||
def build_platform_cover_prompt_refs(
|
||
product,
|
||
*,
|
||
has_model: bool = False,
|
||
base_prompt: str = "",
|
||
platform_id: str = "",
|
||
index: int = 0,
|
||
count: int = 4, # noqa: ARG001 — 透传保留,后续可据张数扩展 slot 选择
|
||
product_ref_count: int = 1,
|
||
) -> str:
|
||
"""平台套图 image_edit 提示词(refs 优化版,对照「平台套图线上提示词优化版.md」):
|
||
参考图1~N=同一件真实商品(锁外形/品牌/配色/Logo/比例),有模特时参考图N+1=出镜模特。
|
||
按 `platform_id` 注入平台块、按 `index` 选 slot 版式,统一服从「头图低信息密度」与「背景反差」,
|
||
内衣 + 有模特时叠加强约束。只出头图 / 主图 / 封面候选,不再出详情。"""
|
||
name = (getattr(product, "title", "") or "商品").strip()
|
||
n = max(1, int(product_ref_count or 1))
|
||
if n <= 1:
|
||
ref_word = "参考图1"
|
||
intro = f"参考图1是「{name}」的真实商品参考图。"
|
||
else:
|
||
rng = f"1-{n}" if n > 2 else "1、2"
|
||
ref_word = f"参考图{rng}"
|
||
intro = f"参考图{rng}是「{name}」同一件真实商品的不同角度参考图。"
|
||
lines = [intro]
|
||
lines.append(
|
||
f"请严格锁定{ref_word}中商品的外形、颜色、材质、结构、Logo、品牌文字与比例,"
|
||
"严禁重新设计 / 改样 / 改品类;这些参考图只锁商品本体,不锁原图里的床品 / 桌面 / 墙面 / 绿植 / 道具与拍摄光线。"
|
||
)
|
||
# 平台块(canonical key 命中则用 §4 平台块,否则回落平台名 / 通用)
|
||
block = _PLATFORM_COVER_BLOCKS.get(platform_id)
|
||
pname = _PLATFORM_NAMES.get(platform_id, "")
|
||
if block:
|
||
lines.append(block)
|
||
elif pname:
|
||
lines.append(f"请生成一张适合「{pname}」平台的商品头图 / 主图 / 封面候选,统一视觉风格。")
|
||
else:
|
||
lines.append("请生成一张电商平台商品头图 / 主图 / 封面候选,统一视觉风格。")
|
||
# 头图低信息密度上限
|
||
lines.append(_COVER_LOW_DENSITY)
|
||
# 模特身份 + 内衣强约束
|
||
if has_model:
|
||
lines.append(
|
||
f"参考图{n + 1}是出镜模特,请让这位模特真实展示该商品;"
|
||
"模特的五官 / 发型 / 肤色 / 身形须与该图高度一致,不要换人。"
|
||
)
|
||
if _is_underwear_product(product):
|
||
lines.append(_UNDERWEAR_ON_MODEL)
|
||
# 本张 slot 版式
|
||
slot_key = _COVER_SLOT_ORDER[index % len(_COVER_SLOT_ORDER)]
|
||
lines.append("本张版式:" + _COVER_SLOTS[slot_key] + "。")
|
||
# 背景反差
|
||
lines.append(_COVER_BG_CONTRAST)
|
||
if base_prompt and base_prompt.strip():
|
||
lines.append(
|
||
"用户补充(只影响氛围 / 构图 / 场景 / 光线 / 表达偏好,不得覆盖商品一致性、平台与版式规则):"
|
||
+ base_prompt.strip()
|
||
)
|
||
lines.append(_COVER_NEGATIVE)
|
||
return " ".join(lines)
|
||
|
||
|
||
def build_free_reference_prompt(base_prompt: str, n_refs: int = 1, index: int = 0) -> str:
|
||
"""图片创作自由模式 · 带用户上传参考图时的提示词。
|
||
纯把用户原话丢给图生图,模型只会松散借个色调、不会真的保留参考图里的主体(背心/商品/人物),
|
||
这正是「没参考我上传的素材」的根因。这里显式把参考图钉成「画面主体的唯一依据」,要求严格保留其
|
||
外形/款式/配色/材质/品牌文字/图案,再在此基础上按用户要求创作。"""
|
||
base = (base_prompt or "").strip()
|
||
if n_refs <= 1:
|
||
ref_intro = "参考图是用户提供的素材,是本次画面主体(商品 / 人物 / 物体)的唯一依据。"
|
||
ref_word = "参考图"
|
||
else:
|
||
rng = f"1-{n_refs}" if n_refs > 2 else "1、2"
|
||
ref_intro = f"参考图{rng}是用户提供的素材(同一主体的不同角度 / 多个主体),是本次画面主体的唯一依据。"
|
||
ref_word = f"参考图{rng}"
|
||
lines = [
|
||
ref_intro,
|
||
f"请严格保留{ref_word}中主体的外形、款式、配色、材质、纹样、品牌文字与 Logo,"
|
||
"不要重新设计、不要换款、不要生成相似但不同的物体;主体须与参考图高度一致。",
|
||
]
|
||
if base:
|
||
lines.append(f"在此基础上,按用户要求创作:{base}")
|
||
else:
|
||
lines.append("在此基础上,生成干净、专业的电商视觉画面。")
|
||
# 同批多张要不同构图/视角,否则雷同(PMC#24);主体一致性已在上面钉死,这里只变表现。
|
||
if index > 0:
|
||
lines.append(_COVER_VARIATIONS[index % len(_COVER_VARIATIONS)] + ",在保持主体一致的前提下换一种表现。")
|
||
lines.append("画面真实、构图协调、细节清晰。")
|
||
return " ".join(lines)
|
||
|
||
|
||
def build_person_frontal_prompt(description: str = "") -> str:
|
||
"""人物正面氛围图提示词:把脚本提取(或用户输入)的人物描述包成统一模板。
|
||
用户钦定格式:电商真人模特，氛围正面全身照，<描述>，自然妆容，柔和影棚光，真实质感，单人，纯色背景。"""
|
||
desc = (description or "").strip()
|
||
# 正文可在 admin「提示词」页改;{描述}=脚本提取/用户输入的人物描述(写死字段)。默认含原质量尾串。
|
||
default = "电商真人模特，氛围正面全身照，{描述}，自然妆容，柔和影棚光，真实质感，单人，纯色背景"
|
||
return render_prompt("person_portrait", default, 描述=desc)
|
||
|
||
|
||
# --------------------------------------------------------------------------- #
|
||
# 出图重试(中转站偶发抖动 → 一次失败就「生成不出来」的根因兜底)
|
||
# --------------------------------------------------------------------------- #
|
||
_IMAGE_GEN_ATTEMPTS = 3
|
||
|
||
|
||
def _is_transient_image_error(exc: Exception) -> bool:
|
||
"""判定出图失败是否「瞬时」(值得重试)。瞬时:网络抖动 / 5xx / 429 限流 / 中转站吐空无 media url。
|
||
永久(不重试,立即退费报错):4xx 内容违规 / 参数错(如 invalid_image_file)、配置缺失等。"""
|
||
if isinstance(exc, (requests.ConnectionError, requests.Timeout)):
|
||
return True
|
||
if isinstance(exc, requests.HTTPError):
|
||
code = getattr(getattr(exc, "response", None), "status_code", 0) or 0
|
||
return code >= 500 or code == 429
|
||
# extract_first_media_url 在响应没有 url 时抛 ValueError —— 多为中转站偶发空返回,重试常能拿到
|
||
if isinstance(exc, ValueError) and "media url" in str(exc).lower():
|
||
return True
|
||
return False
|
||
|
||
|
||
def _run_image_with_retry(make, *, attempts: int = _IMAGE_GEN_ATTEMPTS):
|
||
"""跑一次出图(provider 调用 + 取 media url),瞬时错误按退避重试,永久错误立即抛。
|
||
make() 须返回 (response, media)。退避 2s / 4s,总耗时上限 ~6s + 出图本身,worker 内执行对用户无感。"""
|
||
for i in range(attempts):
|
||
try:
|
||
return make()
|
||
except Exception as exc: # noqa: BLE001
|
||
if i == attempts - 1 or not _is_transient_image_error(exc):
|
||
raise
|
||
time.sleep(2 * (i + 1))
|
||
|
||
|
||
def generate_base_asset(*, project, user, kind: str, prompt: str, label: str = "", group_id: str | None = None) -> AITask:
|
||
"""提交基础资产生成(**异步**):Web 请求只建 RESERVED 任务 + 预留额度(秒级),
|
||
慢出图(文生图 / 商品 image_edit)交给 Celery worker(run_base_asset_task)跑。
|
||
|
||
这样 Web 层(gunicorn)不被 ~30s+ 的出图请求占住 → 健康探针不饿死 → 不再"生成几张就整站 502/卡死"。
|
||
返回 RESERVED 的 AITask,前端拿 id 轮询 /api/ai/generate-image/?ids=… 取结果;出图后刷新项目即见新组。"""
|
||
from apps.ai.tasks import generate_base_asset_task
|
||
|
||
model_config = get_default_model(ModelConfig.Capability.IMAGE)
|
||
if model_config is None:
|
||
raise ValueError("no active image model configured")
|
||
provider = get_image_provider(model_config)
|
||
# 商品三视图:有真实商品主图 → 走 image_edit 以主图为参考,锁定包装(品牌字/配色/外形/Logo)一致;
|
||
# 无主图或当前模型不支持 image_edit → 回落纯文生图(仅凭商品名脑补,不保证还原真实包装)。
|
||
product_ref_url = _product_cover_url(project.product) if kind == BaseAssetGroup.Kind.PRODUCT else ""
|
||
use_edit = bool(product_ref_url) and hasattr(provider, "image_edit")
|
||
if use_edit:
|
||
gen_prompt = build_product_triview_prompt_refs(project.product, prompt)
|
||
elif kind == BaseAssetGroup.Kind.PERSON:
|
||
gen_prompt = build_person_frontal_prompt(prompt) # 人物立绘:包成「电商真人模特/正面全身/纯色背景」统一模板
|
||
else:
|
||
# 场景图 = 空镜。不论 admin「提示词」页的 scene 模板写没写,都强制叠加「无真人」硬约束,并显式压过
|
||
# 场景描述里可能出现的人物词(如「主播在客厅」)—— 否则模型会照着画出真人,下游用这张场景图生成视频时
|
||
# 被火山以「输入图像可能包含真实人物」拒绝(ZWQ#13)。已叠加过(含「空镜」)则不重复。
|
||
gen_prompt = render_prompt("scene", "{场景描述}", 场景描述=(prompt or "").strip())
|
||
if "空镜" not in gen_prompt:
|
||
gen_prompt = (
|
||
f"{gen_prompt}。"
|
||
"重要约束:这是一张空镜场景图,画面里绝对不要出现任何人物 / 真人 / 模特 / 人脸 / 人手 / 人影;"
|
||
"即使上文描述里提到了人,也只画对应的环境、空间与陈设,把人物完全省略。干净构图,9:16 竖屏。"
|
||
)
|
||
payload = {
|
||
"model": model_config.name, "endpoint": model_config.endpoint, "prompt": gen_prompt,
|
||
"kind": kind, "label": label or "", "group_id": str(group_id) if group_id else "",
|
||
"use_edit": use_edit, "reference_image": product_ref_url,
|
||
}
|
||
task = create_ai_task(
|
||
project=project,
|
||
user=user,
|
||
task_type={
|
||
BaseAssetGroup.Kind.PRODUCT: AITask.Type.PRODUCT_IMAGE,
|
||
BaseAssetGroup.Kind.PERSON: AITask.Type.PERSON_IMAGE,
|
||
BaseAssetGroup.Kind.SCENE: AITask.Type.SCENE_IMAGE,
|
||
}[kind],
|
||
model_config=model_config,
|
||
request_payload=payload,
|
||
)
|
||
generate_base_asset_task.delay(str(task.id))
|
||
return task
|
||
|
||
|
||
def run_base_asset_task(*, task_id: str) -> None:
|
||
"""Celery worker 内执行基础资产的慢出图:调模型 → 成功落库扣费并归组 / 失败退费。
|
||
幂等:只处理 RESERVED 任务,重复投递不会二次出图、二次扣费。"""
|
||
task = AITask.objects.select_related("team", "created_by", "project", "model_config").filter(id=task_id).first()
|
||
if task is None or task.status != AITask.Status.RESERVED:
|
||
return
|
||
project = task.project
|
||
user = task.created_by
|
||
payload = task.request_payload or {}
|
||
kind = payload.get("kind")
|
||
prompt = str(payload.get("prompt") or "")
|
||
label = str(payload.get("label") or "")
|
||
group_id = payload.get("group_id") or None
|
||
use_edit = bool(payload.get("use_edit"))
|
||
ref_url = str(payload.get("reference_image") or "")
|
||
model_config = task.model_config
|
||
provider = get_image_provider(model_config)
|
||
reservation = task.credit_reservation
|
||
try:
|
||
def _make():
|
||
if use_edit and ref_url:
|
||
# 商品三视图:比例可在 admin 改(默认横)
|
||
size = prompt_ratio_size("product_triview", "1536x1024")
|
||
resp = provider.image_edit(model=model_config.name, prompt=prompt, images=[ref_url], size=size)
|
||
else:
|
||
# 场景默认横、人物立绘默认竖;两者比例都可在 admin 改
|
||
if kind == BaseAssetGroup.Kind.SCENE:
|
||
gen_size = prompt_ratio_size("scene", "1536x1024")
|
||
else:
|
||
gen_size = prompt_ratio_size("person_portrait", "1024x1536")
|
||
resp = provider.image_generation(model=model_config.name, endpoint=model_config.endpoint, prompt=prompt, size=gen_size)
|
||
return resp, provider.extract_first_media_url(resp)
|
||
|
||
# 中转站偶发抖动(网络/5xx/限流/空返回)按退避重试,避免「一次失败就生成不出来」;永久错误仍立即退费
|
||
response, media = _run_image_with_retry(_make)
|
||
with transaction.atomic():
|
||
task.status = AITask.Status.SUCCEEDED
|
||
task.response_payload = response
|
||
task.actual_cost = task.estimated_cost
|
||
task.completed_at = timezone.now()
|
||
task.save(update_fields=["status", "response_payload", "actual_cost", "completed_at", "updated_at"])
|
||
charge_reserved_credit(reservation=reservation, actual_amount=task.actual_cost)
|
||
category = {
|
||
BaseAssetGroup.Kind.PRODUCT: Asset.Category.PRODUCT_IMAGE,
|
||
BaseAssetGroup.Kind.PERSON: Asset.Category.PERSON,
|
||
BaseAssetGroup.Kind.SCENE: Asset.Category.SCENE,
|
||
}[kind]
|
||
# 商品基础资产 = 商品三视图(prompt 即「生成同一件商品的三视图」):名字带「三视图」并打
|
||
# metadata.view=three_view + product_id,让商品库的三视图查询(products is_triview / ?product 过滤)
|
||
# 能认出并回填——否则视频项目生成的商品三视图同步不回商品库,商品库永远显示「尚未生成」(ZWQ#5)。
|
||
is_product = kind == BaseAssetGroup.Kind.PRODUCT
|
||
asset_name = f"{project.name}-商品三视图" if is_product else f"{project.name}-{kind}"
|
||
asset = _store_generated_media(
|
||
team=project.team,
|
||
user=user,
|
||
project=project,
|
||
task=task,
|
||
media=media,
|
||
name=asset_name,
|
||
category=category,
|
||
asset_type=Asset.Type.IMAGE,
|
||
)
|
||
if is_product:
|
||
meta = dict(asset.metadata or {})
|
||
meta["view"] = "three_view"
|
||
if project.product_id:
|
||
meta["product_id"] = str(project.product_id)
|
||
asset.metadata = meta
|
||
asset.save(update_fields=["metadata", "updated_at"])
|
||
# 复用同实体的组:追加候选 + 采用最新(版本=candidate_assets,采用=adopted_asset);无则新建
|
||
group = _find_entity_group(project, kind, label, group_id)
|
||
if group is None:
|
||
group_meta = {"label": label.strip()} if label and label.strip() else {}
|
||
group = BaseAssetGroup.objects.create(project=project, kind=kind, task=task, prompt=prompt, metadata=group_meta)
|
||
elif prompt:
|
||
group.prompt = prompt
|
||
group.candidate_assets.add(asset)
|
||
group.adopted_asset = asset
|
||
group.save(update_fields=["adopted_asset", "prompt", "updated_at"])
|
||
# 含人脸资产(角色 / 场景 / 商品):事务提交后静默送火山审核(best-effort,网络调用放 on_commit 避免占着事务)。
|
||
# 场景 / 商品也送审,因为它们可能出现真人(模特出镜/上身),拿到 remote_id 后视频路才能换 asset:// 引用,
|
||
# 否则传原始直链会被火山判「疑似真人」拒。submit_asset_for_review 自身按 REVIEW_CATEGORIES 兜底,非送审类不会真送。
|
||
from apps.assets.review import submit_asset_for_review
|
||
|
||
transaction.on_commit(lambda a=asset: submit_asset_for_review(a))
|
||
except Exception as exc: # noqa: BLE001 — 失败要退费并把错误记进 AITask 供前端轮询读取;不向上抛(避免 celery 重试二次扣费)
|
||
task.status = AITask.Status.FAILED
|
||
task.error_message = str(exc)
|
||
task.completed_at = timezone.now()
|
||
task.save(update_fields=["status", "error_message", "completed_at", "updated_at"])
|
||
release_credit(reservation=reservation, reason=str(exc))
|
||
|
||
|
||
def generate_person_triview(*, project, user, portrait_asset) -> AITask:
|
||
"""流程步骤4 · 据「某一版立绘资产」生成它配套的三视图(**异步**:image_edit 慢,交给 worker)。
|
||
Web 请求只建 RESERVED 任务 + 预留额度后秒回;worker 内跑 image_edit 并把三视图归组(run_triview_task)。
|
||
三视图与立绘 1:1 绑定:metadata.triview_of=<立绘 asset id>;同一立绘多次=同组追加候选(版本)。"""
|
||
from apps.ai.tasks import generate_triview_task
|
||
from apps.ai.model_library import THREE_VIEW_PROMPT
|
||
|
||
if portrait_asset is None:
|
||
raise ValueError("该立绘尚未生成,无法据它生成三视图")
|
||
asset_key = str(portrait_asset.id)
|
||
model_config = get_default_model(ModelConfig.Capability.IMAGE)
|
||
if model_config is None:
|
||
raise ValueError("no active image model configured")
|
||
provider = get_image_provider(model_config)
|
||
if not hasattr(provider, "image_edit"):
|
||
raise ValueError(f"当前图像模型 {model_config.provider.name}:{model_config.name} 不支持参考图三视图(image_edit)")
|
||
ref_url = _asset_preview_url(portrait_asset)
|
||
# 人物三视图提示词:正文可在 admin「提示词」页改(无占位符)
|
||
tri_prompt = render_prompt("person_triview", THREE_VIEW_PROMPT)
|
||
payload = {"model": model_config.name, "prompt": tri_prompt, "kind": "person", "triview_of": asset_key, "reference_image": ref_url}
|
||
task = create_ai_task(project=project, user=user, task_type=AITask.Type.PERSON_IMAGE, model_config=model_config, request_payload=payload)
|
||
generate_triview_task.delay(str(task.id))
|
||
return task
|
||
|
||
|
||
def run_triview_task(*, task_id: str) -> None:
|
||
"""Celery worker 内执行三视图慢出图(image_edit 以立绘为参考):成功落库扣费并归到立绘的三视图组 / 失败退费。
|
||
幂等:只处理 RESERVED 任务,重复投递不会二次出图、二次扣费。"""
|
||
from apps.ai.model_library import THREE_VIEW_PROMPT
|
||
|
||
task = AITask.objects.select_related("team", "created_by", "project", "model_config").filter(id=task_id).first()
|
||
if task is None or task.status != AITask.Status.RESERVED:
|
||
return
|
||
project = task.project
|
||
user = task.created_by
|
||
payload = task.request_payload or {}
|
||
asset_key = str(payload.get("triview_of") or "")
|
||
ref_url = str(payload.get("reference_image") or "")
|
||
prompt = str(payload.get("prompt") or THREE_VIEW_PROMPT)
|
||
model_config = task.model_config
|
||
provider = get_image_provider(model_config)
|
||
reservation = task.credit_reservation
|
||
try:
|
||
def _make():
|
||
tri_size = prompt_ratio_size("person_triview", "1536x1024")
|
||
resp = provider.image_edit(model=model_config.name, prompt=prompt, images=[ref_url], size=tri_size)
|
||
return resp, provider.extract_first_media_url(resp)
|
||
|
||
# 同基础资产:瞬时抖动重试,永久错误立即退费报错
|
||
response, media = _run_image_with_retry(_make)
|
||
with transaction.atomic():
|
||
task.status = AITask.Status.SUCCEEDED
|
||
task.response_payload = response
|
||
task.actual_cost = task.estimated_cost
|
||
task.completed_at = timezone.now()
|
||
task.save(update_fields=["status", "response_payload", "actual_cost", "completed_at", "updated_at"])
|
||
charge_reserved_credit(reservation=reservation, actual_amount=task.actual_cost)
|
||
asset = _store_generated_media(
|
||
team=project.team, user=user, project=project, task=task, media=media,
|
||
name=f"{project.name}-三视图", category=Asset.Category.TRI_VIEW, asset_type=Asset.Type.IMAGE,
|
||
)
|
||
# 复用该立绘的三视图组(triview_of==立绘asset id):追加候选 + 采用最新
|
||
group = next((g for g in project.base_asset_groups.filter(kind=BaseAssetGroup.Kind.PERSON).order_by("created_at")
|
||
if (g.metadata or {}).get("triview_of") == asset_key), None)
|
||
if group is None:
|
||
group = BaseAssetGroup.objects.create(
|
||
project=project, kind=BaseAssetGroup.Kind.PERSON, task=task, prompt=prompt,
|
||
metadata={"label": "·三视图", "triview_of": asset_key},
|
||
)
|
||
group.candidate_assets.add(asset)
|
||
group.adopted_asset = asset
|
||
group.save(update_fields=["adopted_asset", "updated_at"])
|
||
# B 路(期3):视频流程新生成的「角色」立绘 + 三视图成套 → 自动加入/更新模特库(团队级可复用)。
|
||
# 幂等:同一立绘只建一次 Model;已存在该立绘的 Model(尤其「真人上传」模特,upload 时只建了 portrait、
|
||
# 还没 triview)→ 回写 triview_asset,否则模特库永远显示「无三视图」(ZWQ#7:AI 角色能更新、真人上传不更新)。
|
||
from apps.assets.models import Model as ModelEntity
|
||
|
||
portrait = Asset.objects.filter(id=asset_key).first()
|
||
if portrait is not None:
|
||
existing_model = ModelEntity.objects.filter(portrait_asset=portrait, is_deleted=False).first()
|
||
if existing_model is None:
|
||
ModelEntity.objects.create(
|
||
team=project.team, created_by=user,
|
||
name=(portrait.name or f"{project.name}-角色").split("·")[0].strip()[:255] or "角色",
|
||
source=ModelEntity.Source.AI,
|
||
portrait_asset=portrait, triview_asset=asset,
|
||
metadata={"from_project": str(project.id), "auto_enrolled": True},
|
||
)
|
||
elif existing_model.triview_asset_id != asset.id:
|
||
# 真人上传 / 已有模特补三视图:回写新三视图(应用后模特库即刷新三视图)
|
||
existing_model.triview_asset = asset
|
||
existing_model.save(update_fields=["triview_asset", "updated_at"])
|
||
# 合规(期3):三视图含人脸,视频生成前必须过火山审核 → 事务提交后静默送审(best-effort)
|
||
from apps.assets.review import submit_asset_for_review
|
||
|
||
transaction.on_commit(lambda a=asset: submit_asset_for_review(a))
|
||
except Exception as exc: # noqa: BLE001 — 失败退费 + 错误记进 AITask 供前端轮询读取;不向上抛(避免 celery 重试二次扣费)
|
||
task.status = AITask.Status.FAILED
|
||
task.error_message = str(exc)
|
||
task.completed_at = timezone.now()
|
||
task.save(update_fields=["status", "error_message", "completed_at", "updated_at"])
|
||
release_credit(reservation=reservation, reason=str(exc))
|
||
|
||
|
||
def _scene_context(project) -> str:
|
||
"""从商品 + 已采用基础资产提炼一句「风格锚点」,贯穿故事板 / 视频,保证各镜内容一致。"""
|
||
product = project.product
|
||
parts = [f"商品:{product.title}"]
|
||
if product.brand:
|
||
parts.append(f"品牌:{product.brand}")
|
||
if product.category:
|
||
parts.append(f"类目:{product.category}")
|
||
if getattr(product, "target_audience", ""):
|
||
parts.append(f"人群:{product.target_audience}")
|
||
adopted_kinds = set(
|
||
project.base_asset_groups.filter(adopted_asset__isnull=False).values_list("kind", flat=True)
|
||
)
|
||
if BaseAssetGroup.Kind.PERSON in adopted_kinds:
|
||
parts.append("真人出镜,保持人物一致")
|
||
if BaseAssetGroup.Kind.SCENE in adopted_kinds:
|
||
parts.append("统一场景与色调")
|
||
return " · ".join(parts)
|
||
|
||
|
||
def _segment_script_text(segment, entities=None, *, with_dialogue: bool = False, with_exposure: bool = True) -> str:
|
||
"""本镜脚本文本(画面 + 台词/口播 + 商品露出),拼进故事板/视频提示词的【分镜脚本】。
|
||
with_dialogue:有结构化对白时按「说话人：台词」原样放(说话人 id 用 entities 解析成名字),回退扁平 narration。
|
||
with_exposure:是否带「商品露出:…」一行。"""
|
||
parts = []
|
||
visual = (segment.visual_prompt or "").strip()
|
||
if visual:
|
||
parts.append(f"画面:{visual}")
|
||
dialogue = getattr(segment, "dialogue", None) if with_dialogue else None
|
||
name_by_id = {e.get("id"): (e.get("name") or "").strip() for e in (entities or []) if isinstance(e, dict)}
|
||
lines = []
|
||
if isinstance(dialogue, list):
|
||
for d in dialogue:
|
||
if not isinstance(d, dict):
|
||
continue
|
||
line = (d.get("line") or "").strip()
|
||
if not line:
|
||
continue
|
||
spk = name_by_id.get(d.get("speaker")) or ""
|
||
lines.append(f"{spk}：{line}" if spk else line)
|
||
if lines: # 有对白 → 用带说话人的台词
|
||
parts.append("台词:\n" + "\n".join(lines))
|
||
else: # 无对白 → 回退扁平口播/旁白
|
||
narration = (segment.narration or "").strip()
|
||
if narration:
|
||
parts.append(f"口播/旁白:{narration}")
|
||
if with_exposure and getattr(segment, "product_exposure", ""):
|
||
parts.append(f"商品露出:{segment.product_exposure.strip()}")
|
||
return "\n".join(parts)
|
||
|
||
|
||
def build_storyboard_frame_prompt(project, segment, extra_prompt: str = "") -> str:
|
||
"""单镜导演故事板提示词(无参考图时的文本版)。与 refs 版共用 admin「提示词·分镜图」模板(设定为空)。
|
||
extra_prompt = 整张风格提示词(原 StoryboardVersion.prompt,现存项目级)。"""
|
||
dur = segment.duration_seconds or 15
|
||
default = (
|
||
"{设定}根据以下脚本生成一个导演故事板，用于指导 seedance 的视频生成。\n{场景上下文}\n"
|
||
"【分镜脚本】(本段时长约 {时长} 秒)\n{脚本}\n"
|
||
"请严格保持各参考图中角色的同一张脸、同一商品的外观与配色；"
|
||
"电商竖屏 9:16 导演故事板，一镜一图，画面清晰，可直接指导视频生成。{补充}"
|
||
)
|
||
rendered = render_prompt(
|
||
"storyboard_frame", default,
|
||
设定="",
|
||
场景上下文=_scene_context(project),
|
||
时长=dur,
|
||
脚本=(_segment_script_text(segment) or f"第 {segment.sort_order + 1} 镜"),
|
||
补充=(("\n" + extra_prompt.strip()) if extra_prompt else ""),
|
||
)
|
||
return "\n".join(line for line in rendered.split("\n") if line.strip())
|
||
|
||
|
||
def build_video_segment_prompt(project, video_segment, scene, refs, user_prompt: str = "") -> str:
|
||
"""单段视频提示词(用户钦定 @图N 格式):
|
||
【设定】@图N 点名 角色/场景/商品;【分镜】根据 @图(分镜图) 生成;【脚本】本镜脚本。
|
||
refs 顺序与传给 seedance 的 reference_images 一致(@图N 对齐不错位)。"""
|
||
product_name = (getattr(project.product, "title", "") or "商品").strip()
|
||
setup_parts = []
|
||
storyboard_idx = None
|
||
for i, r in enumerate(refs or []):
|
||
n = i + 1
|
||
if r.get("type") == "storyboard":
|
||
storyboard_idx = n
|
||
else:
|
||
setup_parts.append(f"@图{n}是{r.get('label') or ''}({_ENTITY_TYPE_CN.get(r.get('type'), '参考')})")
|
||
# 视频脚本:用带说话人的台词(原样照脚本),不带「商品露出」那行(露出靠模板尾句统一要求)
|
||
entities = (project.metadata or {}).get("script_entities", [])
|
||
script_text = _segment_script_text(scene, entities, with_dialogue=True, with_exposure=False) if scene is not None else ""
|
||
extra = (user_prompt or "").strip()
|
||
if extra:
|
||
script_text = (script_text + "\n" + extra).strip() if script_text else extra
|
||
# 正文可在 admin「提示词·视频」页改。占位符:{设定}=@图N点名、{分镜}=分镜图引用、{脚本}=本段脚本、{时长}。
|
||
# 时长/比例靠 API 参数传(duration/ratio),不写进正文;尾句给风格 + 音效/字幕要求。
|
||
default = "{设定}{分镜}【脚本】{脚本}\n电商带货短视频，商品露出清晰，节奏有转化感。不要字幕，不要背景音乐，但是要有音效，逼真的音效。"
|
||
rendered = render_prompt(
|
||
"video_segment", default,
|
||
设定=("【设定】" + "，".join(setup_parts) + "。\n" if setup_parts else ""),
|
||
分镜=(f"【分镜】根据@图{storyboard_idx}分镜图生成「{product_name}」短视频。\n" if storyboard_idx is not None else ""),
|
||
脚本=(script_text or f"第 {video_segment.sort_order + 1} 段"),
|
||
时长=video_segment.target_duration_seconds,
|
||
)
|
||
return "\n".join(line for line in rendered.split("\n") if line.strip())
|
||
|
||
|
||
def ensure_storyboard_shots(project) -> list:
|
||
"""确保「采用版分镜数」= StoryboardShot 数:每镜一个 shot(对标视频段)。
|
||
缺则按 sort_order 建、补绑 script_segment;多出且从没出过图的尾 shot 裁掉(已出图的不动)。"""
|
||
adopted_script = project.script_versions.filter(is_adopted=True).prefetch_related("segments").first()
|
||
if adopted_script is None:
|
||
return []
|
||
segs = list(adopted_script.segments.order_by("sort_order"))
|
||
shots = {s.sort_order: s for s in project.storyboard_shots.all()}
|
||
for seg in segs:
|
||
shot = shots.get(seg.sort_order)
|
||
if shot is None:
|
||
shots[seg.sort_order] = StoryboardShot.objects.create(
|
||
project=project, script_segment=seg, sort_order=seg.sort_order, prompt=seg.visual_prompt or "")
|
||
elif shot.script_segment_id != seg.id:
|
||
shot.script_segment = seg
|
||
shot.save(update_fields=["script_segment", "updated_at"])
|
||
target = len(segs)
|
||
for order, shot in list(shots.items()):
|
||
if order >= target and not shot.versions.exists():
|
||
shot.delete()
|
||
del shots[order]
|
||
return [shots[k] for k in sorted(shots)]
|
||
|
||
|
||
def submit_storyboard(*, project, user, prompt: str = "", shot_ids: list | None = None) -> list:
|
||
"""故事板·提交(分镜制,对标视频「开始生成/单段重跑」):确保每镜一个 shot,把目标场标 QUEUED,
|
||
真正出图由 poll_storyboard 起后台线程。返回受影响的 shots。
|
||
- shot_ids 给定 → 只(重)生成这些场(单场重跑);
|
||
- 不给 → 还没出片的场;若全部已出片 → 视为整批重跑,全部重出。
|
||
prompt = 整张风格提示词,存项目级 metadata,供每场出图带上(原 StoryboardVersion.prompt 的去处)。"""
|
||
adopted_script = project.script_versions.filter(is_adopted=True).first()
|
||
if adopted_script is None:
|
||
raise ValueError("script must be adopted before generating storyboard")
|
||
if get_default_model(ModelConfig.Capability.IMAGE) is None:
|
||
raise ValueError("no active image model configured")
|
||
if prompt:
|
||
meta = dict(project.metadata or {})
|
||
if meta.get("storyboard_prompt") != prompt:
|
||
meta["storyboard_prompt"] = prompt
|
||
project.metadata = meta
|
||
project.save(update_fields=["metadata", "updated_at"])
|
||
shots = ensure_storyboard_shots(project)
|
||
if shot_ids is not None:
|
||
want = {str(x) for x in shot_ids}
|
||
targets = [s for s in shots if str(s.id) in want]
|
||
else:
|
||
pending = [s for s in shots if s.status != StoryboardShot.Status.SUCCEEDED or s.adopted_version_id is None]
|
||
targets = pending if pending else shots # 全部已出片 = 整批重跑
|
||
for s in targets:
|
||
s.status = StoryboardShot.Status.QUEUED
|
||
s.error_message = ""
|
||
s.save(update_fields=["status", "error_message", "updated_at"])
|
||
return targets
|
||
|
||
|
||
_ENTITY_TYPE_CN = {"character": "角色", "scene": "场景", "product": "商品"}
|
||
|
||
|
||
def _product_reference_image(project, groups: list | None = None) -> dict | None:
|
||
"""商品参考图:优先已采用的商品三视图(product 基础资产组)→ 否则商品真实主图。无图返回 None。
|
||
(商品是预创建的真实商品,不从脚本提取;每镜参考图都无条件带上它。)"""
|
||
if groups is None:
|
||
groups = list(
|
||
project.base_asset_groups.filter(adopted_asset__isnull=False).select_related("adopted_asset")
|
||
)
|
||
pg = next(
|
||
(g for g in groups if g.kind == BaseAssetGroup.Kind.PRODUCT and g.adopted_asset_id),
|
||
None,
|
||
)
|
||
if pg is not None:
|
||
url = _asset_preview_url(pg.adopted_asset)
|
||
if url:
|
||
return {"url": url, "label": "商品", "type": "product",
|
||
"review_status": pg.adopted_asset.review_status, "review_remote_id": pg.adopted_asset.review_remote_id}
|
||
cover = _product_cover_url(project.product)
|
||
if cover:
|
||
return {"url": cover, "label": "商品", "type": "product"}
|
||
return None
|
||
|
||
|
||
def _storyboard_reference_images(project, segment) -> list[dict]:
|
||
"""按本镜 entity_refs 取参考图(角色 / 场景 已采用基础资产)+ **无条件带上商品参考图**,
|
||
供 gpt-image-2 多图合成 @图N。返回 [{url,label,type}],最多 4 张(角色/场景 ≤3 + 商品 1)。
|
||
商品不靠 entity_refs(预创建真实商品、不从脚本提),统一用商品三视图 / 主图带上,从根上保证
|
||
商品参考永不缺失。依赖提取步落进 metadata 的 script_entities。"""
|
||
entities = {
|
||
e.get("id"): e
|
||
for e in (project.metadata or {}).get("script_entities", [])
|
||
if isinstance(e, dict)
|
||
}
|
||
kind_by_type = {
|
||
"character": BaseAssetGroup.Kind.PERSON,
|
||
"scene": BaseAssetGroup.Kind.SCENE,
|
||
}
|
||
groups = list(project.base_asset_groups.filter(adopted_asset__isnull=False).select_related("adopted_asset"))
|
||
# 下游只取「已采用」的角色/场景组(C/D):metadata.adopt 显性优先,否则按是否在脚本里推导;
|
||
# 商品组永远保留(不受采用态影响)。顺带把不在脚本里的(含三视图组 label「·三视图」)排除,
|
||
# 避免兜底误抓未采用/三视图组。
|
||
_entity_names = {(e.get("name") or "").strip() for e in entities.values() if (e.get("name") or "").strip()}
|
||
|
||
def _group_adopted(g) -> bool:
|
||
if g.kind == BaseAssetGroup.Kind.PRODUCT:
|
||
return True
|
||
a = (g.metadata or {}).get("adopt")
|
||
if a == "adopted":
|
||
return True
|
||
if a == "unadopted":
|
||
return False
|
||
return (g.metadata or {}).get("label", "").strip() in _entity_names
|
||
|
||
groups = [g for g in groups if _group_adopted(g)]
|
||
out: list[dict] = []
|
||
used: set = set()
|
||
for rid in (segment.entity_refs or []):
|
||
ent = entities.get(rid)
|
||
if not ent:
|
||
continue
|
||
kind = kind_by_type.get(ent.get("type"))
|
||
if kind is None: # 商品 / 未知类型不在此处理,商品统一在末尾无条件带上
|
||
continue
|
||
name = (ent.get("name") or "").strip()
|
||
match = next(
|
||
(g for g in groups if g.kind == kind and (g.metadata or {}).get("label", "").strip() == name and g.id not in used),
|
||
None,
|
||
) or next((g for g in groups if g.kind == kind and g.id not in used), None)
|
||
if match:
|
||
used.add(match.id)
|
||
url = _asset_preview_url(match.adopted_asset)
|
||
if url:
|
||
# 带上审核态/素材库 ID/资产 ID:视频路按需换成 asset:// 引用、过审闸据此定位送审;图像生成路只读 url 忽略它们
|
||
out.append({"url": url, "label": name or _ENTITY_TYPE_CN.get(ent.get("type"), "参考"), "type": ent.get("type"),
|
||
"asset_id": match.adopted_asset.id,
|
||
"review_status": match.adopted_asset.review_status, "review_remote_id": match.adopted_asset.review_remote_id})
|
||
if len(out) >= 3: # 给商品留一个位置(总计最多 4 张)
|
||
break
|
||
# 商品参考图永远带上:优先已采用商品三视图组 → 否则真实主图
|
||
product_ref = _product_reference_image(project, groups)
|
||
if product_ref is not None and product_ref["url"] not in {r["url"] for r in out}:
|
||
out.append(product_ref)
|
||
# 规范 @图N 顺序:角色 → 商品 → 场景(与下游 image_edit 传图顺序一致,标注不错位)
|
||
_ord = {"character": 0, "product": 1, "scene": 2}
|
||
out.sort(key=lambda r: _ord.get(r.get("type"), 9))
|
||
return out[:4]
|
||
|
||
|
||
def build_storyboard_frame_prompt_refs(project, segment, refs: list[dict], extra_prompt: str = "") -> str:
|
||
"""参考图合成版(用户钦定格式):顶部 @图N 点名每张参考图(角色/商品/场景),再给导演故事板指令 + 分镜脚本。
|
||
@图N 顺序与传给 gpt-image-2 的参考图顺序严格一致(refs 即 image_edit 的 images 顺序)。"""
|
||
if not refs:
|
||
return build_storyboard_frame_prompt(project, segment, extra_prompt)
|
||
setup = "，".join(
|
||
f"@图{i + 1}是{r['label']}({_ENTITY_TYPE_CN.get(r.get('type'), '参考')})" for i, r in enumerate(refs)
|
||
)
|
||
dur = segment.duration_seconds or 15
|
||
# 正文可在 admin「提示词」页改。占位符:{设定}=@图N点名、{场景上下文}、{时长}、{脚本}=本镜脚本、{补充}=本版附加。
|
||
default = (
|
||
"{设定}根据以下脚本生成一个导演故事板，用于指导 seedance 的视频生成。\n{场景上下文}\n"
|
||
"【分镜脚本】(本段时长约 {时长} 秒)\n{脚本}\n"
|
||
"请严格保持各参考图中角色的同一张脸、同一商品的外观与配色；"
|
||
"电商竖屏 9:16 导演故事板，一镜一图，画面清晰，可直接指导视频生成。{补充}"
|
||
)
|
||
rendered = render_prompt(
|
||
"storyboard_frame", default,
|
||
设定=(f"【设定】{setup}。\n" if setup else ""),
|
||
场景上下文=_scene_context(project),
|
||
时长=dur,
|
||
脚本=(_segment_script_text(segment) or f"第 {segment.sort_order + 1} 镜"),
|
||
补充=(("\n" + extra_prompt.strip()) if extra_prompt else ""),
|
||
)
|
||
return "\n".join(line for line in rendered.split("\n") if line.strip())
|
||
|
||
|
||
def _is_transient_error(exc: Exception) -> bool:
|
||
"""网络抖动/超时类瞬时错误(可重试),区别于内容审核拦截、参数非法等确定性失败。
|
||
中转站(tokenssr 等)偶发 Read timeout / 连接重置会无谓掐掉单帧,这类才重试。
|
||
★ 不重试 400:实测故事板的 400 主因是 gpt-image-2 内容审核拦截(moderation_blocked,
|
||
safety_violations=[sexual],如脚本写到「蕾丝胸罩/抚摸/贴身」)—— 同一画面/文案重试还是被拦,
|
||
重试只会拖延报错、空耗调用。这类应「快速失败 + 友好提示」让用户改提示词,而非闷头重试。"""
|
||
msg = str(exc).lower()
|
||
return any(
|
||
k in msg
|
||
for k in ("timed out", "timeout", "connection", "reset by peer", "temporarily",
|
||
"bad gateway", "502", "503", "504", "remotedisconnected", "max retries")
|
||
)
|
||
|
||
|
||
def friendly_generation_error(raw: str) -> str:
|
||
"""把模型/中转站的原始报错翻成给用户的中文友好提示(前端直接展示)。原始报错仍记进 AITask 供排查。
|
||
覆盖:内容审核拦截 / 超时 / 限流 / 凭证 / 参考图被拒 等;命不中给通用兜底。"""
|
||
s = (raw or "").lower()
|
||
if any(k in s for k in ("moderation_blocked", "safety system", "safety_violation", "content_policy", "content policy")):
|
||
return "画面或文案被内容审核拦截(疑似敏感内容)。请调整脚本措辞(如避免「胸罩 / 内衣 / 抚摸 / 贴身」等直白表述,改用「产品 / 包装展示」),或更换参考图后重试。"
|
||
if any(k in s for k in ("timed out", "timeout", "read timed out")):
|
||
return "生成超时,可能是网络波动或模型繁忙,请稍后重试。"
|
||
if any(k in s for k in ("429", "too many requests", "rate limit", "forbidden", "403")):
|
||
return "请求过于频繁,模型暂时限流,请稍候片刻再重试。"
|
||
if any(k in s for k in ("api_key", "unauthorized", "401")):
|
||
return "图像服务凭证异常,请联系管理员处理。"
|
||
if any(k in s for k in ("invalid_image", "invalid image", "image_file")):
|
||
return "参考图未被模型接受(可能格式/尺寸问题),请更换参考图后重试。"
|
||
if any(k in s for k in ("400", "bad request")):
|
||
return "生成请求被模型拒绝,请调整提示词或更换参考图后重试。"
|
||
return "生成失败,请重试;若多次失败请联系技术支持。"
|
||
|
||
|
||
def _call_image_with_retry(fn, *, attempts: int = 2, base_delay: float = 2.0):
|
||
"""对一次出图网络调用做有界重试:仅瞬时错误重试(指数退避),确定性失败立即抛出。
|
||
出图无副作用(失败=没拿到图),重试安全;成功一次即返回。
|
||
attempts 默认 2(一次重试):单次 HTTP 超时上限 300s,2 次≈10min,须 < poll 的「在途锁过期窗口」
|
||
(STORYBOARD_INFLIGHT_STALE_MINUTES)否则 worker 重试期间任务被判僵尸 → 重复起线程 + 重复扣费。"""
|
||
import time
|
||
|
||
last: Exception | None = None
|
||
for i in range(attempts):
|
||
try:
|
||
return fn()
|
||
except Exception as exc: # noqa: BLE001
|
||
last = exc
|
||
if i == attempts - 1 or not _is_transient_error(exc):
|
||
raise
|
||
time.sleep(base_delay * (i + 1))
|
||
raise last # 理论不可达(循环内已 return/raise)
|
||
|
||
|
||
def _storyboard_shot_worker(task_id, shot_id, user_id) -> None:
|
||
"""后台线程:为一个 StoryboardShot 出一张图 → 落成一条 StoryboardShotVersion 并采用。HTTP 永远秒回。"""
|
||
from django.db import connections
|
||
|
||
from apps.accounts.models import User
|
||
|
||
try:
|
||
task = AITask.objects.select_related("model_config__provider").get(id=task_id)
|
||
shot = StoryboardShot.objects.select_related("project__team", "script_segment").get(id=shot_id)
|
||
user = User.objects.get(id=user_id)
|
||
project = shot.project
|
||
segment = shot.script_segment
|
||
model_config = task.model_config
|
||
reservation = task.credit_reservation
|
||
extra_prompt = (project.metadata or {}).get("storyboard_prompt", "") or ""
|
||
task.status = AITask.Status.SUBMITTED
|
||
task.save(update_fields=["status", "updated_at"])
|
||
try:
|
||
provider = get_image_provider(model_config)
|
||
refs = _storyboard_reference_images(project, segment) if segment is not None else []
|
||
ref_urls = [r["url"] for r in refs]
|
||
if ref_urls and hasattr(provider, "image_edit"):
|
||
# gpt-image-2 多图参考:必须用 refs 版提示词(点名「参考图N=角色/场景/商品」+锁脸锁商品)
|
||
frame_prompt = build_storyboard_frame_prompt_refs(project, segment, refs, extra_prompt)
|
||
response = _call_image_with_retry(
|
||
lambda: provider.image_edit(model=model_config.name, prompt=frame_prompt, images=ref_urls, size="1024x1536")
|
||
)
|
||
else:
|
||
frame_prompt = (
|
||
build_storyboard_frame_prompt(project, segment, extra_prompt) if segment is not None
|
||
else (task.request_payload.get("prompt") or "")
|
||
)
|
||
response = _call_image_with_retry(
|
||
lambda: provider.image_generation(model=model_config.name, endpoint=model_config.endpoint, prompt=frame_prompt)
|
||
)
|
||
media = provider.extract_first_media_url(response)
|
||
asset = _store_generated_media(
|
||
team=project.team, user=user, project=project, task=task, media=media,
|
||
name=f"{project.name}-storyboard-{shot.sort_order + 1}",
|
||
category=Asset.Category.STORYBOARD, asset_type=Asset.Type.IMAGE,
|
||
)
|
||
with transaction.atomic():
|
||
task.status = AITask.Status.SUCCEEDED
|
||
task.response_payload = response
|
||
task.actual_cost = task.estimated_cost
|
||
task.completed_at = timezone.now()
|
||
task.save(update_fields=["status", "response_payload", "actual_cost", "completed_at", "updated_at"])
|
||
charge_reserved_credit(reservation=reservation, actual_amount=task.actual_cost)
|
||
# 落一条新版本并采用(反采用同 shot 其余版本)= 该场历史 +1,采用最新
|
||
version = StoryboardShotVersion.objects.create(
|
||
shot=shot, task=task, asset=asset,
|
||
prompt=(segment.visual_prompt if segment is not None else ""), is_adopted=True,
|
||
)
|
||
shot.versions.exclude(id=version.id).update(is_adopted=False)
|
||
StoryboardShot.objects.filter(id=shot.id).update(
|
||
adopted_version=version, status=StoryboardShot.Status.SUCCEEDED, error_message="", updated_at=timezone.now())
|
||
# 合规:分镜图含人脸,视频生成前必须过火山审核 → 事务提交后静默送审(best-effort)
|
||
from apps.assets.review import submit_asset_for_review
|
||
|
||
transaction.on_commit(lambda a=asset: submit_asset_for_review(a))
|
||
except Exception as exc: # noqa: BLE001 — 失败回滚额度,标记任务+shot 失败供 poll 上报
|
||
raw = str(exc)
|
||
hint = friendly_generation_error(raw) # 翻成前端友好提示;原始报错(含 yunqi body)留在 task 供排查
|
||
task.status = AITask.Status.FAILED
|
||
task.error_message = raw[:2000]
|
||
task.completed_at = timezone.now()
|
||
task.save(update_fields=["status", "error_message", "completed_at", "updated_at"])
|
||
release_credit(reservation=reservation, reason=raw[:200])
|
||
# 重跑失败时保留旧 adopted_version(画面不丢),只把状态标 FAILED + 友好提示供前端显示
|
||
StoryboardShot.objects.filter(id=shot.id).update(
|
||
status=StoryboardShot.Status.FAILED, error_message=hint, updated_at=timezone.now())
|
||
finally:
|
||
connections.close_all() # 释放该线程的 DB 连接
|
||
|
||
|
||
def poll_storyboard(*, project, user) -> dict:
|
||
"""异步故事板·轮询(秒回):为 QUEUED/在制 shot 起线程出图;报告进度。永不阻塞在 ARK 调用上。
|
||
返回 {status: generating|succeeded|failed, done, total}。done = 已出片(有采用版且 SUCCEEDED)的场数。"""
|
||
import threading
|
||
|
||
from django.conf import settings as dj_settings
|
||
|
||
shots = list(project.storyboard_shots.select_related("script_segment").order_by("sort_order"))
|
||
if not shots:
|
||
return {"status": "succeeded", "done": 0, "total": 0}
|
||
total = len(shots)
|
||
done = sum(1 for s in shots if s.adopted_version_id is not None and s.status == StoryboardShot.Status.SUCCEEDED)
|
||
active = [s for s in shots if s.status in (StoryboardShot.Status.QUEUED, StoryboardShot.Status.RUNNING)]
|
||
if not active:
|
||
failed = [s for s in shots if s.status == StoryboardShot.Status.FAILED]
|
||
if failed:
|
||
return {"status": "failed", "done": done, "total": total, "error": failed[0].error_message or "storyboard shot failed"}
|
||
return {"status": "succeeded", "done": done, "total": total}
|
||
|
||
# 每镜独立「占位锁」:近 N 分钟内有 CREATED/RESERVED/SUBMITTED 任务的 shot = 在生成中(僵尸超时后释放)。
|
||
STORYBOARD_MAX_PARALLEL = int(getattr(dj_settings, "STORYBOARD_MAX_PARALLEL", 4))
|
||
stale_minutes = int(getattr(dj_settings, "STORYBOARD_INFLIGHT_STALE_MINUTES", 12))
|
||
stale_cutoff = timezone.now() - timedelta(minutes=stale_minutes)
|
||
inflight_shot_ids = {
|
||
str(v)
|
||
for v in AITask.objects.filter(
|
||
project=project, task_type=AITask.Type.STORYBOARD,
|
||
status__in=[AITask.Status.CREATED, AITask.Status.RESERVED, AITask.Status.SUBMITTED],
|
||
created_at__gte=stale_cutoff,
|
||
).values_list("request_payload__storyboard_shot", flat=True)
|
||
if v
|
||
}
|
||
model_config = get_default_model(ModelConfig.Capability.IMAGE)
|
||
extra_prompt = (project.metadata or {}).get("storyboard_prompt", "") or ""
|
||
spawnable = [s for s in active if str(s.id) not in inflight_shot_ids]
|
||
slots = max(0, STORYBOARD_MAX_PARALLEL - len(inflight_shot_ids))
|
||
for shot in spawnable[:slots]:
|
||
segment = shot.script_segment
|
||
task = create_ai_task(
|
||
project=project, user=user, task_type=AITask.Type.STORYBOARD, model_config=model_config,
|
||
request_payload={
|
||
"model": model_config.name, "endpoint": model_config.endpoint,
|
||
"prompt": build_storyboard_frame_prompt(project, segment, extra_prompt) if segment is not None else "",
|
||
"storyboard_shot": str(shot.id),
|
||
},
|
||
)
|
||
StoryboardShot.objects.filter(id=shot.id).update(status=StoryboardShot.Status.RUNNING, updated_at=timezone.now())
|
||
threading.Thread(
|
||
target=_storyboard_shot_worker, args=(str(task.id), str(shot.id), str(user.id)), daemon=True
|
||
).start()
|
||
return {"status": "generating", "done": done, "total": total}
|
||
|
||
|
||
def adopt_storyboard_shot_version(*, shot: StoryboardShot, version: StoryboardShotVersion) -> None:
|
||
"""采用某场的某个历史版本(对标 adopt-video-version):反采用同场其余版本,置该场为采用版+SUCCEEDED。"""
|
||
shot.versions.exclude(id=version.id).update(is_adopted=False)
|
||
if not version.is_adopted:
|
||
version.is_adopted = True
|
||
version.save(update_fields=["is_adopted", "updated_at"])
|
||
shot.adopted_version = version
|
||
shot.status = StoryboardShot.Status.SUCCEEDED
|
||
shot.error_message = ""
|
||
shot.save(update_fields=["adopted_version", "status", "error_message", "updated_at"])
|
||
|
||
|
||
def _asset_preview_url(asset) -> str:
|
||
"""资产主文件的可公开访问 URL(已写绝对 URL 优先,否则实时签 TOS GET)。"""
|
||
if asset is None:
|
||
return ""
|
||
primary = asset.files.filter(is_primary=True).first() or asset.files.first()
|
||
if primary is None:
|
||
return ""
|
||
if primary.preview_url:
|
||
return primary.preview_url
|
||
try:
|
||
return TosStorage().presigned_get_url(object_key=primary.object_key)
|
||
except Exception:
|
||
return ""
|
||
|
||
|
||
def _seedance_ref_url(raw_url: str, review_status: str = "", review_remote_id: str = "") -> str:
|
||
"""Seedance 参考图 URL:只要资产已进素材库(有 remote_id)就用 asset:// 素材库引用 —— 火山认的是
|
||
「在不在素材库」,与审核态无关(实测 processing 也能引用);未进库才回落原始直链。
|
||
火山对写实人脸的原始直链会以 InputImageSensitiveContentDetected 直接拒,必须走同账号素材库引用。
|
||
人物立绘 / 分镜图(含脸)进库走 asset://;场景 / 商品(无脸、未送审)保持直链。
|
||
注意不能只放过 active:故事板刚重生成时分镜图还是 processing,卡 active 会回落直链 → 视频被火山拒。"""
|
||
if review_remote_id:
|
||
return f"asset://{review_remote_id}"
|
||
return raw_url
|
||
|
||
|
||
def _video_reference_images(project, video_segment) -> list[dict]:
|
||
"""视频参考图(带类型,供 @图N):角色/场景/商品 基础资产 + 本镜故事板帧。
|
||
顺序:角色 → 场景 → 商品 → 分镜图(与用户钦定 @图1角色@图2场景@图3商品@图4分镜图 一致)。
|
||
返回 [{url,label,type}];url 对过审人脸资产为 asset:// 素材库引用。都取不到时兜底商品图。"""
|
||
out: list[dict] = []
|
||
scene = None
|
||
adopted_script = project.script_versions.filter(is_adopted=True).prefetch_related("segments").first()
|
||
if adopted_script is not None:
|
||
scene = adopted_script.segments.filter(sort_order=video_segment.sort_order).first()
|
||
if scene is not None:
|
||
refs = list(_storyboard_reference_images(project, scene)) # 角色/商品/场景 实体图
|
||
_vord = {"character": 0, "scene": 1, "product": 2}
|
||
refs.sort(key=lambda r: _vord.get(r.get("type"), 9))
|
||
out = refs
|
||
# 末位追加本镜故事板分镜图(@图N 末位 = 分镜图):取该 sort_order 的 shot 的采用版资产
|
||
shot = (
|
||
project.storyboard_shots.filter(sort_order=video_segment.sort_order, adopted_version__isnull=False).select_related("adopted_version__asset").first()
|
||
or project.storyboard_shots.filter(adopted_version__isnull=False).select_related("adopted_version__asset").order_by("sort_order").first()
|
||
)
|
||
if shot is not None and shot.adopted_version_id:
|
||
frame_asset = shot.adopted_version.asset
|
||
url = _asset_preview_url(frame_asset)
|
||
if url:
|
||
out.append({"url": url, "label": "分镜图", "type": "storyboard",
|
||
"review_status": frame_asset.review_status, "review_remote_id": frame_asset.review_remote_id})
|
||
if not out:
|
||
product_group = (
|
||
project.base_asset_groups.filter(kind=BaseAssetGroup.Kind.PRODUCT, adopted_asset__isnull=False)
|
||
.order_by("-created_at").first()
|
||
)
|
||
if product_group is not None:
|
||
url = _asset_preview_url(product_group.adopted_asset)
|
||
if url:
|
||
out.append({"url": url, "label": "商品", "type": "product"})
|
||
# 写实人脸(已过审)换素材库引用,避免火山「疑似真人」400 拒 → 退文生 → 人物/场景全错
|
||
for r in out:
|
||
r["url"] = _seedance_ref_url(r["url"], r.get("review_status", ""), r.get("review_remote_id", ""))
|
||
return out
|
||
|
||
|
||
def collect_video_review_blockers(project, only_segment: "VideoSegment | None" = None) -> list[dict]:
|
||
"""点「生成视频」前的过审闸:列出本次将生成的段里,含真人脸的参考图(人物立绘 + 该镜故事板分镜)
|
||
中**尚未过审(review_status != 'active')** 的项。火山视频对含真人脸的图,只接受素材库已过审的引用,
|
||
否则报 InputImageSensitiveContentDetected。这里在调火山前先拦,弹窗指明是哪一镜的哪个人物/分镜未过审。
|
||
|
||
only_segment 给定 → 只校验该段(单段重跑);为 None → 校验全部「未出片」段(整批生成)。
|
||
返回 [{video_segment_id, sort_order, scene_no, kind:'person'|'storyboard', name, asset_id, review_status}];
|
||
空列表 = 全部已过审,可放行。"""
|
||
segs = [only_segment] if only_segment is not None else list(project.video_segments.order_by("sort_order"))
|
||
adopted_script = project.script_versions.filter(is_adopted=True).prefetch_related("segments").first()
|
||
shots_by_order = {
|
||
s.sort_order: s
|
||
for s in project.storyboard_shots.filter(adopted_version__isnull=False).select_related("adopted_version__asset")
|
||
}
|
||
blockers: list[dict] = []
|
||
for seg in segs:
|
||
# 整批生成只校验还没出片的段;单段重跑则无论状态都校验(用户主动要重出这一段)
|
||
if only_segment is None and seg.status == VideoSegment.Status.SUCCEEDED:
|
||
continue
|
||
scene_no = seg.sort_order + 1
|
||
scene = adopted_script.segments.filter(sort_order=seg.sort_order).first() if adopted_script else None
|
||
# 人物立绘(与视频实际取图同一套逻辑,保证「拦的」就是「会传给火山的」)
|
||
if scene is not None:
|
||
for ref in _storyboard_reference_images(project, scene):
|
||
if ref.get("type") != "character":
|
||
continue
|
||
if (ref.get("review_status") or "") != "active":
|
||
blockers.append({
|
||
"video_segment_id": str(seg.id), "sort_order": seg.sort_order, "scene_no": scene_no,
|
||
"kind": "person", "name": ref.get("label") or "人物",
|
||
"asset_id": str(ref.get("asset_id") or ""), "review_status": ref.get("review_status") or "",
|
||
})
|
||
# 该镜故事板分镜图(分镜图里也有同一张脸,是另一个需过审的资产)= 对应 shot 的采用版资产
|
||
shot = shots_by_order.get(seg.sort_order)
|
||
if shot is not None and shot.adopted_version_id:
|
||
frame_asset = shot.adopted_version.asset
|
||
if (frame_asset.review_status or "") != "active":
|
||
blockers.append({
|
||
"video_segment_id": str(seg.id), "sort_order": seg.sort_order, "scene_no": scene_no,
|
||
"kind": "storyboard", "name": f"场{scene_no} 分镜图",
|
||
"asset_id": str(frame_asset.id), "review_status": frame_asset.review_status or "",
|
||
})
|
||
return blockers
|
||
|
||
|
||
def submit_video_segment(*, video_segment: VideoSegment, user, prompt: str) -> VideoSegmentVersion | None:
|
||
model_config = get_default_model(ModelConfig.Capability.VIDEO)
|
||
if model_config is None:
|
||
raise ValueError("no active video model configured")
|
||
project = video_segment.project
|
||
|
||
# 衔接:按 sort_order 把视频段绑到对应脚本镜,并织出跟住该镜的提示词。
|
||
scene = None
|
||
adopted_script = project.script_versions.filter(is_adopted=True).prefetch_related("segments").first()
|
||
if adopted_script is not None:
|
||
scene = adopted_script.segments.filter(sort_order=video_segment.sort_order).first()
|
||
if scene is not None and video_segment.script_segment_id != scene.id:
|
||
video_segment.script_segment = scene
|
||
video_segment.save(update_fields=["script_segment", "updated_at"])
|
||
# 参考图(带类型):角色/场景/商品 基础资产 + 本镜故事板帧;@图N 提示词与传图顺序严格对齐。
|
||
refs = _video_reference_images(project, video_segment)
|
||
reference_images = [r["url"] for r in refs]
|
||
final_prompt = build_video_segment_prompt(project, video_segment, scene, refs, prompt)
|
||
|
||
task = create_ai_task(
|
||
project=project,
|
||
user=user,
|
||
task_type=AITask.Type.VIDEO_SEGMENT,
|
||
model_config=model_config,
|
||
request_payload={
|
||
"model": model_config.name,
|
||
"endpoint": model_config.endpoint,
|
||
"prompt": final_prompt,
|
||
"duration": video_segment.target_duration_seconds,
|
||
"ratio": "9:16",
|
||
"video_segment_id": str(video_segment.id),
|
||
"reference_images": reference_images,
|
||
},
|
||
)
|
||
try:
|
||
provider = build_provider(model_config)
|
||
# 不再静默退文生兜底:火山报错(如人脸需走素材库的 InputImageSensitiveContentDetected)直接抛,
|
||
# 由下方 except 落 error_message → 前端如实显示真实报错,绝不假装成功出一条没锁脸的视频。
|
||
response = provider.create_video_task(
|
||
model=model_config.name,
|
||
endpoint=model_config.endpoint,
|
||
prompt=final_prompt,
|
||
duration=video_segment.target_duration_seconds,
|
||
ratio="9:16",
|
||
resolution="720p",
|
||
reference_images=reference_images or None,
|
||
)
|
||
task.provider_task_id = str(response.get("id") or response.get("task_id") or "")
|
||
task.response_payload = response
|
||
task.status = AITask.Status.SUBMITTED
|
||
task.submitted_at = timezone.now()
|
||
task.save(update_fields=["provider_task_id", "response_payload", "status", "submitted_at", "updated_at"])
|
||
video_segment.status = VideoSegment.Status.RUNNING
|
||
video_segment.save(update_fields=["status", "updated_at"])
|
||
return None
|
||
except Exception as exc:
|
||
task.status = AITask.Status.FAILED
|
||
task.error_message = str(exc)
|
||
task.completed_at = timezone.now()
|
||
task.save(update_fields=["status", "error_message", "completed_at", "updated_at"])
|
||
release_credit(reservation=task.credit_reservation, reason=str(exc))
|
||
video_segment.status = VideoSegment.Status.FAILED
|
||
video_segment.error_message = str(exc)
|
||
video_segment.save(update_fields=["status", "error_message", "updated_at"])
|
||
raise
|
||
|
||
|
||
def poll_video_segment(*, video_segment: VideoSegment, user) -> VideoSegmentVersion | None:
|
||
# 幂等:已完成的段直接回采用版;已失败的段不再 poll。避免对已成功 task 再 poll → 二次建版 / 二次扣费。
|
||
if video_segment.status == VideoSegment.Status.SUCCEEDED:
|
||
return video_segment.adopted_version or video_segment.versions.order_by("-created_at").first()
|
||
if video_segment.status == VideoSegment.Status.FAILED:
|
||
return None
|
||
|
||
# ★ 先找「在途任务」再回退旧版本的任务。旧实现反过来:重跑时段上已有(旧)版本,
|
||
# 取到旧版本挂的已成功任务 → 短路返回旧版,在途的新任务永远没人轮询,
|
||
# 段永远卡「生成中」、新视频取不回来(实测重跑卡 40 分钟,ARK 侧其实早已生成完)。
|
||
ai_task = video_segment.project.ai_tasks.filter(
|
||
task_type=AITask.Type.VIDEO_SEGMENT,
|
||
request_payload__video_segment_id=str(video_segment.id),
|
||
status__in=[AITask.Status.SUBMITTED, AITask.Status.POLLING],
|
||
).order_by("-created_at").first()
|
||
if ai_task is None:
|
||
latest_version = video_segment.versions.order_by("-created_at").first()
|
||
ai_task = latest_version.task if latest_version else None
|
||
if ai_task is None:
|
||
raise ValueError("no active video generation task")
|
||
|
||
# task 已终态(可能被并发的 worker / 另一次 poll 处理过):直接回已有版,不再调 ARK。
|
||
if ai_task.status == AITask.Status.SUCCEEDED:
|
||
return video_segment.versions.filter(task=ai_task).order_by("-created_at").first()
|
||
if ai_task.status in (AITask.Status.FAILED, AITask.Status.CANCELLED):
|
||
return None
|
||
|
||
provider = build_provider(ai_task.model_config)
|
||
response = provider.poll_video_task(endpoint=ai_task.model_config.endpoint, provider_task_id=ai_task.provider_task_id)
|
||
remote_status = response.get("status")
|
||
if remote_status in {"queued", "running", "processing"}:
|
||
# 仍在生成:只在状态首次进入 POLLING 时落一次库。旧实现每次 poll(5s 一次)都把完整
|
||
# response JSON 回写远程 MySQL——纯浪费写带宽,终态时反正会存完整 payload。
|
||
if ai_task.status != AITask.Status.POLLING:
|
||
ai_task.status = AITask.Status.POLLING
|
||
ai_task.save(update_fields=["status", "updated_at"])
|
||
return None
|
||
if remote_status in {"failed", "expired", "cancelled"}:
|
||
ai_task.status = AITask.Status.FAILED
|
||
ai_task.response_payload = response
|
||
ai_task.error_message = response.get("error", {}).get("message", "video generation failed")
|
||
ai_task.completed_at = timezone.now()
|
||
ai_task.save(update_fields=["status", "response_payload", "error_message", "completed_at", "updated_at"])
|
||
release_credit(reservation=ai_task.credit_reservation, reason=ai_task.error_message)
|
||
video_segment.status = VideoSegment.Status.FAILED
|
||
video_segment.error_message = ai_task.error_message
|
||
video_segment.save(update_fields=["status", "error_message", "updated_at"])
|
||
return None
|
||
|
||
media = provider.extract_first_media_url(response)
|
||
asset = _store_generated_media(
|
||
team=video_segment.project.team,
|
||
user=user,
|
||
project=video_segment.project,
|
||
task=ai_task,
|
||
media=media,
|
||
name=f"{video_segment.project.name}-segment-{video_segment.sort_order + 1}",
|
||
category=Asset.Category.VIDEO_CLIP,
|
||
asset_type=Asset.Type.VIDEO,
|
||
)
|
||
# 终态化必须持锁原子做:两个并发 poll(前端 5s 静默轮询 × 提交后轮询/worker)同时走到这里时,
|
||
# 旧实现会同 task 建两个版本 + charge_reserved_credit 双扣费(实测 03:08:25 同秒双版本)。
|
||
# select_for_update 锁 task 行,后到者看到 SUCCEEDED 直接回已有版,不再建版/扣费。
|
||
with transaction.atomic():
|
||
locked_task = AITask.objects.select_for_update().get(id=ai_task.id)
|
||
if locked_task.status == AITask.Status.SUCCEEDED:
|
||
existing = video_segment.versions.filter(task=locked_task).order_by("-created_at").first()
|
||
if existing is not None:
|
||
return existing
|
||
locked_task.status = AITask.Status.SUCCEEDED
|
||
locked_task.response_payload = response
|
||
locked_task.actual_cost = locked_task.estimated_cost
|
||
locked_task.completed_at = timezone.now()
|
||
locked_task.save(update_fields=["status", "response_payload", "actual_cost", "completed_at", "updated_at"])
|
||
charge_reserved_credit(reservation=locked_task.credit_reservation, actual_amount=locked_task.actual_cost)
|
||
version = VideoSegmentVersion.objects.create(
|
||
video_segment=video_segment,
|
||
task=locked_task,
|
||
asset=asset,
|
||
prompt=locked_task.request_payload.get("prompt", ""),
|
||
is_adopted=True,
|
||
)
|
||
video_segment.versions.exclude(id=version.id).update(is_adopted=False)
|
||
video_segment.adopted_version = version
|
||
video_segment.status = VideoSegment.Status.SUCCEEDED
|
||
video_segment.error_message = ""
|
||
video_segment.save(update_fields=["adopted_version", "status", "error_message", "updated_at"])
|
||
return version
|
||
|
||
|
||
def create_export_job(*, timeline, user) -> ExportJob:
|
||
return ExportJob.objects.create(timeline=timeline, status=ExportJob.Status.QUEUED)
|
||
|
||
|
||
# 图片趴三类(模特库+资产模型重构 期2):
|
||
# · model + product → model_tryon(模特上身图);model 无 product → person(视频角色「生成演员」,留期3)
|
||
# · cover → platform_kit(平台套图);image → free_create(自由创作)
|
||
_STANDALONE_CATEGORY = {
|
||
"model": Asset.Category.PERSON,
|
||
"cover": Asset.Category.PLATFORM_KIT,
|
||
"image": Asset.Category.FREE_CREATE,
|
||
}
|
||
_STANDALONE_TASK_TYPE = {
|
||
"model": AITask.Type.PERSON_IMAGE,
|
||
"cover": AITask.Type.PRODUCT_IMAGE,
|
||
"image": AITask.Type.PRODUCT_IMAGE,
|
||
}
|
||
|
||
|
||
def _reap_stale_standalone_image_tasks(*, team) -> None:
|
||
"""兜底:worker 崩溃/重启(OOM、部署)可能留下卡在 RESERVED 的出图任务,额度被一直占住、
|
||
前端轮询也永远等不到结果。超过 10 分钟(远大于单张真实出图耗时 ~60s)仍 RESERVED 的判为僵尸:
|
||
标记失败并退还预留额度。趁每次新提交时顺手回收,无需额外的定时任务(与导出僵尸清理同思路)。"""
|
||
cutoff = timezone.now() - timedelta(minutes=10)
|
||
stale = AITask.objects.filter(
|
||
team=team,
|
||
project__isnull=True,
|
||
task_type__in=[AITask.Type.PERSON_IMAGE, AITask.Type.PRODUCT_IMAGE],
|
||
status=AITask.Status.RESERVED,
|
||
updated_at__lt=cutoff,
|
||
)
|
||
for task in stale:
|
||
try:
|
||
with transaction.atomic():
|
||
task.status = AITask.Status.FAILED
|
||
task.error_message = "worker 未在预期时间内完成(僵尸任务自动回收)"
|
||
task.completed_at = timezone.now()
|
||
task.save(update_fields=["status", "error_message", "completed_at", "updated_at"])
|
||
try:
|
||
reservation = task.credit_reservation
|
||
except ObjectDoesNotExist:
|
||
reservation = None
|
||
if reservation is not None:
|
||
release_credit(reservation=reservation, reason="僵尸出图任务自动回收")
|
||
except Exception: # noqa: BLE001 — 单个回收失败不应阻断新任务提交
|
||
continue
|
||
|
||
|
||
def enqueue_standalone_images(*, team, user, prompt: str, mode: str = "image", count: int = 1, product_id: str | None = None, reference_product: bool = False, model_id: str | None = None, model_entity_id: str | None = None, ratio: str | None = None, image_model: str | None = None, conversation=None, reference_image_ids: list[str] | None = None, platform_id: str | None = None) -> list[AITask]:
|
||
"""独立生图(图片创作 / 模特上身图 / 平台套图)改为**异步**:本函数在 Web 请求里只做「建任务 +
|
||
预留额度」这种秒级的活,真正 ~30s 的 ARK 出图交给 Celery worker(generate_standalone_image_task)。
|
||
|
||
这样 Web 层(gunicorn)不会被慢出图请求占住 worker → 健康探针不会被饿死 → 根治"几张图就整站 502"。
|
||
且任务一旦提交(额度已预留),浏览器关掉 / 断网都不影响——worker 照样把图生成并落库,扣费/退费在
|
||
worker 内闭环。返回已 RESERVED 的 AITask 列表,前端拿 id 轮询 GET /api/ai/generate-image/?ids=… 取结果。"""
|
||
from apps.ai.tasks import generate_standalone_image_task
|
||
|
||
_reap_stale_standalone_image_tasks(team=team)
|
||
# 用户在工作室选的生图模型(火山 / gpt-image)优先;未选或解析不到则回落系统默认
|
||
model_config = resolve_image_model(image_model) or get_default_model(ModelConfig.Capability.IMAGE)
|
||
if model_config is None:
|
||
raise ValueError("no active image model configured")
|
||
task_type = _STANDALONE_TASK_TYPE.get(mode, AITask.Type.PRODUCT_IMAGE)
|
||
count = max(1, min(int(count or 1), 12))
|
||
ref_ids = [str(r) for r in (reference_image_ids or []) if r]
|
||
# 带参考图的图片创作必须走 image_edit(gpt-image)才真的还原上传素材——火山 Seedream 的图生图只松散
|
||
# 借色调、不锁主体(与三视图/模特上身图同理,那两条流程也强制 image_edit)。当前模型不支持 image_edit
|
||
# 就自动切到 gpt-image-2,否则用户选了火山却传了参考图 → 生成完全不参考素材(用户实测的根因)。
|
||
if ref_ids and not hasattr(build_provider(model_config), "image_edit"):
|
||
alt = resolve_image_model("gpt-image")
|
||
if alt is not None:
|
||
model_config = alt
|
||
# 本次提交 = 一组(模特上身图组 / 平台套图组):同一 batch_id 串起这批图,前端可成组展示。
|
||
batch_id = str(uuid.uuid4())
|
||
# 平台套图:规范化平台 id(前端 dy/tb… → canonical),用于注入平台版式块(优化版);非 cover 模式忽略。
|
||
platform_key = str(platform_id or "").strip() if mode == "cover" else ""
|
||
platform_name = _PLATFORM_NAMES.get(platform_key, "")
|
||
tasks: list[AITask] = []
|
||
for index in range(count):
|
||
cost = estimate_cost(model_config)
|
||
task = AITask.objects.create(
|
||
team=team,
|
||
created_by=user,
|
||
project=None,
|
||
conversation=conversation,
|
||
task_type=task_type,
|
||
status=AITask.Status.CREATED,
|
||
model_config=model_config,
|
||
idempotency_key=f"standalone-image:{team.id}:{uuid.uuid4()}",
|
||
request_payload={"model": model_config.name, "endpoint": model_config.endpoint, "prompt": prompt, "mode": mode, "index": index, "product_id": str(product_id) if product_id else None, "reference_product": bool(reference_product), "model_id": str(model_id) if model_id else None, "model_entity_id": str(model_entity_id) if model_entity_id else None, "batch_id": batch_id, "ratio": str(ratio) if ratio else None, "reference_image_ids": ref_ids, "platform_id": platform_key or None, "platform_name": platform_name or None},
|
||
estimated_cost=cost,
|
||
)
|
||
# 预留额度若余额不足会抛 ValueError,在同步的 Web 请求里立刻反馈给前端(不会先建半套任务)
|
||
reserve_credit(team=team, user=user, task=task, amount=cost)
|
||
task.status = AITask.Status.RESERVED
|
||
task.save(update_fields=["status", "updated_at"])
|
||
tasks.append(task)
|
||
# 额度都预留成功后再统一派发,避免"派发了任务但后面某张预留失败"的半成品状态
|
||
for task in tasks:
|
||
generate_standalone_image_task.delay(str(task.id))
|
||
return tasks
|
||
|
||
|
||
def run_standalone_image_task(*, task_id: str) -> None:
|
||
"""Celery worker 内执行**单张**图的慢活:调 ARK → 成功落库扣费 / 失败退费。
|
||
幂等:只处理 RESERVED 状态的任务,重复投递(celery retry / 重启重放)不会二次出图、二次扣费。"""
|
||
task = AITask.objects.select_related("team", "created_by", "model_config").filter(id=task_id).first()
|
||
if task is None or task.status != AITask.Status.RESERVED:
|
||
return
|
||
team = task.team
|
||
user = task.created_by
|
||
payload = task.request_payload or {}
|
||
prompt = str(payload.get("prompt") or "")
|
||
mode = str(payload.get("mode") or "image")
|
||
index = int(payload.get("index") or 0)
|
||
product_id = payload.get("product_id") or None
|
||
# 模特上身图(mode=model 且绑了商品)= 图片趴「模特上身图」(引用模特库),归 model_tryon、不送审;
|
||
# 「生成演员」同样走 mode=model 但无 product_id = 视频角色,仍归 person(送审,留期3 收编为「角色」)。
|
||
if mode == "model" and product_id:
|
||
category = Asset.Category.MODEL_TRYON
|
||
else:
|
||
category = _STANDALONE_CATEGORY.get(mode, Asset.Category.UNCATEGORIZED)
|
||
model_config = task.model_config
|
||
provider = get_image_provider(model_config)
|
||
reservation = task.credit_reservation
|
||
# 出图策略(都优先 image_edit 锁真实素材,模型不支持/无素材才回落纯文生图):
|
||
# · 模特上身图(mode=model):参考图1=商品真实主图 + 参考图2=选中模特 → 生成「该模特用该商品」效果图;
|
||
# · 平台套图(mode=cover):参考图1=商品真实主图(+ 有模特则参考图2=模特)→ 锁包装一致性出套图;
|
||
# · 商品三视图(reference_product):参考图1=商品真实主图 → 锁包装一致性;
|
||
# · 其余(图片创作):纯文生图。
|
||
product = None
|
||
if product_id:
|
||
from apps.products.models import Product
|
||
|
||
product = Product.objects.filter(id=product_id).first()
|
||
can_edit = hasattr(provider, "image_edit")
|
||
model_url = ""
|
||
if payload.get("model_id"):
|
||
model_asset = Asset.objects.filter(id=payload.get("model_id")).first()
|
||
if model_asset is not None:
|
||
model_url = _asset_preview_url(model_asset)
|
||
product_url = _product_cover_url(product) if product is not None else ""
|
||
# 模特上身图:真实上传图优先、排除 AI 生成图,可多张(多角度更易锁外形/品牌)
|
||
product_urls = _product_reference_urls(product, limit=3) if product is not None else []
|
||
# 图片创作自由模式:用户上传的参考图(已先传成 Asset)→ 取直链,作多图参考(image_edit / 图生图)
|
||
ref_urls: list[str] = []
|
||
for rid in (payload.get("reference_image_ids") or []):
|
||
ref_asset = Asset.objects.filter(id=rid).first()
|
||
if ref_asset is not None:
|
||
u = _asset_preview_url(ref_asset)
|
||
if u:
|
||
ref_urls.append(u)
|
||
|
||
# 参考图收集与 provider 无关:先把「该用哪些参考图 + 哪条提示词」定下来,再按模型能力选调用方式。
|
||
edit_images: list[str] = []
|
||
edit_prompt = ""
|
||
if mode == "model" and product_urls:
|
||
# 模特上身图:参考图1~N=商品真实图(多角度),参考图N+1=模特(模特图可缺则让模型自取真人模特)
|
||
edit_images = product_urls + ([model_url] if model_url else [])
|
||
edit_prompt = build_model_tryon_prompt_refs(
|
||
product, has_model=bool(model_url), base_prompt=prompt,
|
||
index=index, n_product=len(product_urls),
|
||
)
|
||
elif mode == "cover" and product_urls:
|
||
# 平台套图:参考图1~N=商品真实图(多角度,_product_reference_urls 已优先真实上传图/排除 AI 图),
|
||
# 有模特则参考图N+1=模特(锁人脸/身形)。platform_id 注入平台版式块(优化版)。
|
||
edit_images = product_urls + ([model_url] if model_url else [])
|
||
edit_prompt = build_platform_cover_prompt_refs(
|
||
product,
|
||
has_model=bool(model_url),
|
||
base_prompt=prompt,
|
||
platform_id=str(payload.get("platform_id") or ""),
|
||
index=index,
|
||
product_ref_count=len(product_urls),
|
||
)
|
||
elif bool(payload.get("reference_product")) and product_url:
|
||
edit_images = [product_url]
|
||
edit_prompt = build_product_triview_prompt_refs(product, "")
|
||
elif ref_urls:
|
||
# 图片创作自由模式:以用户上传的参考图为基底出图。提示词必须显式要求「保留参考图主体」,
|
||
# 否则图生图只会松散借个色调、不真的还原上传的素材(=用户反馈的「没参考我的图」)。
|
||
edit_images = ref_urls
|
||
edit_prompt = build_free_reference_prompt(prompt, n_refs=len(ref_urls), index=index)
|
||
use_edit = bool(edit_images)
|
||
try:
|
||
if use_edit and can_edit:
|
||
# gpt-image 等支持 image_edit(多图参考编辑接口)
|
||
if payload.get("reference_product"):
|
||
size = "1536x1024" # 三视图固定横向
|
||
else:
|
||
size = _ratio_to_image_size(str(payload.get("ratio") or "")) # 模特图按选中比例
|
||
response = provider.image_edit(model=model_config.name, prompt=edit_prompt, images=edit_images, size=size)
|
||
elif use_edit:
|
||
# 火山 Seedream 无 image_edit:走 image_generation 带 image=参考图(图生图多参考);
|
||
# 尺寸按选中比例换算成火山可接受的 ~2K 尺寸(三视图固定横向)
|
||
vsize = "2304x1728" if payload.get("reference_product") else _ratio_to_volcano_size(str(payload.get("ratio") or ""))
|
||
response = provider.image_generation(model=model_config.name, endpoint=model_config.endpoint, prompt=edit_prompt, image=edit_images, size=vsize)
|
||
else:
|
||
# 纯文生图同批多张要不同构图,否则雷同(PMC#24);index>0 追加换版式指令。
|
||
gen_prompt = prompt
|
||
if index > 0:
|
||
gen_prompt = f"{prompt}。{_COVER_VARIATIONS[index % len(_COVER_VARIATIONS)]},换一种构图与视角。"
|
||
response = provider.image_generation(model=model_config.name, endpoint=model_config.endpoint, prompt=gen_prompt)
|
||
media = provider.extract_first_media_url(response)
|
||
with transaction.atomic():
|
||
task.status = AITask.Status.SUCCEEDED
|
||
task.response_payload = response
|
||
task.actual_cost = task.estimated_cost
|
||
task.completed_at = timezone.now()
|
||
task.save(update_fields=["status", "response_payload", "actual_cost", "completed_at", "updated_at"])
|
||
charge_reserved_credit(reservation=reservation, actual_amount=task.actual_cost)
|
||
fileobj, content_type = VolcanoArkProvider.media_to_bytes(media)
|
||
suffix = ".jpg" if "jpeg" in content_type else (".webp" if "webp" in content_type else ".png")
|
||
asset_id = uuid.uuid4()
|
||
object_key = f"teams/{team.id}/standalone/{asset_id}{suffix}"
|
||
stored = TosStorage().upload_fileobj(fileobj=fileobj, object_key=object_key, content_type=content_type)
|
||
# 商品三视图(reference_product):与视频项目基础资产阶段生成的商品三视图同款标记 ——
|
||
# category=product_image + 名字含「商品三视图」+ metadata.view=three_view + product_id。
|
||
# 否则商品库自己生成的三视图认不出(无 view 标记),刷新即丢、视频项目第二步也读不到「商品已有三视图」(ZWQ#5 续)。
|
||
is_product_triview = bool(payload.get("reference_product"))
|
||
asset_label = "商品三视图" if is_product_triview else {"model": "模特上身图", "cover": "平台套图", "image": "图片创作"}.get(mode, mode)
|
||
asset_category = Asset.Category.PRODUCT_IMAGE if is_product_triview else category
|
||
# 资产元数据:product_id(商品详情页据此只展示该商品素材)+ batch_id(成组)+ mode + model_entity_id(上身图溯源模特库)
|
||
asset_meta: dict = {"mode": mode}
|
||
if is_product_triview:
|
||
asset_meta["view"] = "three_view"
|
||
if product_id:
|
||
asset_meta["product_id"] = str(product_id)
|
||
if payload.get("batch_id"):
|
||
asset_meta["batch_id"] = str(payload["batch_id"])
|
||
if payload.get("model_entity_id"):
|
||
asset_meta["model_entity_id"] = str(payload["model_entity_id"])
|
||
asset = Asset.objects.create(
|
||
id=asset_id, team=team, created_by=user, name=f"AI 生成 · {asset_label} · {index + 1}",
|
||
asset_type=Asset.Type.IMAGE, source=Asset.Source.AI_GENERATED, category=asset_category, origin_task=task,
|
||
metadata=asset_meta,
|
||
# 工作台生成的图默认不进资产库列表,只在工作台展示;用户「加入资产库」后才置 True
|
||
in_library=False,
|
||
)
|
||
AssetFile.objects.create(asset=asset, object_key=stored.object_key, bucket=stored.bucket, content_type=stored.content_type, size_bytes=stored.size_bytes, is_primary=True)
|
||
except Exception as exc: # noqa: BLE001 — 失败要退费并把错误记进 AITask 供前端轮询读取;不向上抛(避免 celery 重试二次扣费)
|
||
task.status = AITask.Status.FAILED
|
||
task.error_message = str(exc)
|
||
task.completed_at = timezone.now()
|
||
task.save(update_fields=["status", "error_message", "completed_at", "updated_at"])
|
||
release_credit(reservation=reservation, reason=str(exc))
|
||
|
||
|
||
# ── 旁白配音(TTS):每镜旁白合成一段语音,导出时作为人声轨混在 BGM 之上 ──
|
||
|
||
# 音色按「语音合成(经典版)」试用包实测可用清单配置;大模型音色(*_bigtts)需另开通「语音合成大模型」服务,当前账号 403
|
||
VOICEOVER_VOICES = [
|
||
{"key": "BV700_streaming", "label": "灿灿 · 活力女声"},
|
||
{"key": "BV034_streaming", "label": "知性姐姐 · 沉稳女声"},
|
||
{"key": "BV001_streaming", "label": "通用女声"},
|
||
{"key": "BV056_streaming", "label": "阳光男声"},
|
||
{"key": "BV102_streaming", "label": "儒雅青年 · 解说男声"},
|
||
{"key": "BV002_streaming", "label": "通用男声"},
|
||
]
|
||
DEFAULT_VOICEOVER_VOICE = VOICEOVER_VOICES[0]["key"]
|
||
|
||
|
||
def synthesize_project_voiceover(*, project, user, items: list[dict], voice_type: str, speed_ratio: float = 1.0) -> dict:
|
||
"""每镜旁白 → **逐句** TTS 配音资产(一句一段音频,带句内起点 offset_ms),映射写入
|
||
timeline.metadata["voiceover"]。逐句才能支持「拖动字幕块 = 字幕和它的语音一起移动」;
|
||
一次调用 = 一个 AITask = 计一次费;任何一句失败则整体失败并释放预留(不留半套配音)。"""
|
||
from apps.projects.services.export import _split_subtitle_text
|
||
|
||
texts = [] # (片段 index, 句序 cue, 句文本)
|
||
for n, item in enumerate(items or []):
|
||
text = str(item.get("text") or "").strip()
|
||
if not text:
|
||
continue
|
||
idx = int(item.get("index", n))
|
||
pieces = _split_subtitle_text(text) or [text]
|
||
for j, piece in enumerate(pieces):
|
||
texts.append((idx, j, piece))
|
||
if not texts:
|
||
raise ValueError("没有可配音的旁白文本")
|
||
provider = VolcanoTtsProvider()
|
||
if not provider.configured:
|
||
raise TtsNotConfigured(
|
||
"语音合成未配置:请在后端环境变量设置 VOLC_TTS_APPID 和 VOLC_TTS_ACCESS_TOKEN"
|
||
"(火山引擎控制台 → 语音技术 → 语音合成大模型 → 创建应用)"
|
||
)
|
||
voice_type = voice_type or DEFAULT_VOICEOVER_VOICE
|
||
model_config = get_default_model(ModelConfig.Capability.AUDIO)
|
||
if model_config is None:
|
||
raise ValueError("no active audio model configured")
|
||
task = create_ai_task(
|
||
project=project,
|
||
user=user,
|
||
task_type=AITask.Type.VOICEOVER,
|
||
model_config=model_config,
|
||
request_payload={
|
||
"voice_type": voice_type,
|
||
"speed_ratio": float(speed_ratio or 1.0),
|
||
"items": [{"index": idx, "cue": j, "text": text} for idx, j, text in texts],
|
||
},
|
||
)
|
||
reservation = task.credit_reservation
|
||
try:
|
||
synthesized = []
|
||
for idx, j, text in texts:
|
||
audio, duration_ms = provider.synthesize(text=text, voice_type=voice_type, speed_ratio=speed_ratio, uid=str(user.id))
|
||
synthesized.append((idx, j, text, audio, duration_ms))
|
||
with transaction.atomic():
|
||
task.status = AITask.Status.SUCCEEDED
|
||
task.actual_cost = task.estimated_cost
|
||
task.completed_at = timezone.now()
|
||
task.response_payload = {"segments": len(synthesized)}
|
||
task.save(update_fields=["status", "actual_cost", "completed_at", "response_payload", "updated_at"])
|
||
charge_reserved_credit(reservation=reservation, actual_amount=task.actual_cost)
|
||
vo_items = []
|
||
offset_acc: dict[int, int] = {} # 同一片段内逐句顺排:句 j 的默认起点 = 前面句时长之和
|
||
for idx, j, text, audio, duration_ms in synthesized:
|
||
asset_id = uuid.uuid4()
|
||
object_key = f"teams/{project.team_id}/projects/{project.id}/voiceover/{asset_id}.mp3"
|
||
stored = TosStorage().upload_fileobj(fileobj=BytesIO(audio), object_key=object_key, content_type="audio/mpeg")
|
||
asset = Asset.objects.create(
|
||
id=asset_id, team=project.team, created_by=user,
|
||
name=f"配音 · 场 {idx + 1} · 句 {j + 1}", asset_type=Asset.Type.AUDIO,
|
||
source=Asset.Source.AI_GENERATED, category=Asset.Category.UNCATEGORIZED,
|
||
origin_task=task, description=text,
|
||
)
|
||
AssetFile.objects.create(
|
||
asset=asset, object_key=stored.object_key, bucket=stored.bucket,
|
||
content_type=stored.content_type, size_bytes=stored.size_bytes, is_primary=True,
|
||
)
|
||
offset_ms = offset_acc.get(idx, 0)
|
||
offset_acc[idx] = offset_ms + (duration_ms or 0)
|
||
vo_items.append({
|
||
"index": idx, "cue": j, "text": text, "asset": str(asset.id),
|
||
"duration_ms": duration_ms, "offset_ms": offset_ms,
|
||
})
|
||
timeline, _ = Timeline.objects.get_or_create(
|
||
project=project, defaults={"name": f"{project.name} Timeline", "duration_seconds": 60}
|
||
)
|
||
metadata = dict(timeline.metadata or {})
|
||
metadata["voiceover"] = {
|
||
"enabled": True,
|
||
"voice_type": voice_type,
|
||
"speed_ratio": float(speed_ratio or 1.0),
|
||
"items": vo_items,
|
||
}
|
||
timeline.metadata = metadata
|
||
timeline.save(update_fields=["metadata", "updated_at"])
|
||
return metadata["voiceover"]
|
||
except Exception as exc:
|
||
task.status = AITask.Status.FAILED
|
||
task.error_message = str(exc)
|
||
task.completed_at = timezone.now()
|
||
task.save(update_fields=["status", "error_message", "completed_at", "updated_at"])
|
||
release_credit(reservation=reservation, reason=str(exc))
|
||
raise
|