优化脚本
This commit is contained in:
Azmat@qq.com
@@ -404,6 +404,17 @@ def classify_generation_error(
|
||||
return _build_error("invalid_input", operation, reference_id=reference_id)
|
||||
if _contains(compact_code, "postprocesserror"):
|
||||
return _build_error("processing_failed", operation, reference_id=reference_id)
|
||||
if _contains(
|
||||
signal,
|
||||
"没有用上商品卖点原词",
|
||||
"没有出现品牌原词",
|
||||
"必须原词用上",
|
||||
"旁白太短",
|
||||
"画面描写太短",
|
||||
"按秒拆分镜",
|
||||
"切两次景别",
|
||||
):
|
||||
return _build_error("processing_failed", operation, reference_id=reference_id)
|
||||
return _build_error("unknown", operation, reference_id=reference_id)
|
||||
|
||||
|
||||
|
||||
@@ -0,0 +1,47 @@
|
||||
from django.db import migrations
|
||||
|
||||
STORYBOARD = (
|
||||
"{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。"
|
||||
"脚本里的 visual 是本段秒级分镜清单；请画商品用法正确、最能看清商品的那一拍作为关键帧。"
|
||||
"禁止画出违背常识的用法（例如把茶包丢进冷白开、悬浮的手）。{补充}"
|
||||
)
|
||||
STORYBOARD_OLD = (
|
||||
"{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}"
|
||||
)
|
||||
VIDEO = (
|
||||
"{设定}{分镜}【脚本】{脚本}\n"
|
||||
"严格按照脚本里的秒级分镜切换景别和动作，商品用法必须真实，不要诡异姿势或错误容器。"
|
||||
"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。"
|
||||
)
|
||||
VIDEO_OLD = (
|
||||
"{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。"
|
||||
)
|
||||
|
||||
|
||||
def _set(apps, key, template):
|
||||
PromptTemplate = apps.get_model("ai", "PromptTemplate")
|
||||
PromptTemplate.objects.filter(key=key).update(template=template)
|
||||
|
||||
|
||||
def forwards(apps, schema_editor):
|
||||
_set(apps, "storyboard_frame", STORYBOARD)
|
||||
_set(apps, "video_segment", VIDEO)
|
||||
|
||||
|
||||
def backwards(apps, schema_editor):
|
||||
_set(apps, "storyboard_frame", STORYBOARD_OLD)
|
||||
_set(apps, "video_segment", VIDEO_OLD)
|
||||
|
||||
|
||||
class Migration(migrations.Migration):
|
||||
|
||||
dependencies = [("ai", "0030_alter_aitask_task_type")]
|
||||
|
||||
operations = [migrations.RunPython(forwards, backwards)]
|
||||
@@ -19,8 +19,8 @@ SSE 事件(每帧 `data: {json}\n\n`,json 带 type):
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
import math
|
||||
import re
|
||||
import uuid
|
||||
from decimal import Decimal
|
||||
from functools import lru_cache
|
||||
from pathlib import Path
|
||||
@@ -36,10 +36,11 @@ VALID_TONES = ["种草", "测评", "剧情", "痛点"]
|
||||
VALID_ROLES = ["钩子", "痛点", "卖点", "CTA"]
|
||||
VALID_ENTITY_TYPES = ["character", "scene", "product"]
|
||||
|
||||
# 时长:总时长 5–60 秒按 5 秒步进;单镜 4–15 秒(15 是出片模型硬上限,越界下游直接拒片)。
|
||||
TOTAL_DURATION_MIN = 5
|
||||
# 时长:主流程每镜固定 15 秒;总时长只能是 15/30/45/60。
|
||||
# 4–15 仍是出片模型合法区间,手改单镜 / 精准改一镜(旧稿)才用。
|
||||
TOTAL_DURATION_MIN = 15
|
||||
TOTAL_DURATION_MAX = 60
|
||||
TOTAL_DURATION_STEP = 5
|
||||
TOTAL_DURATION_STEP = 15
|
||||
SEGMENT_DURATION_MIN = 4
|
||||
SEGMENT_DURATION_MAX = 15
|
||||
DEFAULT_TOTAL_DURATION = 30
|
||||
@@ -58,22 +59,84 @@ DEFAULT_VIDEO_STRUCTURE = "pain"
|
||||
# 唯一禁用组合:短剧 × 测评验证。演出来的实测没有可信度,详见 playbooks/combo-matrix.md。
|
||||
FORBIDDEN_COMBOS: set[tuple[str, str]] = {("drama", "review")}
|
||||
|
||||
# 表现形式推荐的单镜节奏(秒)。镜数 ≈ 总时长 / 该值,再夹到 4–15 秒的合法区间。
|
||||
FORMAT_SHOT_PACE: dict[str, int] = {"oral": 12, "drama": 8, "vlog": 7}
|
||||
# 表现形式推荐的默认总时长:口播短平快,短剧要装下三幕,Vlog 要铺氛围。
|
||||
FORMAT_DEFAULT_DURATION: dict[str, int] = {"oral": 30, "drama": 45, "vlog": 30}
|
||||
# 各结构能压到的最短总时长(低于此值证据/氛围不成立),见 playbooks/combo-matrix.md。
|
||||
STRUCTURE_MIN_DURATION: dict[str, int] = {"pain": 15, "contrast": 10, "review": 20, "scene": 20}
|
||||
# 各结构能压到的最短总时长(须落在 15 秒步进上),见 playbooks/combo-matrix.md。
|
||||
STRUCTURE_MIN_DURATION: dict[str, int] = {"pain": 15, "contrast": 15, "review": 30, "scene": 30}
|
||||
|
||||
# 可懂语速上限 3.5 字/秒 —— 旁白字数按这一镜自己的秒数算,不再全场 55 字一刀切。
|
||||
# 设定卡人物 key → 中文(与前端 WIZ_PERSONA_LABEL / 模板 coerce_persona 对齐)。
|
||||
PERSONA_LABELS: dict[str, str] = {
|
||||
"urban": "都市白领女性",
|
||||
"bestie": "闺蜜种草",
|
||||
"ceo": "总裁亲选",
|
||||
"reviewer": "专业测评师",
|
||||
"mom": "实用宝妈",
|
||||
"genz": "学生党",
|
||||
}
|
||||
PERSONA_BRIEFS: dict[str, str] = {
|
||||
"urban": "25–32岁都市白领,工位或下班回家,说话像跟同事吐槽,不要主播腔",
|
||||
"bestie": "闺蜜分享口吻,带点兴奋,爱用「你懂的」「我跟你讲」",
|
||||
"ceo": "利落、判断句、少形容词,像拍板不是带货",
|
||||
"reviewer": "先讲怎么试的再给结论,允许提一个小缺点才可信",
|
||||
"mom": "带娃/家务间隙,讲省事,孩子或家人能沾边",
|
||||
"genz": "宿舍或通勤,短句,像给朋友发语音条",
|
||||
}
|
||||
_PERSONA_KEY_BY_LABEL = {label: key for key, label in PERSONA_LABELS.items()}
|
||||
|
||||
# 可懂语速上限 3.5 字/秒;15 秒口播还要有下限,否则模型写一句 20 字就收,撑不满。
|
||||
NARRATION_CHARS_PER_SECOND = 3.5
|
||||
NARRATION_CHARS_PER_SECOND_MIN = 2.4
|
||||
NARRATION_CHARS_HARD_CAP = 55
|
||||
VISUAL_CHARS_MIN = 72
|
||||
SHOT_BEATS_MIN = 3
|
||||
BEAT_SPAN_RE = re.compile(
|
||||
r"(?P<start>\d{1,2})\s*[-–—~到至]\s*(?P<end>\d{1,2})\s*(?:s|秒)?\s*[：:]",
|
||||
re.IGNORECASE,
|
||||
)
|
||||
_SHOT_SIZE_MARKERS = (
|
||||
"特写", "近景", "中近景", "中景", "全景", "远景", "胸上", "过肩",
|
||||
"手持", "跟拍", "俯拍", "仰拍", "推近", "拉远", "摇",
|
||||
)
|
||||
|
||||
|
||||
_FORMAT_KEY_BY_LABEL = {label: key for key, label in PRESENTATION_FORMATS.items()}
|
||||
_STRUCTURE_KEY_BY_LABEL = {label: key for key, label in VIDEO_STRUCTURES.items()}
|
||||
|
||||
|
||||
def coerce_persona(value: str | None) -> str:
|
||||
"""把设定卡 key 或中文标签收成 key;空/未知原样(未知时当补充描述用)。"""
|
||||
raw = (value or "").strip()
|
||||
if not raw:
|
||||
return ""
|
||||
if raw in PERSONA_LABELS:
|
||||
return raw
|
||||
return _PERSONA_KEY_BY_LABEL.get(raw, raw)
|
||||
|
||||
|
||||
def persona_label(value: str | None) -> str:
|
||||
key = coerce_persona(value)
|
||||
return PERSONA_LABELS.get(key, key)
|
||||
|
||||
|
||||
def _wizard_meta(project) -> dict:
|
||||
meta = getattr(project, "metadata", None) or {}
|
||||
wizard = meta.get("wizard") if isinstance(meta, dict) else None
|
||||
return wizard if isinstance(wizard, dict) else {}
|
||||
|
||||
|
||||
def _resolve_selling_point_ids(project, selling_point_ids: list | None) -> list[str]:
|
||||
if selling_point_ids:
|
||||
return [str(item) for item in selling_point_ids if item]
|
||||
raw = _wizard_meta(project).get("selling_point_ids") or []
|
||||
if not isinstance(raw, list):
|
||||
return []
|
||||
return [str(item) for item in raw if item]
|
||||
|
||||
|
||||
def _resolve_persona(project, persona: str | None) -> str:
|
||||
return coerce_persona(persona) or coerce_persona(_wizard_meta(project).get("persona"))
|
||||
|
||||
|
||||
def combo_keys(value_format, value_structure) -> tuple[str, str]:
|
||||
"""把「中文标签或 ASCII key」都归一成 key。落库存的是中文,请求传的是 key,两边都要认。"""
|
||||
fmt = _FORMAT_KEY_BY_LABEL.get(value_format, value_format)
|
||||
@@ -101,6 +164,61 @@ def narration_limit(duration: int) -> int:
|
||||
return max(1, min(NARRATION_CHARS_HARD_CAP, int(duration * NARRATION_CHARS_PER_SECOND)))
|
||||
|
||||
|
||||
def narration_floor(duration: int) -> int:
|
||||
"""15 秒口播至少要说到这个字数,再短就撑不满镜头。"""
|
||||
cap = narration_limit(duration)
|
||||
return max(1, min(cap - 6, int(duration * NARRATION_CHARS_PER_SECOND_MIN)))
|
||||
|
||||
|
||||
def _compact_len(text: str) -> int:
|
||||
return len(re.sub(r"\s+", "", text or ""))
|
||||
|
||||
|
||||
def _speech_text(seg: dict) -> str:
|
||||
dialogue = seg.get("dialogue") if isinstance(seg.get("dialogue"), list) else []
|
||||
lines = [
|
||||
str(item.get("line") or "")
|
||||
for item in dialogue
|
||||
if isinstance(item, dict) and (item.get("line") or "").strip()
|
||||
]
|
||||
if lines:
|
||||
return "".join(lines)
|
||||
return str(seg.get("narration") or "")
|
||||
|
||||
|
||||
def min_beats_for_duration(duration: int) -> int:
|
||||
if duration >= 12:
|
||||
return SHOT_BEATS_MIN
|
||||
if duration >= 8:
|
||||
return 2
|
||||
return 0
|
||||
|
||||
|
||||
def parse_visual_beats(text: str) -> list[tuple[int, int, str]]:
|
||||
"""从 visual 文本里抽出「0-3s：…」分镜。接得上才能喂下游视频。"""
|
||||
raw = (text or "").strip()
|
||||
if not raw:
|
||||
return []
|
||||
matches = list(BEAT_SPAN_RE.finditer(raw))
|
||||
if not matches:
|
||||
return []
|
||||
beats: list[tuple[int, int, str]] = []
|
||||
for index, match in enumerate(matches):
|
||||
start, end = int(match.group("start")), int(match.group("end"))
|
||||
if end < start:
|
||||
start, end = end, start
|
||||
body_to = matches[index + 1].start() if index + 1 < len(matches) else len(raw)
|
||||
body = re.sub(r"[\n\r]+", " ", raw[match.end():body_to]).strip(" ；;，,、")
|
||||
if start == end or not body:
|
||||
continue
|
||||
beats.append((start, end, body))
|
||||
return beats
|
||||
|
||||
|
||||
def format_visual_beats(beats: list[tuple[int, int, str]]) -> str:
|
||||
return "\n".join(f"{start}-{end}s：{body}" for start, end, body in beats)
|
||||
|
||||
|
||||
# --------------------------------------------------------------------------- #
|
||||
# skill 加载(缓存)
|
||||
# --------------------------------------------------------------------------- #
|
||||
@@ -174,14 +292,16 @@ _OUTPUT_PROTOCOL = """
|
||||
## 运行时输出协议(AirShelf 流式展示专用,优先级高于技能正文的「只输出 JSON」)
|
||||
|
||||
严格按以下顺序输出,不要有别的内容:
|
||||
1. 先用 **1 句中文口语**告诉用户你正在做什么(≤40 字,例:「在为这款保温杯生成 4 镜痛点脚本…」),让用户看到进展;
|
||||
1. 先用 **1 句中文口语**告诉用户你正在做什么(≤40 字,例:「在为这款保温杯生成 2 镜痛点脚本…」),让用户看到进展;
|
||||
2. 紧接着输出**且仅输出一个** ```json 代码块,内容为符合技能契约(铁律1)的 ScriptDraft 对象;
|
||||
3. json 代码块**收尾之后另起一行**,用 **1–2 句中文口语**跟用户交付这一版:做了什么、为什么这么改、可以怎么接着调(像同事汇报,**别复述 JSON 字段、别再写代码块**)。这句会作为你的回复气泡展示给用户。
|
||||
|
||||
### 字段名锚定(硬性 · 下游靠它取数,跑偏即数据全空)
|
||||
- 分镜数组的键名**必须**叫 `segments`(禁止用 scenes / script / shots / 分镜 等同义词)。
|
||||
- 每镜口播键名**必须**叫 `narration`(禁止用 voiceover / audio / line)。
|
||||
- 每镜画面键名**必须**叫 `visual`,且为**一句话字符串**(禁止用 scene / screen / 画面,也禁止写成 {setting,camera,...} 对象)。
|
||||
- 每镜画面键名**必须**叫 `visual`(禁止用 scene / screen / 画面 当键名)。
|
||||
- `visual` 是**多行字符串**,每行一条秒级分镜:`0-3s：近景，……`(禁止只写一句静态动作,也禁止只给 setting/camera 对象而不给 visual)。
|
||||
- 允许额外给 `beats`/`shots` 数组,后端会折进 visual;有数组也必须能折成「起-止秒」格式。
|
||||
- 即使你额外附带了 scenes / shots 等创作结构,也**必须同时**给出标准 `segments` 数组,并把口播填进 `narration`、画面填进 `visual`,否则视为不合格。
|
||||
"""
|
||||
|
||||
@@ -189,25 +309,87 @@ _OUTPUT_PROTOCOL = """
|
||||
# --------------------------------------------------------------------------- #
|
||||
# 提示词构建(3 模式)
|
||||
# --------------------------------------------------------------------------- #
|
||||
def _product_context(project, selling_point_ids: list[str] | None) -> str:
|
||||
def _specs_lines(specs) -> str:
|
||||
if not isinstance(specs, dict) or not specs:
|
||||
return ""
|
||||
skip = {"source"}
|
||||
labels = {"price": "价格"}
|
||||
lines: list[str] = []
|
||||
for key, val in specs.items():
|
||||
if key in skip or val in (None, "", [], {}):
|
||||
continue
|
||||
if isinstance(val, (dict, list)):
|
||||
rendered = json.dumps(val, ensure_ascii=False)
|
||||
else:
|
||||
rendered = str(val).strip()
|
||||
if not rendered:
|
||||
continue
|
||||
lines.append(f"- {labels.get(key, key)}:{rendered}")
|
||||
return "\n".join(lines)
|
||||
|
||||
|
||||
def _product_facts(project, selling_point_ids: list[str] | None):
|
||||
"""商品事实 + 本次勾选的卖点(供提示词和落库前原词校验共用)。
|
||||
|
||||
新建向导历史上把卖点**标题**(如「茶」)写进 metadata.wizard.selling_point_ids,
|
||||
字段名叫 ids 但不是 UUID。按 id__in 过滤会直接 ValidationError 把 SSE 打崩。
|
||||
这里 UUID 和标题都认;对不上就回落全部卖点,绝不抛。
|
||||
"""
|
||||
product = project.product
|
||||
selling_points = product.selling_points.all()
|
||||
if selling_point_ids:
|
||||
selling_points = selling_points.filter(id__in=selling_point_ids)
|
||||
selling_text = "\n".join(f"- {sp.title}:{sp.detail}" for sp in selling_points)
|
||||
keys = [str(item).strip() for item in _resolve_selling_point_ids(project, selling_point_ids) if str(item).strip()]
|
||||
all_points = list(product.selling_points.all())
|
||||
if not keys:
|
||||
return product, all_points
|
||||
uuid_keys: set[str] = set()
|
||||
title_keys: set[str] = set()
|
||||
for key in keys:
|
||||
try:
|
||||
uuid_keys.add(str(uuid.UUID(key)))
|
||||
except (ValueError, AttributeError, TypeError):
|
||||
title_keys.add(key)
|
||||
matched = [
|
||||
point
|
||||
for point in all_points
|
||||
if str(getattr(point, "id", "") or "") in uuid_keys
|
||||
or (getattr(point, "title", "") or "").strip() in title_keys
|
||||
]
|
||||
return product, matched or all_points
|
||||
|
||||
|
||||
def _product_context(project, selling_point_ids: list[str] | None, persona: str | None = None) -> str:
|
||||
product, selling_points = _product_facts(project, selling_point_ids)
|
||||
selling_text = "\n".join(f"- {sp.title}:{sp.detail or sp.title}" for sp in selling_points)
|
||||
business_type = getattr(product, "business_type", "") or "ecommerce"
|
||||
if business_type == "local_life":
|
||||
type_line = "业务类型:本地生活(团购/到店核销,无实物,按虚拟商品走主流程 SOP)\n"
|
||||
else:
|
||||
type_line = "业务类型:电商(实物商品)\n"
|
||||
specs_text = _specs_lines(getattr(product, "specs", None) or {})
|
||||
persona_key = _resolve_persona(project, persona)
|
||||
persona_text = persona_label(persona_key)
|
||||
persona_brief = PERSONA_BRIEFS.get(persona_key, "")
|
||||
must = [f"商品名「{product.title}」"] if product.title else []
|
||||
if product.brand:
|
||||
must.append(f"品牌「{product.brand}」")
|
||||
must.extend(f"卖点「{sp.title.strip()}」" for sp in selling_points if (sp.title or "").strip())
|
||||
must_line = "、".join(must) if must else "无(根据商品描述自行提炼,禁止空话套话)"
|
||||
specs_block = f"\n{specs_text}" if specs_text else "未填写"
|
||||
desc = (product.description or "").strip() or "未填写"
|
||||
return (
|
||||
f"商品标题:{product.title}\n"
|
||||
f"品牌:{product.brand or '未填写'}\n"
|
||||
f"{type_line}"
|
||||
f"品类:{product.category or '未填写'}\n"
|
||||
f"目标人群:{product.target_audience or '未填写'}\n"
|
||||
f"商品描述:{product.description or '未填写'}\n"
|
||||
f"卖点:\n{selling_text or '未勾选卖点,请根据商品信息自行提炼。'}"
|
||||
f"人物设定:{persona_text or '未指定,按目标人群自行定一个具体身份'}\n"
|
||||
f"人物口吻:{persona_brief or '按人物设定自己定一个具体身份,不要用万能主播腔'}\n"
|
||||
f"商品描述:{desc}\n"
|
||||
f"规格:{specs_block}\n"
|
||||
f"本次必须用上的卖点:\n{selling_text or '未勾选卖点,请根据商品信息自行提炼,禁止空话。'}\n"
|
||||
f"【必须原词用上】旁白/对白里要出现:{must_line}。"
|
||||
f"禁止用「补水/好用/值得买/宝藏」这类空卖点替换上面的原词。"
|
||||
f"商品名全片点名 1–2 次即可,其余镜用卖点原词和可感知细节(口感/气味/动作/使用场景),不要每句重复商品名。"
|
||||
f"描述若未填写,就从标题+卖点把使用感写具体,禁止只喊商品名。"
|
||||
)
|
||||
|
||||
|
||||
@@ -223,23 +405,80 @@ def build_agent_messages(
|
||||
presentation_format: str = DEFAULT_PRESENTATION_FORMAT,
|
||||
video_structure: str = DEFAULT_VIDEO_STRUCTURE,
|
||||
target_index: int | None = None,
|
||||
persona: str | None = None,
|
||||
) -> list[dict[str, str]]:
|
||||
fmt, structure = coerce_combo(presentation_format, video_structure)
|
||||
if target_index is not None:
|
||||
try:
|
||||
total = int(total_duration)
|
||||
except (TypeError, ValueError):
|
||||
total = DEFAULT_TOTAL_DURATION
|
||||
if total <= 0:
|
||||
total = DEFAULT_TOTAL_DURATION
|
||||
else:
|
||||
total = coerce_total_duration(total_duration)
|
||||
system = load_ecommerce_skill(fmt, structure) + _OUTPUT_PROTOCOL
|
||||
|
||||
# 给一组建议时长(不是硬性),模型可以按内容调整,只要每镜 4–15 秒且加总不变。
|
||||
suggested = plan_segment_durations(total, fmt)
|
||||
pace_hint = "+".join(str(d) for d in suggested)
|
||||
shot_n = len(suggested)
|
||||
extra = (user_prompt or "").strip()
|
||||
if target_index is not None and base_draft:
|
||||
existing_n = len(base_draft.get("segments") or [])
|
||||
duration_line = (
|
||||
f"【分镜时长】保持现有 {existing_n} 镜和每镜秒数,"
|
||||
f"只改第 {target_index + 1} 镜的文案与画面,不要加减镜、不要改时长。\n"
|
||||
)
|
||||
else:
|
||||
duration_line = (
|
||||
f"【分镜时长】每镜必须 {SEGMENT_DURATION_MAX} 秒,一共 {shot_n} 镜;"
|
||||
f"禁止写成 8/10/12 这种不等长,也禁止加减镜数。\n"
|
||||
)
|
||||
speech_floor = narration_floor(SEGMENT_DURATION_MAX)
|
||||
speech_cap = narration_limit(SEGMENT_DURATION_MAX)
|
||||
beats_line = (
|
||||
f"【秒级分镜】每个 {SEGMENT_DURATION_MAX} 秒场必须拆成 3–5 个分镜,visual 写成多行,"
|
||||
f"格式`0-3s：景别，谁在做什么`,最后一条接到 {SEGMENT_DURATION_MAX}s。"
|
||||
"每条写清手/商品/容器的空间关系和真实用法"
|
||||
"(茶=热水+蒸汽+茶汤变色,禁止茶包丢进冷白开;手从真实方向入画,禁止悬浮肢体)。"
|
||||
"禁止一句空画面撑满 15 秒。允许另给 beats 数组,后端会折进 visual。\n"
|
||||
)
|
||||
if fmt == "oral":
|
||||
writing_line = (
|
||||
f"【写法硬约束】口播每镜必须说满 {speech_floor}–{speech_cap} 字(2–4 句短句,"
|
||||
f"禁止一句 20 字收工);visual 至少 {VISUAL_CHARS_MIN} 字。"
|
||||
"第一镜用人物口吻交代身份,不要万能主播腔。\n"
|
||||
)
|
||||
elif fmt == "drama":
|
||||
writing_line = (
|
||||
f"【写法硬约束】对白每镜合计 {speech_floor}–{speech_cap} 字;"
|
||||
f"visual 至少 {VISUAL_CHARS_MIN} 字。\n"
|
||||
)
|
||||
else:
|
||||
writing_line = (
|
||||
f"【写法硬约束】有人声的镜口播/对白 {speech_floor}–{speech_cap} 字,允许个别镜纯画面;"
|
||||
f"visual 至少 {VISUAL_CHARS_MIN} 字。\n"
|
||||
)
|
||||
structure_line = ""
|
||||
if structure == "scene":
|
||||
structure_line = (
|
||||
"【场景种草骨架】按镜走完:场景建立(商品先不出现)→商品被使用着入场→场景因它更好/软CTA。"
|
||||
"每镜写出具体时刻+地点+光/声音/物件。口播时每镜仍要说满字数,留白靠停顿不是靠少写字。\n"
|
||||
)
|
||||
combo_line = (
|
||||
f"严格按已加载的「{PRESENTATION_FORMATS[fmt]} × {VIDEO_STRUCTURES[structure]}」套路写,"
|
||||
f"不要串成别的结构的套话。\n"
|
||||
)
|
||||
head = (
|
||||
f"【画幅】{aspect_ratio}\n"
|
||||
f"【表现形式】{PRESENTATION_FORMATS[fmt]}(套路见 playbooks/format-{fmt}.md,已加载)\n"
|
||||
f"【视频结构】{VIDEO_STRUCTURES[structure]}(套路见 playbooks/structure-{structure}.md,已加载)\n"
|
||||
f"【总时长】{total} 秒\n"
|
||||
f"【分镜时长】单镜 4–15 秒,可以不等长;各镜相加必须精确等于 {total} 秒。\n"
|
||||
f" 建议切成 {len(suggested)} 镜({pace_hint}),这是按「{PRESENTATION_FORMATS[fmt]}」的节奏算的;\n"
|
||||
f" 你可以按内容调整镜数与每镜长短(该长的给足、该短的压短),但必须守住上面两条硬约束。\n"
|
||||
f"【商品信息】\n{_product_context(project, selling_point_ids)}"
|
||||
f"{duration_line}"
|
||||
f"{writing_line}"
|
||||
f"{beats_line}"
|
||||
f"{structure_line}"
|
||||
f"{combo_line}"
|
||||
f"【商品信息】\n{_product_context(project, selling_point_ids, persona)}"
|
||||
)
|
||||
if mode == "revise" and base_draft and target_index is not None:
|
||||
# 精准改一镜:读全脚本上下文,只重写第 N 镜,强制与前后镜衔接;其余镜后端会强制保持原样。
|
||||
@@ -247,7 +486,7 @@ def build_agent_messages(
|
||||
f"【任务】只重写第 {target_index + 1} 镜(共 {len(base_draft.get('segments', []))} 镜),其余镜保持不变。\n"
|
||||
f"{head}\n\n"
|
||||
f"【现有完整脚本 JSON(读它保证与前后镜衔接)】\n{json.dumps(base_draft, ensure_ascii=False)}\n\n"
|
||||
f"【对第 {target_index + 1} 镜的修改意见】{user_prompt.strip() or '让这一镜更有吸引力、表达更清晰。'}\n\n"
|
||||
f"【对第 {target_index + 1} 镜的修改意见】{extra or '让这一镜更有吸引力、表达更清晰。'}\n\n"
|
||||
"仍输出**完整** ScriptDraft(我只会采用第 "
|
||||
f"{target_index + 1} 镜的改动);若意见涉及角色对白,就给这一镜填 dialogue。"
|
||||
)
|
||||
@@ -256,21 +495,25 @@ def build_agent_messages(
|
||||
"【任务】改稿(模式③):在保留用户原意的前提下,增强钩子/节奏/卖点/CTA,并归一化到契约 JSON。\n"
|
||||
f"{head}\n\n"
|
||||
f"【现有脚本 JSON】\n{json.dumps(base_draft, ensure_ascii=False)}\n\n"
|
||||
f"【用户修改意见】{user_prompt.strip() or '让整体更有吸引力、转化感更强,并保持各镜衔接连贯。'}\n\n"
|
||||
f"【用户修改意见】{extra or '让整体更有吸引力、转化感更强,并保持各镜衔接连贯。'}\n\n"
|
||||
"请输出修订后的**完整** ScriptDraft。"
|
||||
)
|
||||
elif mode == "theme" or (user_prompt and user_prompt.strip()):
|
||||
elif mode == "theme":
|
||||
user = (
|
||||
"【任务】一句话主题扩写(模式②):以用户主题为脚本主轴,其余自动补全。\n"
|
||||
"【任务】一句话主题扩写(模式②):以用户主题为脚本主轴,但商品事实和必须原词仍要全部用上。\n"
|
||||
f"{head}\n\n"
|
||||
f"【用户主题】{user_prompt.strip()}\n\n"
|
||||
f"【用户主题】{extra or '按商品最强卖点选题'}\n\n"
|
||||
"请按技能流程一次性产出 ScriptDraft。"
|
||||
)
|
||||
else:
|
||||
extra_block = f"\n\n【补充要求】{extra}\n" if extra else "\n"
|
||||
user = (
|
||||
"【任务】全自动(模式①):仅凭商品与前置条件,按指定的表现形式与视频结构套路"
|
||||
"自动定镜数/选 tone/造 entity/填结构骨架。\n"
|
||||
f"{head}\n\n"
|
||||
"【任务】全自动(模式①):仅凭上面的商品事实与前置条件,按指定的表现形式与视频结构套路"
|
||||
"自动定镜/选 tone/造 entity/填结构骨架。"
|
||||
"不要另起一个空主题;钩子、痛点、卖点必须能对上这份商品,而不是品类套话。"
|
||||
"每镜旁白要能撑满 15 秒;画面必须按秒拆分镜,够导演在 15 秒里切 3–5 刀。\n"
|
||||
f"{head}"
|
||||
f"{extra_block}\n"
|
||||
"请按技能流程一次性产出 ScriptDraft。"
|
||||
)
|
||||
return [{"role": "system", "content": system}, {"role": "user", "content": user}]
|
||||
@@ -333,7 +576,7 @@ def _extract_json(text: str) -> str | None:
|
||||
|
||||
|
||||
def coerce_total_duration(value) -> int:
|
||||
"""总时长夹到 5–60 秒、5 秒步进。空值/0/非法输入一律回落默认 30。"""
|
||||
"""总时长夹到 15–60 秒、15 秒步进(15/30/45/60)。空值/0/非法输入一律回落默认 30。"""
|
||||
if value in (None, "", 0):
|
||||
return DEFAULT_TOTAL_DURATION
|
||||
try:
|
||||
@@ -347,21 +590,11 @@ def coerce_total_duration(value) -> int:
|
||||
return max(TOTAL_DURATION_MIN, min(TOTAL_DURATION_MAX, stepped))
|
||||
|
||||
|
||||
def plan_segment_durations(total_duration: int, presentation_format: str) -> list[int]:
|
||||
"""把总时长切成每镜 4–15 秒、加总精确等于总时长的一组时长。
|
||||
|
||||
镜数按表现形式的推荐节奏定(口播 12s/镜、短剧 8s/镜、Vlog 7s/镜),再夹进
|
||||
ceil(total/15) ~ total//4 的合法区间。余数摊到前面几镜,所以镜与镜之间最多差 1 秒——
|
||||
这只是**兜底**,模型自己给的不等长时长只要合法就照用。
|
||||
"""
|
||||
def plan_segment_durations(total_duration: int, presentation_format: str | None = None) -> list[int]:
|
||||
"""主流程每镜固定 15 秒。镜数 = 总时长 / 15。presentation_format 保留签名以免调用方改动。"""
|
||||
total = coerce_total_duration(total_duration)
|
||||
fmt = presentation_format if presentation_format in FORMAT_SHOT_PACE else DEFAULT_PRESENTATION_FORMAT
|
||||
count_min = math.ceil(total / SEGMENT_DURATION_MAX)
|
||||
count_max = max(count_min, total // SEGMENT_DURATION_MIN)
|
||||
count = max(1, round(total / FORMAT_SHOT_PACE[fmt]))
|
||||
count = max(count_min, min(count_max, count))
|
||||
base, remainder = divmod(total, count)
|
||||
return [base + 1 if i < remainder else base for i in range(count)]
|
||||
count = max(1, total // SEGMENT_DURATION_MAX)
|
||||
return [SEGMENT_DURATION_MAX] * count
|
||||
|
||||
|
||||
def plan_roles(count: int) -> list[str]:
|
||||
@@ -376,45 +609,106 @@ def plan_roles(count: int) -> list[str]:
|
||||
|
||||
|
||||
def _fit_segment_durations(raw: list, total: int, presentation_format: str) -> list[int]:
|
||||
"""采纳模型给的每镜时长(允许不等长),非法就修;修不动就整组回落到 plan_segment_durations。
|
||||
|
||||
合法定义:每镜 4–15 秒的整数,且加总 == 总时长。模型很容易把总数算错一两秒,
|
||||
所以先夹单镜范围,再把差额摊到还有余量的镜上,尽量保住模型的节奏意图。
|
||||
"""
|
||||
if not raw:
|
||||
"""主流程强制每镜 15 秒。模型给多少秒都丢掉,按总时长切成 N 个 15。"""
|
||||
return plan_segment_durations(total, presentation_format)
|
||||
|
||||
|
||||
def _layout_durations(preserve_layout: dict | None, count: int) -> list[int] | None:
|
||||
"""精准改一镜:保住原稿每镜秒数,避免 3 镜旧稿被压成 2×15。"""
|
||||
if not preserve_layout:
|
||||
return None
|
||||
raw = preserve_layout.get("durations")
|
||||
if not isinstance(raw, list) or len(raw) != count:
|
||||
return None
|
||||
durations: list[int] = []
|
||||
for value in raw:
|
||||
try:
|
||||
seconds = int(value)
|
||||
except (TypeError, ValueError):
|
||||
seconds = 0
|
||||
durations.append(max(SEGMENT_DURATION_MIN, min(SEGMENT_DURATION_MAX, seconds or SEGMENT_DURATION_MIN)))
|
||||
|
||||
# 镜数本身就装不下总时长(太少会超 15s/镜,太多会低于 4s/镜)→ 模型节奏不可用,整组重排。
|
||||
count = len(durations)
|
||||
if not (count * SEGMENT_DURATION_MIN <= total <= count * SEGMENT_DURATION_MAX):
|
||||
return plan_segment_durations(total, presentation_format)
|
||||
|
||||
diff = total - sum(durations)
|
||||
while diff != 0:
|
||||
step = 1 if diff > 0 else -1
|
||||
# 每轮只给「还有余量」的镜加/减 1 秒,均匀铺开,避免把某一镜顶到边界
|
||||
movable = [
|
||||
i for i, d in enumerate(durations)
|
||||
if (step > 0 and d < SEGMENT_DURATION_MAX) or (step < 0 and d > SEGMENT_DURATION_MIN)
|
||||
]
|
||||
if not movable:
|
||||
return plan_segment_durations(total, presentation_format)
|
||||
for i in movable:
|
||||
if diff == 0:
|
||||
break
|
||||
durations[i] += step
|
||||
diff -= step
|
||||
durations.append(max(SEGMENT_DURATION_MIN, min(SEGMENT_DURATION_MAX, seconds or SEGMENT_DURATION_MAX)))
|
||||
return durations
|
||||
|
||||
|
||||
def _script_plaintext(draft: dict) -> str:
|
||||
parts = [str(draft.get("hook") or "")]
|
||||
for seg in draft.get("segments") or []:
|
||||
if not isinstance(seg, dict):
|
||||
continue
|
||||
parts.append(str(seg.get("narration") or ""))
|
||||
parts.append(str(seg.get("visual") or ""))
|
||||
parts.append(str(seg.get("product_exposure") or ""))
|
||||
for item in seg.get("dialogue") or []:
|
||||
if isinstance(item, dict):
|
||||
parts.append(str(item.get("line") or ""))
|
||||
return "".join(parts)
|
||||
|
||||
|
||||
def _fact_mentioned(text: str, fact: str) -> bool:
|
||||
fact = (fact or "").strip()
|
||||
if not fact:
|
||||
return False
|
||||
if fact in text:
|
||||
return True
|
||||
# 超长卖点/标题模型常念前半截,允许前 4 字命中
|
||||
return len(fact) >= 6 and fact[:4] in text
|
||||
|
||||
|
||||
def assert_product_facts_used(draft: dict, *, brand: str = "", selling_titles: list[str] | None = None) -> None:
|
||||
"""整版生成时至少用上一个勾选卖点原词。单字卖点(如「茶」)也算。
|
||||
|
||||
品牌不强制:口播经常只喊卖点不念品牌,硬卡会把已经写对的稿整单作废。
|
||||
"""
|
||||
del brand # 保留调用方签名,避免改一串入口
|
||||
text = _script_plaintext(draft)
|
||||
titles = [t.strip() for t in (selling_titles or []) if (t or "").strip()]
|
||||
if titles and not any(_fact_mentioned(text, title) for title in titles):
|
||||
raise ValueError("脚本没有用上商品卖点原词,请按【必须原词用上】重写")
|
||||
|
||||
|
||||
def assert_shot_density(draft: dict, presentation_format: str = DEFAULT_PRESENTATION_FORMAT) -> None:
|
||||
"""口播/短剧每镜要说满、画面要写厚;Vlog 允许个别镜无声,但开口就不能偷懒。"""
|
||||
fmt, _ = coerce_combo(_FORMAT_KEY_BY_LABEL.get(presentation_format, presentation_format), DEFAULT_VIDEO_STRUCTURE)
|
||||
allow_silent = fmt == "vlog"
|
||||
for index, seg in enumerate(draft.get("segments") or []):
|
||||
if not isinstance(seg, dict):
|
||||
continue
|
||||
try:
|
||||
duration = int(seg.get("duration") or SEGMENT_DURATION_MAX)
|
||||
except (TypeError, ValueError):
|
||||
duration = SEGMENT_DURATION_MAX
|
||||
visual_n = _compact_len(str(seg.get("visual") or ""))
|
||||
if visual_n < VISUAL_CHARS_MIN:
|
||||
raise ValueError(f"第 {index + 1} 镜画面描写太短,撑不满 {duration} 秒")
|
||||
assert_intra_shot_beats(seg, index, duration)
|
||||
speech_n = _compact_len(_speech_text(seg))
|
||||
if speech_n == 0 and allow_silent:
|
||||
continue
|
||||
floor = narration_floor(duration)
|
||||
if speech_n < floor:
|
||||
raise ValueError(f"第 {index + 1} 镜旁白太短,口播至少要说到 {floor} 字才能撑满 {duration} 秒")
|
||||
|
||||
|
||||
def assert_intra_shot_beats(seg: dict, index: int, duration: int) -> None:
|
||||
"""15 秒场必须按秒拆出分镜,否则下游视频只能对着一句空描述乱编。"""
|
||||
need = min_beats_for_duration(duration)
|
||||
if need <= 0:
|
||||
return
|
||||
visual = str(seg.get("visual") or "")
|
||||
beats = parse_visual_beats(visual)
|
||||
if len(beats) < need:
|
||||
raise ValueError(
|
||||
f"第 {index + 1} 镜画面必须按秒拆分镜,至少 {need} 条「0-3s：景别，动作」,接到 {duration}s"
|
||||
)
|
||||
if beats[0][0] > 1:
|
||||
raise ValueError(f"第 {index + 1} 镜画面必须按秒拆分镜,第一条要从 0 秒起")
|
||||
if beats[-1][1] < duration - 1:
|
||||
raise ValueError(f"第 {index + 1} 镜画面必须按秒拆分镜,最后一条要接到 {duration}s")
|
||||
markers = [mark for mark in _SHOT_SIZE_MARKERS if mark in visual]
|
||||
if len(set(markers)) < 2:
|
||||
raise ValueError(f"第 {index + 1} 镜 15 秒内至少要切两次景别,并写进秒级分镜")
|
||||
|
||||
|
||||
# 模型每次生成都可能换字段名(scene/screenDescription/visual…、dialogue/lines/caption…),
|
||||
# 与其逐一追变体,不如「优先键命中 → 否则按关键词模糊匹配」通用解析。SKIP 掉明显的非内容键,
|
||||
# 避免误抓(shotNo/duration/bgMusic/note 等)。
|
||||
@@ -422,6 +716,7 @@ _PICK_SKIP_KEYS = {
|
||||
"shotno", "shotsize", "shotsizetype", "duration", "duration_seconds", "starttime", "endtime",
|
||||
"timerange", "time_range", "bgmusic", "bg_music", "music", "sound", "sfx", "note", "notes",
|
||||
"tips", "index", "role", "speaker", "entity_refs", "product_exposure", "id", "no", "transition",
|
||||
"beats", "timeline",
|
||||
# 标题/编号类:含 scene 字样会误命中画面 fuzzy,显式跳过(注意:不跳裸 scene,它常=画面)
|
||||
"scene_id", "sceneid", "scene_no", "sceneno", "scene_number", "scenenumber",
|
||||
"scene_title", "scenetitle", "title", "scene_title_type",
|
||||
@@ -463,6 +758,113 @@ def _pick_field(seg: dict, exact: tuple[str, ...], fuzzy: tuple[str, ...]) -> st
|
||||
return ""
|
||||
|
||||
|
||||
def _opt_int(value) -> int | None:
|
||||
try:
|
||||
if value is None or value == "":
|
||||
return None
|
||||
return int(value)
|
||||
except (TypeError, ValueError):
|
||||
return None
|
||||
|
||||
|
||||
def _beat_from_dict(item: dict) -> dict | None:
|
||||
action = _flatten_text(
|
||||
item.get("action")
|
||||
or item.get("shot")
|
||||
or item.get("visual")
|
||||
or item.get("description")
|
||||
or item.get("画面")
|
||||
or ""
|
||||
)
|
||||
shot_size = str(item.get("shot_size") or item.get("camera") or item.get("景别") or "").strip()
|
||||
start = _opt_int(item.get("start") if item.get("start") is not None else item.get("from"))
|
||||
end = _opt_int(item.get("end") if item.get("end") is not None else item.get("to"))
|
||||
time_s = str(item.get("time") or item.get("t") or "").strip()
|
||||
if start is None and end is None and time_s:
|
||||
parsed = parse_visual_beats(f"{time_s}：{action or '画面'}")
|
||||
if parsed:
|
||||
start, end, parsed_body = parsed[0]
|
||||
action = action or parsed_body
|
||||
if not action and not shot_size:
|
||||
return None
|
||||
piece = f"{shot_size}，{action}" if shot_size and action else (action or shot_size)
|
||||
return {"start": start, "end": end, "action": piece}
|
||||
|
||||
|
||||
def coerce_beats(raw) -> list[dict]:
|
||||
if raw is None or raw == "":
|
||||
return []
|
||||
if isinstance(raw, str):
|
||||
return [{"start": start, "end": end, "action": body} for start, end, body in parse_visual_beats(raw)]
|
||||
if isinstance(raw, dict):
|
||||
nested = raw.get("beats") or raw.get("shots") or raw.get("timeline") or raw.get("key_shots") or raw.get("分镜")
|
||||
if nested:
|
||||
return coerce_beats(nested)
|
||||
beat = _beat_from_dict(raw)
|
||||
return [beat] if beat else []
|
||||
if isinstance(raw, list):
|
||||
out: list[dict] = []
|
||||
for item in raw:
|
||||
if isinstance(item, dict):
|
||||
nested = item.get("beats") or item.get("shots")
|
||||
if nested and not (item.get("action") or item.get("shot") or item.get("visual")):
|
||||
out.extend(coerce_beats(nested))
|
||||
continue
|
||||
beat = _beat_from_dict(item)
|
||||
if beat:
|
||||
out.append(beat)
|
||||
else:
|
||||
out.extend(coerce_beats(item))
|
||||
return out
|
||||
return []
|
||||
|
||||
|
||||
def _assign_beat_times(beats: list[dict], duration: int) -> list[tuple[int, int, str]]:
|
||||
n = len(beats)
|
||||
if n == 0:
|
||||
return []
|
||||
out: list[tuple[int, int, str]] = []
|
||||
for index, beat in enumerate(beats):
|
||||
start = beat.get("start")
|
||||
end = beat.get("end")
|
||||
if start is None:
|
||||
start = int(round(index * duration / n))
|
||||
if end is None:
|
||||
end = duration if index == n - 1 else int(round((index + 1) * duration / n))
|
||||
if end <= start:
|
||||
end = start + 1
|
||||
out.append((int(start), int(end), str(beat.get("action") or "").strip()))
|
||||
return [item for item in out if item[2]]
|
||||
|
||||
|
||||
def compose_segment_visual(seg: dict, duration: int = SEGMENT_DURATION_MAX) -> str:
|
||||
"""把 beats 数组或已写好的秒级 visual 收成多行「0-3s：…」。"""
|
||||
visual_raw = seg.get("visual")
|
||||
if isinstance(visual_raw, str) and visual_raw.strip():
|
||||
parsed = parse_visual_beats(visual_raw)
|
||||
if len(parsed) >= 2:
|
||||
return format_visual_beats(parsed)
|
||||
candidates = [
|
||||
seg.get("beats"),
|
||||
seg.get("timeline"),
|
||||
seg.get("分镜"),
|
||||
visual_raw if isinstance(visual_raw, (list, dict)) else None,
|
||||
]
|
||||
shots_raw = seg.get("shots")
|
||||
if isinstance(shots_raw, list) and shots_raw:
|
||||
sample = shots_raw[0]
|
||||
looks_like_nested_segments = isinstance(sample, dict) and (
|
||||
sample.get("narration") or sample.get("role") in VALID_ROLES
|
||||
)
|
||||
if not looks_like_nested_segments:
|
||||
candidates.append(shots_raw)
|
||||
for raw in candidates:
|
||||
assigned = _assign_beat_times(coerce_beats(raw), duration)
|
||||
if len(assigned) >= 2:
|
||||
return format_visual_beats(assigned)
|
||||
return _pick_field(seg, _VISUAL_EXACT, _VISUAL_FUZZY)
|
||||
|
||||
|
||||
# 模型给分镜数组的键名五花八门(segments/scenes/script/shots…),且常同时给一个**空的**
|
||||
# segments 骨架 + 真内容放在 scenes 里。所以不能「见 segments 是 list 就用」,要在所有候选里
|
||||
# 挑「能解析出最多非空旁白/画面」的那个。segments(契约本名)排第一,同分时优先。
|
||||
@@ -501,6 +903,7 @@ def normalize_draft(
|
||||
total_duration: int,
|
||||
presentation_format: str = DEFAULT_PRESENTATION_FORMAT,
|
||||
video_structure: str = DEFAULT_VIDEO_STRUCTURE,
|
||||
preserve_layout: dict | None = None,
|
||||
) -> dict:
|
||||
"""把模型输出抽成 JSON 并按铁律1契约规范化。宽容:小问题就地修,不轻易抛错。"""
|
||||
blob = _extract_json(raw_text)
|
||||
@@ -564,13 +967,17 @@ def normalize_draft(
|
||||
draft["entities"] = norm_entities
|
||||
valid_ids = {e["id"] for e in norm_entities}
|
||||
|
||||
# segments 规范化:镜数交给模型(只夹进合法区间),role 枚举,引用合法
|
||||
# segments 规范化:主流程镜数 = 总时长/15;精准改一镜则保住原稿镜数。
|
||||
segments = draft.get("segments") if isinstance(draft.get("segments"), list) else []
|
||||
# 镜数上下限由「单镜 4–15 秒」倒推:少于 count_min 会有镜超 15 秒,多于 count_max 会有镜不足 4 秒。
|
||||
count_min = math.ceil(dur / SEGMENT_DURATION_MAX)
|
||||
count_max = max(count_min, dur // SEGMENT_DURATION_MIN)
|
||||
segments = segments[:count_max]
|
||||
seg_count = max(count_min, len(segments))
|
||||
layout_count = 0
|
||||
if preserve_layout:
|
||||
try:
|
||||
layout_count = int(preserve_layout.get("count") or 0)
|
||||
except (TypeError, ValueError):
|
||||
layout_count = 0
|
||||
expected = layout_count or max(1, dur // SEGMENT_DURATION_MAX)
|
||||
segments = segments[:expected]
|
||||
seg_count = expected
|
||||
role_plan = plan_roles(seg_count)
|
||||
norm_segments: list[dict] = []
|
||||
for i, seg in enumerate(segments):
|
||||
@@ -604,12 +1011,12 @@ def normalize_draft(
|
||||
narration = " ".join(d["line"] for d in dialogue) # 扁平拼接,兼容下游字幕/配音
|
||||
if not narration:
|
||||
narration = _pick_field(seg, _NARRATION_EXACT, _NARRATION_FUZZY)
|
||||
# 画面:通用解析(visual/scene/screenDescription/画面… 都能命中,跳过 shotNo/bgMusic 等)
|
||||
visual = _pick_field(seg, _VISUAL_EXACT, _VISUAL_FUZZY)
|
||||
# 画面:优先收成秒级分镜;beats 数组会折进 visual,下游故事板/视频直接读这一段。
|
||||
visual = compose_segment_visual(seg)
|
||||
norm_segments.append(
|
||||
{
|
||||
"index": i,
|
||||
"duration": seg.get("duration"), # 先原样收着,等镜数定了再统一夹进 4–15 秒并配平总时长
|
||||
"duration": seg.get("duration"), # 先原样收着,后面统一写成 15 或保住原稿秒数
|
||||
"role": role,
|
||||
"narration": narration,
|
||||
"speaker": speaker,
|
||||
@@ -638,16 +1045,36 @@ def normalize_draft(
|
||||
if not norm_segments:
|
||||
raise ValueError("脚本没有任何分镜")
|
||||
|
||||
# 每镜时长:采纳模型的不等长意图,非法就修,修不动整组回落到按表现形式节奏均切。
|
||||
fitted = _fit_segment_durations([s["duration"] for s in norm_segments], dur, fmt)
|
||||
for seg, seconds in zip(norm_segments, fitted):
|
||||
seg["duration"] = seconds
|
||||
kept = _layout_durations(preserve_layout, len(norm_segments))
|
||||
if kept:
|
||||
fitted = kept
|
||||
draft["total_duration"] = sum(fitted)
|
||||
else:
|
||||
fitted = plan_segment_durations(dur, fmt)
|
||||
if len(fitted) != len(norm_segments):
|
||||
fitted = [SEGMENT_DURATION_MAX] * len(norm_segments)
|
||||
draft["total_duration"] = sum(fitted)
|
||||
for index, (norm, seconds) in enumerate(zip(norm_segments, fitted)):
|
||||
norm["duration"] = seconds
|
||||
if index < len(segments) and isinstance(segments[index], dict):
|
||||
composed = compose_segment_visual(segments[index], seconds)
|
||||
if composed:
|
||||
norm["visual"] = composed
|
||||
|
||||
draft["segments"] = norm_segments
|
||||
draft["segment_count"] = len(norm_segments)
|
||||
return draft
|
||||
|
||||
|
||||
def _preserve_layout_from(draft: dict | None) -> dict | None:
|
||||
if not draft:
|
||||
return None
|
||||
segs = [s for s in (draft.get("segments") or []) if isinstance(s, dict)]
|
||||
if not segs:
|
||||
return None
|
||||
return {"count": len(segs), "durations": [s.get("duration") for s in segs]}
|
||||
|
||||
|
||||
def _merge_single_segment(
|
||||
base: dict,
|
||||
new: dict,
|
||||
@@ -677,12 +1104,14 @@ def _merge_single_segment(
|
||||
target["index"] = idx
|
||||
segs[idx] = target
|
||||
merged["segments"] = segs
|
||||
layout = _preserve_layout_from(base)
|
||||
return normalize_draft(
|
||||
json.dumps(merged, ensure_ascii=False),
|
||||
aspect_ratio=aspect_ratio,
|
||||
total_duration=total_duration,
|
||||
presentation_format=presentation_format,
|
||||
video_structure=video_structure,
|
||||
preserve_layout=layout,
|
||||
)
|
||||
|
||||
|
||||
@@ -802,6 +1231,7 @@ def stream_script_agent(
|
||||
video_structure: str = DEFAULT_VIDEO_STRUCTURE,
|
||||
target_index: int | None = None,
|
||||
entry_source: str = "",
|
||||
persona: str | None = None,
|
||||
):
|
||||
"""生成 SSE 帧字符串的同步生成器,供 StreamingHttpResponse 包裹。
|
||||
target_index 非空 = 精准只改第 N 镜(读全脚本上下文,后端强制保留其余镜原样)。"""
|
||||
@@ -820,6 +1250,10 @@ def stream_script_agent(
|
||||
if base_draft is None:
|
||||
target_index = None # 没有基准稿就退回整版生成,单镜改无从谈起
|
||||
# 改稿以基准稿的时长/镜数为准,避免请求侧默认值把长稿的尾镜挤掉
|
||||
# 精准改一镜不把总时长夹成 15 步进,否则旧的不等长稿会被提示词误导切镜。
|
||||
if target_index is not None and base_draft is not None:
|
||||
effective_duration = int(base_draft.get("total_duration") or total_duration or DEFAULT_TOTAL_DURATION)
|
||||
else:
|
||||
effective_duration = coerce_total_duration(
|
||||
(base_draft.get("total_duration") if base_draft else None) or total_duration
|
||||
)
|
||||
@@ -829,6 +1263,11 @@ def stream_script_agent(
|
||||
if not (0 <= target_index < seg_n):
|
||||
yield _sse({"type": "error", "detail": f"镜号越界:第 {target_index + 1} 镜(共 {seg_n} 镜)"})
|
||||
return
|
||||
selling_point_ids = _resolve_selling_point_ids(project, selling_point_ids)
|
||||
persona = _resolve_persona(project, persona)
|
||||
product, selling_points = _product_facts(project, selling_point_ids)
|
||||
selling_titles = [sp.title for sp in selling_points]
|
||||
layout = _preserve_layout_from(base_draft) if target_index is not None else None
|
||||
messages = build_agent_messages(
|
||||
project=project,
|
||||
mode=mode,
|
||||
@@ -840,6 +1279,7 @@ def stream_script_agent(
|
||||
presentation_format=fmt,
|
||||
video_structure=structure,
|
||||
target_index=target_index,
|
||||
persona=persona,
|
||||
)
|
||||
yield _sse({"type": "tool", "id": "analyze", "status": "done"})
|
||||
|
||||
@@ -886,6 +1326,7 @@ def stream_script_agent(
|
||||
total_duration=effective_duration,
|
||||
presentation_format=fmt,
|
||||
video_structure=structure,
|
||||
preserve_layout=layout,
|
||||
)
|
||||
if target_index is not None and base_draft:
|
||||
return _merge_single_segment(
|
||||
@@ -897,6 +1338,13 @@ def stream_script_agent(
|
||||
fmt,
|
||||
structure,
|
||||
)
|
||||
if mode != "revise":
|
||||
assert_product_facts_used(
|
||||
candidate,
|
||||
brand=getattr(product, "brand", "") or "",
|
||||
selling_titles=selling_titles,
|
||||
)
|
||||
assert_shot_density(candidate, fmt)
|
||||
return candidate
|
||||
|
||||
routed_stream = stream_routed_text_request(
|
||||
@@ -1069,7 +1517,7 @@ def _draft_from_version(version) -> dict:
|
||||
(用户增删/改镜后 content 不一定同步)。total_duration 按各镜真实秒数加总,避免 normalize 按 stale 值截/补镜。"""
|
||||
meta = version.metadata or {}
|
||||
segs = list(version.segments.order_by("sort_order"))
|
||||
# 镜可以不等长了,总时长必须按实际相加(旧写法 15×镜数 会在不等长稿上算出错误总时长)
|
||||
# 总时长按各镜真实秒数加总(旧稿可能不等长;新稿每镜 15 秒)。
|
||||
actual_total = sum(s.duration_seconds or SEGMENT_DURATION_MAX for s in segs)
|
||||
return {
|
||||
"hook": meta.get("hook", ""),
|
||||
@@ -1119,13 +1567,14 @@ def regenerate_segment_via_agent(*, project, user, model_config: ModelConfig, se
|
||||
project=project,
|
||||
mode="revise",
|
||||
user_prompt=instruction,
|
||||
selling_point_ids=None,
|
||||
selling_point_ids=_resolve_selling_point_ids(project, None),
|
||||
base_draft=base_draft,
|
||||
aspect_ratio=aspect_ratio,
|
||||
total_duration=total_duration,
|
||||
presentation_format=fmt,
|
||||
video_structure=structure,
|
||||
target_index=target_index,
|
||||
persona=_resolve_persona(project, None),
|
||||
)
|
||||
task = create_ai_task(
|
||||
project=project,
|
||||
@@ -1152,12 +1601,14 @@ def regenerate_segment_via_agent(*, project, user, model_config: ModelConfig, se
|
||||
task.submitted_at = timezone.now()
|
||||
task.save(update_fields=["status", "submitted_at", "updated_at"])
|
||||
def validate_segment_text(raw_text: str) -> dict:
|
||||
layout = _preserve_layout_from(base_draft)
|
||||
candidate = normalize_draft(
|
||||
raw_text,
|
||||
aspect_ratio=aspect_ratio,
|
||||
total_duration=total_duration,
|
||||
presentation_format=fmt,
|
||||
video_structure=structure,
|
||||
preserve_layout=layout,
|
||||
)
|
||||
return _merge_single_segment(
|
||||
base_draft, candidate, target_index, aspect_ratio, total_duration, fmt, structure
|
||||
|
||||
@@ -2508,7 +2508,8 @@ def _segment_script_text(segment, entities=None, *, with_dialogue: bool = False,
|
||||
parts = []
|
||||
visual = (segment.visual_prompt or "").strip()
|
||||
if visual:
|
||||
parts.append(f"画面:{visual}")
|
||||
label = "秒级分镜(按时间切镜,商品用法必须真实,禁止诡异动作/错误容器)" if ("s：" in visual or "s:" in visual or "秒" in visual[:12]) else "画面"
|
||||
parts.append(f"{label}:\n{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 = []
|
||||
@@ -2540,7 +2541,9 @@ def build_storyboard_frame_prompt(project, segment, extra_prompt: str = "") -> s
|
||||
"{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}"
|
||||
"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。"
|
||||
"脚本里的 visual 是本段秒级分镜清单；请画商品用法正确、最能看清商品的那一拍作为关键帧。"
|
||||
"禁止画出违背常识的用法（例如把茶包丢进冷白开、悬浮的手）。{补充}"
|
||||
)
|
||||
rendered = render_prompt(
|
||||
"storyboard_frame", default,
|
||||
@@ -2574,7 +2577,11 @@ def build_video_segment_prompt(project, video_segment, scene, refs, user_prompt:
|
||||
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。"
|
||||
default = (
|
||||
"{设定}{分镜}【脚本】{脚本}\n"
|
||||
"严格按照脚本里的秒级分镜切换景别和动作，商品用法必须真实，不要诡异姿势或错误容器。"
|
||||
"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 ""),
|
||||
@@ -2744,7 +2751,9 @@ def build_storyboard_frame_prompt_refs(project, segment, refs: list[dict], extra
|
||||
"{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}"
|
||||
"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。"
|
||||
"脚本里的 visual 是本段秒级分镜清单；请画商品用法正确、最能看清商品的那一拍作为关键帧。"
|
||||
"禁止画出违背常识的用法（例如把茶包丢进冷白开、悬浮的手）。{补充}"
|
||||
)
|
||||
rendered = render_prompt(
|
||||
"storyboard_frame", default,
|
||||
|
||||
@@ -322,3 +322,11 @@ class GenerationErrorClassifierTests(SimpleTestCase):
|
||||
self.assertEqual(input_error.code, "invalid_input")
|
||||
self.assertEqual(processing_error.code, "processing_failed")
|
||||
self.assertNotIn("storage", processing_error.fallback_message)
|
||||
|
||||
def test_short_script_copy_is_processing_failed_not_unknown(self):
|
||||
error = classify_generation_error(
|
||||
ValueError("第 1 镜旁白太短,口播至少要说到 36 字才能撑满 15 秒"),
|
||||
operation="script_generate",
|
||||
)
|
||||
self.assertEqual(error.code, "processing_failed")
|
||||
self.assertNotEqual(error.code, "unknown")
|
||||
|
||||
@@ -1,32 +1,42 @@
|
||||
"""二期第2段:时长自由化 + 表现形式 × 视频结构 的纯函数单测(不碰 DB / 不调模型)。
|
||||
"""主流程:每镜固定 15 秒 + 表现形式 × 视频结构 的纯函数单测(不碰 DB / 不调模型)。
|
||||
|
||||
覆盖三件容易悄悄坏掉的事:
|
||||
1. 切镜结果必须「每镜 4–15 秒」且「加总精确等于总时长」—— 越界下游出片会直接拒。
|
||||
2. 模型给的每镜时长可能是错的(加总对不上/单镜越界),后端必须能修回来而不是原样落库。
|
||||
1. 切镜结果必须「每镜 15 秒」且「加总精确等于总时长」—— 不再允许 8/10/12。
|
||||
2. 模型给的每镜时长可能是错的,后端必须改写成 15×N,而不是原样落库。
|
||||
3. 套路必须按组合**选择性**加载,全量灌进去会让系统提示词翻倍且互相打架。
|
||||
"""
|
||||
from unittest.mock import Mock
|
||||
|
||||
from django.test import SimpleTestCase
|
||||
|
||||
from apps.ai.script_agent import (
|
||||
DEFAULT_TOTAL_DURATION,
|
||||
PRESENTATION_FORMATS,
|
||||
SEGMENT_DURATION_MAX,
|
||||
SEGMENT_DURATION_MIN,
|
||||
TOTAL_DURATION_MAX,
|
||||
TOTAL_DURATION_MIN,
|
||||
VIDEO_STRUCTURES,
|
||||
VISUAL_CHARS_MIN,
|
||||
_fit_segment_durations,
|
||||
_product_facts,
|
||||
allowed_structures,
|
||||
assert_product_facts_used,
|
||||
assert_shot_density,
|
||||
build_agent_messages,
|
||||
coerce_combo,
|
||||
coerce_total_duration,
|
||||
combo_keys,
|
||||
compose_segment_visual,
|
||||
load_ecommerce_skill,
|
||||
narration_floor,
|
||||
narration_limit,
|
||||
normalize_draft,
|
||||
parse_visual_beats,
|
||||
plan_roles,
|
||||
plan_segment_durations,
|
||||
)
|
||||
|
||||
ALL_TOTALS = list(range(TOTAL_DURATION_MIN, TOTAL_DURATION_MAX + 1, 5))
|
||||
ALL_TOTALS = [15, 30, 45, 60]
|
||||
|
||||
|
||||
class TotalDurationCoercionTests(SimpleTestCase):
|
||||
@@ -37,28 +47,24 @@ class TotalDurationCoercionTests(SimpleTestCase):
|
||||
def test_clamped_into_range_and_snapped_to_step(self):
|
||||
self.assertEqual(coerce_total_duration(3), TOTAL_DURATION_MIN)
|
||||
self.assertEqual(coerce_total_duration(999), TOTAL_DURATION_MAX)
|
||||
self.assertEqual(coerce_total_duration(7), 5)
|
||||
self.assertEqual(coerce_total_duration(8), 10)
|
||||
self.assertEqual(coerce_total_duration(7), 15)
|
||||
self.assertEqual(coerce_total_duration(20), 15)
|
||||
self.assertEqual(coerce_total_duration(25), 30)
|
||||
self.assertEqual(coerce_total_duration(45), 45)
|
||||
|
||||
|
||||
class SegmentPlanningTests(SimpleTestCase):
|
||||
def test_every_total_and_format_yields_legal_shots(self):
|
||||
def test_every_total_and_format_yields_fixed_15s_shots(self):
|
||||
for total in ALL_TOTALS:
|
||||
for fmt in PRESENTATION_FORMATS:
|
||||
with self.subTest(total=total, fmt=fmt):
|
||||
durations = plan_segment_durations(total, fmt)
|
||||
self.assertEqual(sum(durations), total)
|
||||
for seconds in durations:
|
||||
self.assertGreaterEqual(seconds, SEGMENT_DURATION_MIN)
|
||||
self.assertLessEqual(seconds, SEGMENT_DURATION_MAX)
|
||||
self.assertEqual(durations, [SEGMENT_DURATION_MAX] * (total // 15))
|
||||
|
||||
def test_faster_format_produces_more_shots(self):
|
||||
# 同样 60 秒,Vlog(碎片)镜数应多于口播(要把话说完)
|
||||
self.assertGreater(
|
||||
len(plan_segment_durations(60, "vlog")),
|
||||
len(plan_segment_durations(60, "oral")),
|
||||
)
|
||||
def test_format_no_longer_changes_shot_count(self):
|
||||
self.assertEqual(plan_segment_durations(60, "vlog"), plan_segment_durations(60, "oral"))
|
||||
self.assertEqual(plan_segment_durations(30, "drama"), [15, 15])
|
||||
|
||||
def test_roles_follow_shot_count(self):
|
||||
self.assertEqual(plan_roles(1), ["钩子"])
|
||||
@@ -76,32 +82,195 @@ class SegmentPlanningTests(SimpleTestCase):
|
||||
|
||||
|
||||
class SegmentDurationFittingTests(SimpleTestCase):
|
||||
def test_legal_uneven_durations_are_kept_as_is(self):
|
||||
# 模型的不等长节奏意图只要合法就不该被抹平
|
||||
self.assertEqual(_fit_segment_durations([12, 8, 10], 30, "oral"), [12, 8, 10])
|
||||
def test_uneven_durations_are_forced_to_15(self):
|
||||
self.assertEqual(_fit_segment_durations([12, 8, 10], 30, "oral"), [15, 15])
|
||||
|
||||
def test_wrong_total_is_repaired(self):
|
||||
fitted = _fit_segment_durations([15, 15, 15, 15], 30, "oral")
|
||||
self.assertEqual(sum(fitted), 30)
|
||||
self.assertTrue(all(SEGMENT_DURATION_MIN <= d <= SEGMENT_DURATION_MAX for d in fitted))
|
||||
def test_too_many_shots_are_collapsed(self):
|
||||
self.assertEqual(_fit_segment_durations([15, 15, 15, 15], 30, "oral"), [15, 15])
|
||||
|
||||
def test_out_of_range_shot_is_repaired(self):
|
||||
fitted = _fit_segment_durations([99, 1], 30, "drama")
|
||||
self.assertEqual(sum(fitted), 30)
|
||||
self.assertTrue(all(SEGMENT_DURATION_MIN <= d <= SEGMENT_DURATION_MAX for d in fitted))
|
||||
self.assertEqual(_fit_segment_durations([99, 1], 30, "drama"), [15, 15])
|
||||
|
||||
def test_empty_falls_back_to_planner(self):
|
||||
self.assertEqual(_fit_segment_durations([], 20, "vlog"), plan_segment_durations(20, "vlog"))
|
||||
self.assertEqual(_fit_segment_durations([], 30, "vlog"), plan_segment_durations(30, "vlog"))
|
||||
|
||||
def test_garbage_values_never_escape_the_legal_range(self):
|
||||
def test_garbage_values_never_escape_15s_grid(self):
|
||||
for raw in ([None, None], ["a", "b"], [0, 0, 0], [3, 3], [40, 40, 40]):
|
||||
for total in (10, 30, 60):
|
||||
for total in ALL_TOTALS:
|
||||
with self.subTest(raw=raw, total=total):
|
||||
fitted = _fit_segment_durations(list(raw), total, "oral")
|
||||
self.assertEqual(sum(fitted), total)
|
||||
for seconds in fitted:
|
||||
self.assertGreaterEqual(seconds, SEGMENT_DURATION_MIN)
|
||||
self.assertLessEqual(seconds, SEGMENT_DURATION_MAX)
|
||||
self.assertTrue(all(d == SEGMENT_DURATION_MAX for d in fitted))
|
||||
|
||||
|
||||
class NormalizeDurationTests(SimpleTestCase):
|
||||
def test_extra_shots_are_truncated_to_duration_grid(self):
|
||||
raw = {
|
||||
"segments": [
|
||||
{"role": "钩子", "narration": "口播1", "visual": "画面1", "duration": 12},
|
||||
{"role": "痛点", "narration": "口播2", "visual": "画面2", "duration": 8},
|
||||
{"role": "CTA", "narration": "口播3", "visual": "画面3", "duration": 10},
|
||||
]
|
||||
}
|
||||
import json
|
||||
|
||||
draft = normalize_draft(json.dumps(raw, ensure_ascii=False), aspect_ratio="9:16", total_duration=30)
|
||||
self.assertEqual(len(draft["segments"]), 2)
|
||||
self.assertEqual([s["duration"] for s in draft["segments"]], [15, 15])
|
||||
|
||||
def test_preserve_layout_keeps_legacy_shot_count(self):
|
||||
import json
|
||||
|
||||
raw = {
|
||||
"segments": [
|
||||
{"role": "钩子", "narration": "口播1", "visual": "画面1"},
|
||||
{"role": "痛点", "narration": "口播2", "visual": "画面2"},
|
||||
{"role": "CTA", "narration": "口播3", "visual": "画面3"},
|
||||
]
|
||||
}
|
||||
draft = normalize_draft(
|
||||
json.dumps(raw, ensure_ascii=False),
|
||||
aspect_ratio="9:16",
|
||||
total_duration=30,
|
||||
preserve_layout={"count": 3, "durations": [12, 8, 10]},
|
||||
)
|
||||
self.assertEqual(len(draft["segments"]), 3)
|
||||
self.assertEqual([s["duration"] for s in draft["segments"]], [12, 8, 10])
|
||||
self.assertEqual(draft["total_duration"], 30)
|
||||
|
||||
|
||||
class ProductFactTests(SimpleTestCase):
|
||||
def test_missing_selling_point_is_rejected(self):
|
||||
with self.assertRaises(ValueError):
|
||||
assert_product_facts_used(
|
||||
{"hook": "夏天好热", "segments": [{"narration": "补水很重要"}]},
|
||||
selling_titles=["控油12小时"],
|
||||
)
|
||||
|
||||
def test_selling_point_original_word_passes(self):
|
||||
assert_product_facts_used(
|
||||
{"hook": "控油12小时真的扛得住", "segments": [{"narration": "晚上八点还不油"}]},
|
||||
selling_titles=["控油12小时"],
|
||||
)
|
||||
|
||||
def test_single_character_selling_point_counts(self):
|
||||
assert_product_facts_used(
|
||||
{"hook": "这杯茶真香", "segments": [{"narration": "下午一杯茶"}]},
|
||||
selling_titles=["茶"],
|
||||
)
|
||||
|
||||
def test_brand_is_not_required(self):
|
||||
assert_product_facts_used(
|
||||
{"hook": "控油12小时", "segments": [{"narration": "控油12小时"}]},
|
||||
brand="AirBrand",
|
||||
selling_titles=["控油12小时"],
|
||||
)
|
||||
|
||||
|
||||
class ProductFactsLookupTests(SimpleTestCase):
|
||||
def _project(self, points):
|
||||
class _QS(list):
|
||||
def filter(self, **kwargs):
|
||||
raise AssertionError("不要再用 id__in 过滤,标题不是 UUID")
|
||||
|
||||
product = Mock()
|
||||
product.selling_points.all.return_value = _QS(points)
|
||||
project = Mock()
|
||||
project.product = product
|
||||
project.metadata = {}
|
||||
return project
|
||||
|
||||
def test_title_keys_match_without_uuid_filter(self):
|
||||
tea = Mock(id="11111111-1111-1111-1111-111111111111", title="茶", detail="茶香")
|
||||
oil = Mock(id="22222222-2222-2222-2222-222222222222", title="控油", detail="x")
|
||||
_product, matched = _product_facts(self._project([tea, oil]), ["茶"])
|
||||
self.assertEqual([item.title for item in matched], ["茶"])
|
||||
|
||||
def test_uuid_keys_still_match(self):
|
||||
tea = Mock(id="11111111-1111-1111-1111-111111111111", title="茶", detail="茶香")
|
||||
oil = Mock(id="22222222-2222-2222-2222-222222222222", title="控油", detail="x")
|
||||
_product, matched = _product_facts(self._project([tea, oil]), [str(tea.id)])
|
||||
self.assertEqual([item.title for item in matched], ["茶"])
|
||||
|
||||
|
||||
class PromptAssemblyTests(SimpleTestCase):
|
||||
def _project(self):
|
||||
sp = Mock(title="控油12小时", detail="到晚八点")
|
||||
|
||||
class _QS(list):
|
||||
def filter(self, **kwargs):
|
||||
return self
|
||||
|
||||
product = Mock()
|
||||
product.title = "玻尿酸精华"
|
||||
product.brand = "AirBrand"
|
||||
product.business_type = "ecommerce"
|
||||
product.category = "美妆"
|
||||
product.target_audience = "油皮"
|
||||
product.description = "夜间修护"
|
||||
product.specs = {"price": "99", "source": "product-create-upload"}
|
||||
product.selling_points.all.return_value = _QS([sp])
|
||||
project = Mock()
|
||||
project.product = product
|
||||
project.metadata = {"wizard": {"persona": "urban", "selling_point_ids": ["1"]}}
|
||||
return project
|
||||
|
||||
def test_auto_mode_with_prompt_is_not_theme(self):
|
||||
messages = build_agent_messages(
|
||||
project=self._project(),
|
||||
mode="auto",
|
||||
user_prompt="节奏紧凑,适合短视频投放",
|
||||
selling_point_ids=["1"],
|
||||
base_draft=None,
|
||||
aspect_ratio="9:16",
|
||||
total_duration=30,
|
||||
persona="urban",
|
||||
)
|
||||
user = messages[1]["content"]
|
||||
self.assertIn("全自动", user)
|
||||
self.assertNotIn("一句话主题扩写", user)
|
||||
self.assertIn("【补充要求】节奏紧凑,适合短视频投放", user)
|
||||
self.assertIn("控油12小时", user)
|
||||
self.assertIn("都市白领女性", user)
|
||||
self.assertIn("价格:99", user)
|
||||
self.assertIn("每镜必须 15 秒", user)
|
||||
self.assertIn("一共 2 镜", user)
|
||||
self.assertIn("必须原词用上", user)
|
||||
self.assertIn("跟同事吐槽", user)
|
||||
self.assertIn("禁止一句 20 字收工", user)
|
||||
self.assertIn("秒级分镜", user)
|
||||
|
||||
def test_scene_oral_prompt_locks_skeleton_and_density(self):
|
||||
messages = build_agent_messages(
|
||||
project=self._project(),
|
||||
mode="auto",
|
||||
user_prompt="",
|
||||
selling_point_ids=["1"],
|
||||
base_draft=None,
|
||||
aspect_ratio="9:16",
|
||||
total_duration=45,
|
||||
presentation_format="oral",
|
||||
video_structure="scene",
|
||||
persona="urban",
|
||||
)
|
||||
user = messages[1]["content"]
|
||||
self.assertIn("场景种草骨架", user)
|
||||
self.assertIn("口播每镜必须说满", user)
|
||||
self.assertIn("一共 3 镜", user)
|
||||
self.assertIn("秒级分镜", user)
|
||||
|
||||
def test_theme_mode_still_exists(self):
|
||||
messages = build_agent_messages(
|
||||
project=self._project(),
|
||||
mode="theme",
|
||||
user_prompt="主打熬夜党",
|
||||
selling_point_ids=None,
|
||||
base_draft=None,
|
||||
aspect_ratio="9:16",
|
||||
total_duration=15,
|
||||
)
|
||||
self.assertIn("一句话主题扩写", messages[1]["content"])
|
||||
self.assertIn("主打熬夜党", messages[1]["content"])
|
||||
|
||||
|
||||
class NarrationLimitTests(SimpleTestCase):
|
||||
@@ -113,6 +282,107 @@ class NarrationLimitTests(SimpleTestCase):
|
||||
def test_never_exceeds_hard_cap(self):
|
||||
self.assertLessEqual(narration_limit(60), 55)
|
||||
|
||||
def test_fifteen_second_floor_is_about_thirty_six(self):
|
||||
self.assertEqual(narration_floor(15), 36)
|
||||
self.assertLess(narration_floor(15), narration_limit(15))
|
||||
|
||||
|
||||
class ShotDensityTests(SimpleTestCase):
|
||||
_VISUAL = (
|
||||
"0-3s：近景，女主对镜头抬眼，工位键盘还亮着，窗外是下午的光\n"
|
||||
"3-8s：手持跟到桌面，右手把茶包放进盛了热水的玻璃杯，水面起蒸汽\n"
|
||||
"8-12s：杯壁特写，茶汤从浅金慢慢变深，标签贴在杯沿\n"
|
||||
"12-15s：拉回中近景，女主喝一口，眼神松下来"
|
||||
)
|
||||
|
||||
def test_short_oral_narration_is_rejected(self):
|
||||
with self.assertRaises(ValueError) as ctx:
|
||||
assert_shot_density(
|
||||
{
|
||||
"segments": [
|
||||
{
|
||||
"duration": 15,
|
||||
"narration": "下午三点来杯100numur茶解乏。",
|
||||
"visual": self._VISUAL,
|
||||
}
|
||||
]
|
||||
},
|
||||
"oral",
|
||||
)
|
||||
self.assertIn("旁白太短", str(ctx.exception))
|
||||
|
||||
def test_oral_meeting_floor_passes(self):
|
||||
narration = "下午三点工位犯困，键盘都敲不利索。我倒了杯茶，第一口是回甘不是苦，整个人才慢慢醒过来。"
|
||||
self.assertGreaterEqual(len(narration.replace(" ", "")), narration_floor(15))
|
||||
self.assertGreaterEqual(len(self._VISUAL.replace(" ", "")), VISUAL_CHARS_MIN)
|
||||
assert_shot_density(
|
||||
{"segments": [{"duration": 15, "narration": narration, "visual": self._VISUAL}]},
|
||||
"oral",
|
||||
)
|
||||
|
||||
def test_short_visual_is_rejected(self):
|
||||
with self.assertRaises(ValueError) as ctx:
|
||||
assert_shot_density(
|
||||
{
|
||||
"segments": [
|
||||
{
|
||||
"duration": 15,
|
||||
"narration": "下午三点工位犯困，键盘都敲不利索。我倒了杯茶，第一口是回甘不是苦，整个人才慢慢醒过来。",
|
||||
"visual": "女主举起杯子。",
|
||||
}
|
||||
]
|
||||
},
|
||||
"oral",
|
||||
)
|
||||
self.assertIn("画面描写太短", str(ctx.exception))
|
||||
|
||||
def test_vlog_allows_silent_shot(self):
|
||||
assert_shot_density(
|
||||
{"segments": [{"duration": 15, "narration": "", "visual": self._VISUAL}]},
|
||||
"vlog",
|
||||
)
|
||||
|
||||
def test_paragraph_visual_without_timestamps_is_rejected(self):
|
||||
narration = "下午三点工位犯困，键盘都敲不利索。我倒了杯茶，第一口是回甘不是苦，整个人才慢慢醒过来。"
|
||||
long_visual = (
|
||||
"近景对镜头，女主从键盘抬眼，端起杯子喝一口，再切杯壁特写，蒸汽从杯口升起，"
|
||||
"茶汤慢慢变深，窗外下午的光打在桌面上，手没有离开杯柄，整段都不写起止秒。"
|
||||
)
|
||||
self.assertGreaterEqual(len(long_visual.replace(" ", "")), VISUAL_CHARS_MIN)
|
||||
with self.assertRaises(ValueError) as ctx:
|
||||
assert_shot_density(
|
||||
{"segments": [{"duration": 15, "narration": narration, "visual": long_visual}]},
|
||||
"oral",
|
||||
)
|
||||
self.assertIn("按秒拆分镜", str(ctx.exception))
|
||||
|
||||
def test_beats_array_flattens_into_visual(self):
|
||||
import json
|
||||
|
||||
raw = {
|
||||
"segments": [
|
||||
{
|
||||
"role": "钩子",
|
||||
"narration": "下午三点工位犯困，键盘都敲不利索。",
|
||||
"beats": [
|
||||
{"start": 0, "end": 3, "shot_size": "近景", "action": "女主对镜头抬眼，键盘还亮着"},
|
||||
{"start": 3, "end": 8, "shot_size": "手持", "action": "右手把茶包放进热水杯，蒸汽升起"},
|
||||
{"start": 8, "end": 12, "shot_size": "特写", "action": "茶汤从浅金慢慢变深"},
|
||||
{"start": 12, "end": 15, "shot_size": "中近景", "action": "女主喝一口眼神松开"},
|
||||
],
|
||||
}
|
||||
]
|
||||
}
|
||||
draft = normalize_draft(json.dumps(raw, ensure_ascii=False), aspect_ratio="9:16", total_duration=15)
|
||||
visual = draft["segments"][0]["visual"]
|
||||
self.assertIn("0-3s：", visual)
|
||||
self.assertIn("近景", visual)
|
||||
self.assertGreaterEqual(len(parse_visual_beats(visual)), 3)
|
||||
|
||||
def test_compose_keeps_timestamped_visual(self):
|
||||
visual = compose_segment_visual({"visual": self._VISUAL})
|
||||
self.assertGreaterEqual(len(parse_visual_beats(visual)), 3)
|
||||
|
||||
|
||||
class ComboTests(SimpleTestCase):
|
||||
def test_drama_cannot_pick_review(self):
|
||||
@@ -160,3 +430,9 @@ class SkillLoadingTests(SimpleTestCase):
|
||||
def test_skill_text_is_substantial(self):
|
||||
# skills 没随镜像打进去时会退化成一句兜底,这里守住「提示词没丢」
|
||||
self.assertGreater(len(load_ecommerce_skill("oral", "pain")), 5000)
|
||||
|
||||
def test_skill_now_requires_fixed_15s(self):
|
||||
text = load_ecommerce_skill("oral", "pain")
|
||||
self.assertIn("每镜固定 15 秒", text)
|
||||
self.assertIn("商品原词", text)
|
||||
self.assertNotIn("单镜 4–15 秒，允许不等长", text)
|
||||
|
||||
@@ -518,14 +518,13 @@ class ProjectViewSet(TeamScopedViewSetMixin, ModelViewSet):
|
||||
@action(detail=True, methods=["post"], url_path="script-agent-stream", renderer_classes=[ServerSentEventRenderer])
|
||||
def script_agent_stream(self, request, pk=None):
|
||||
"""对话式脚本 agent · 流式(SSE)。出稿 + 改稿一体,多模型可选。
|
||||
请求体:mode(auto|theme|revise)、prompt、model_config_id、selling_point_ids、
|
||||
base_version_id(改稿)、aspect_ratio、total_duration(5-60,5 秒步进)、
|
||||
请求体:mode(auto|theme|revise)、prompt、model_config_id、selling_point_ids、persona、
|
||||
base_version_id(改稿)、aspect_ratio、total_duration(15/30/45/60)、
|
||||
presentation_format(oral|drama|vlog)、video_structure(pain|contrast|review|scene)。
|
||||
响应:text/event-stream,逐帧吐 tool/delta/draft/saved/done/error。"""
|
||||
project = self.get_object()
|
||||
mode = str(request.data.get("mode") or "auto")
|
||||
prompt = str(request.data.get("prompt") or "")
|
||||
selling_point_ids = request.data.get("selling_point_ids") or []
|
||||
base_version_id = request.data.get("base_version_id") or None
|
||||
aspect_ratio = str(request.data.get("aspect_ratio") or "9:16")
|
||||
# 非法值一律由 agent 侧 coerce 兜底(夹区间/回落默认),这里不做 400,避免生成被参数噪声打断
|
||||
@@ -534,6 +533,9 @@ class ProjectViewSet(TeamScopedViewSetMixin, ModelViewSet):
|
||||
request.data.get("presentation_format"),
|
||||
request.data.get("video_structure"),
|
||||
)
|
||||
wizard = ((project.metadata or {}).get("wizard") if isinstance(project.metadata, dict) else None) or {}
|
||||
selling_point_ids = request.data.get("selling_point_ids") or wizard.get("selling_point_ids") or []
|
||||
persona = str(request.data.get("persona") or wizard.get("persona") or "")
|
||||
target_index = request.data.get("target_index")
|
||||
try:
|
||||
target_index = int(target_index) if target_index is not None else None
|
||||
@@ -568,6 +570,7 @@ class ProjectViewSet(TeamScopedViewSetMixin, ModelViewSet):
|
||||
video_structure=video_structure,
|
||||
target_index=target_index,
|
||||
entry_source=str(request.data.get("source") or ""),
|
||||
persona=persona,
|
||||
)
|
||||
response = StreamingHttpResponse(stream, content_type="text/event-stream")
|
||||
response["Cache-Control"] = "no-cache"
|
||||
@@ -894,8 +897,8 @@ class ProjectViewSet(TeamScopedViewSetMixin, ModelViewSet):
|
||||
"""采用版分镜数变化时,同步 VideoSegment 的**数量和时长**:不足则尾部补 NOT_STARTED,
|
||||
多出且尾部是「从未生成过」的段则裁掉(已生成的段绝不动)。
|
||||
|
||||
时长同步是二期补的:脚本镜可以不等长了,而出片、计价、时间线读的都是
|
||||
VideoSegment.target_duration_seconds。不同步的话脚本写了 8 秒、出片仍按 15 秒跑。
|
||||
时长同步:脚本镜主流程固定 15 秒,而出片、计价、时间线读的都是
|
||||
VideoSegment.target_duration_seconds。不同步的话脚本写了 15 秒、出片仍按旧默认跑。
|
||||
已出片的段不改时长——改了会跟已渲染的成片对不上。"""
|
||||
if not script.is_adopted:
|
||||
return
|
||||
|
||||
@@ -5,7 +5,7 @@ description: >
|
||||
服务对象不是人类编剧，而是 AirShelf 产品后端的「脚本生成 agent」——在运行时按需加载本技能作为领域知识。
|
||||
能力：把【商品信息 + 前置条件】或【一句话主题】或【用户已有脚本】，
|
||||
自动收敛成一份结构化的带货短视频脚本 JSON
|
||||
（默认 9:16 竖屏可改 / 总时长 5–60 秒按 5 秒步进 / 单镜 4–15 秒可不等长 /
|
||||
（默认 9:16 竖屏可改 / 总时长 15/30/45/60 秒 / 每镜固定 15 秒 /
|
||||
按「表现形式 × 视频结构」套路生成）。
|
||||
当任务为「生成带货脚本 / 扩写主题 / 优化已有脚本 / 商品转视频脚本」时使用本技能。
|
||||
核心目标是「电商小白点一下按钮就出能吸睛、能转化的脚本」，不是影视级艺术性。
|
||||
@@ -57,11 +57,11 @@ description: >
|
||||
"segments": [
|
||||
{
|
||||
"index": 0,
|
||||
"duration": 12,
|
||||
"duration": 15,
|
||||
"role": "钩子|痛点|卖点|CTA",
|
||||
"narration": "这一镜被说出来的台词/旁白,字数上限=duration×3.5",
|
||||
"narration": "这一镜被说出来的台词/旁白,15秒口播目标38-47字(下限36,上限52)",
|
||||
"speaker": "可选,指向某 entity 的 id;画外旁白时为 null",
|
||||
"visual": "这一镜的画面:主体+动作+景别/运镜(特写/全景/手持跟拍/推拉摇)+一个画面或情绪的变化,够导演撑满这一镜的秒数,约40-70字,别只写一句静态动作",
|
||||
"visual": "0-3s：近景，……\\n3-8s：手持跟拍，开始使用商品\\n8-12s：特写，商品用法过程\\n12-15s：拉回中近景，使用后的反应",
|
||||
"product_exposure": "商品露出方式(手持/特写/使用中)",
|
||||
"entity_refs": ["c1", "s1"],
|
||||
"dialogue": []
|
||||
@@ -76,11 +76,12 @@ description: >
|
||||
`video_structure`（`痛点解决|前后对比|测评验证|场景种草`）**原样回填到输出里**，不要自己改。
|
||||
输入没给时，按商品品类与人群自行选一组最合适的，并如实填进这两个字段。
|
||||
- **画幅由输入给定**：`aspect_ratio` 默认 `"9:16"`（电商竖屏主场景），但**不写死**——输入指定了其他比例（如 `"16:9"`、`"1:1"`、`"4:5"`）就照用，原样透传给下游。画幅只影响 `visual` 的构图措辞，不改变结构与镜数。
|
||||
- **总时长由输入给定**：`total_duration` 取 **5–60 秒之间的 5 的倍数**（5/10/15/…/60），输入未指定时默认 `30`。**不要写死。**
|
||||
- **单镜时长 4–15 秒，允许不等长**：每个 `segments[].duration` 必须是 **4 到 15 之间的整数**
|
||||
（15 秒是下游出片模型的硬上限，越界下游直接拒绝出片）。
|
||||
该长的镜给足、该短的镜压短，**不要机械均分**。
|
||||
- **总时长由输入给定**：`total_duration` 取 **15 / 30 / 45 / 60**（15 秒步进），输入未指定时默认 `30`。**不要写死成别的数。**
|
||||
- **每镜固定 15 秒**：每个 `segments[].duration` **必须是 15**，镜数 = `total_duration / 15`（15 秒 1 镜、30 秒 2 镜、45 秒 3 镜、60 秒 4 镜）。
|
||||
禁止 8/10/12 这种不等长，禁止自行加减镜。15 秒也是下游出片模型的硬上限。
|
||||
- **镜时长必须精确加总**：`sum(segments[].duration) == total_duration`，一秒都不能差。
|
||||
- **商品原词铁律**：钩子/旁白/对白必须用上输入里给出的**商品名、品牌、本次勾选卖点的原词**。
|
||||
禁止用「补水 / 好用 / 值得买 / 宝藏 / 闭眼入」这类空话替换真实卖点。品类话术和钩子库只是写法参考，不是正文。
|
||||
- **`segment_count` = `segments` 的实际长度**，且 `index` 从 0 连续递增。
|
||||
- 镜数怎么定、role 怎么按镜数分配，见 `references/methodology.md`「时长 → 镜数 → 黄金结构映射」。
|
||||
- `entities[].id` 全局唯一，`segments[].entity_refs` 与 `speaker` 只能引用已声明的 id。
|
||||
@@ -94,7 +95,11 @@ description: >
|
||||
- **每个声明的 entity 至少被一个 segment 引用**（不留孤儿 entity）。
|
||||
- **场景必抽，且每镜必绑一个场景**：每条脚本**至少声明 1 个 `type:"scene"` 实体**表示画面所在环境；**每个 segment 的 `entity_refs` 必须恰好引用一个 scene**。多镜在同一环境就**复用同一个 scene id**（绝不为同一环境写两份 visual_prompt，否则下游背景漂移）；只有真正换了环境才新建另一个 scene。纯产品特写镜也要绑它所处环境的 scene（如「宿舍书桌」「厨房台面」），没有合适环境时复用主场景。
|
||||
- `visual_prompt` 由你自动生成，小白无需打字。
|
||||
- **每镜 `visual` 要够厚撑满这一镜的秒数**：一段话写清 ①主体+动作 ②景别/运镜（特写/全景/手持跟拍/推拉摇，至少给一个镜头语言）③一个画面或情绪的变化（从…到…）。约 **40–70 字**，**禁止只写一句静态动作**（如「女主举起商品展示」撑不住十几秒，要补镜头与变化）。4–6 秒的短镜可以只给一个动作 + 一个镜头语言，但仍要写清景别。注意：这是给生图/视频导演的画面，**不占 narration 的字数额度**。
|
||||
- **每镜 `visual` 必须按秒拆分镜**：15 秒一场里至少 3 刀、目标 4 刀，写成多行：
|
||||
`0-3s：景别，谁在做什么，手和商品的空间关系`
|
||||
第一条从 0 秒起，最后一条接到 15s。每条都要有景别/运镜，15 秒内至少切两次景别。
|
||||
**禁止**用一句静态动作撑满 15 秒（如「女主举起商品展示」）。这是给故事板和 Seedance 的导演说明书，**不占 narration 字数**。
|
||||
- **画面物理常识**：商品用法必须是真人会做的。茶/咖啡 = 热水、蒸汽、茶汤渐染，禁止茶包丢进看起来像冷白开的杯子；护肤品 = 挤出/涂抹；食品 = 打开/入口。手从真实方向入画，禁止悬浮肢体、反关节、商品凭空出现。每条分镜写清容器/包装长什么样、液体或材质当前状态。
|
||||
- 不要输出 schema 之外的字段，也不要省略必填字段。
|
||||
|
||||
### 铁律 2 · 输出前自检
|
||||
@@ -106,10 +111,12 @@ description: >
|
||||
|
||||
详见 `references/methodology.md`「旁白红线」。最关键的几条：
|
||||
- **口语化**，像真人对着镜头说话，不准书面腔 / AI 腔。
|
||||
- **每镜 narration ≤ `duration × 3.5` 字，且绝不超过 55 字**（出片模型在镜内直接发声，
|
||||
3.5 字/秒是可懂语速上限）。短镜装不下就拆到下一镜，或干脆留空走纯画面。
|
||||
- **口播 15 秒镜要说满**：`narration` **36–52 字**（写作目标 **38–47 字，2–4 句短句**）。
|
||||
上限仍是 `duration × 3.5` 且不超过 55 字。禁止一句 20 字收工。
|
||||
商品名全片点名 1–2 次即可，其余镜用卖点原词和可感知细节（口感/气味/动作/使用场景）。
|
||||
Vlog 才允许个别镜纯画面；口播禁止整镜无声。
|
||||
- **禁违规词**：医疗功效（治疗/根治/抗癌…）、绝对化用语（最/第一/100%/国家级…）一律不写。
|
||||
- 不浮夸、不空喊，卖点要落到「商品怎么解决痛点」。
|
||||
- 不浮夸、不空喊，卖点要落到「商品怎么解决痛点」。人物口吻跟输入的人物设定走，不要万能主播腔。
|
||||
|
||||
### 铁律 4 · 一键自动化（与影视母版相反）
|
||||
|
||||
@@ -124,10 +131,10 @@ description: >
|
||||
| 参数 | 值 |
|
||||
| ---- | ---- |
|
||||
| 画幅 | **默认 9:16 竖屏**，由输入可覆盖（16:9 / 1:1 / 4:5 等照用） |
|
||||
| 总时长 | **5–60 秒，5 秒步进**，由输入给定；未指定时默认 30 |
|
||||
| 总时长 | **15 / 30 / 45 / 60 秒**，由输入给定；未指定时默认 30 |
|
||||
| 表现形式 | **口播 / 短剧 / Vlog**，由输入给定；未指定时按品类人群自选 |
|
||||
| 视频结构 | **痛点解决 / 前后对比 / 测评验证 / 场景种草**，由输入给定；未指定时按品类人群自选 |
|
||||
| 分镜 | **单镜 4–15 秒，可不等长**；镜数按表现形式的推荐节奏定（见方法论） |
|
||||
| 分镜 | **每镜固定 15 秒**；镜数 = 总时长 / 15 |
|
||||
| 镜头功能 | 钩子 → 痛点 → 卖点 → CTA（黄金结构；具体骨架以视频结构套路为准） |
|
||||
| 发声方式 | 出片模型直接生成画面+音效+人声（**不走 TTS**） |
|
||||
|
||||
@@ -163,13 +170,14 @@ description: >
|
||||
## 生成流程（内部执行，一次走完，不暂停）
|
||||
|
||||
1. **路由** — 判定输入模式（①/②/③），加载对应 references；再按输入的表现形式与视频结构加载那两份套路。
|
||||
2. **定镜** — 读画幅（`aspect_ratio` 默认 9:16）与总时长（5–60 秒，未指定默认 30）；
|
||||
按表现形式的推荐单镜节奏定出镜数，再给每镜分配 4–15 秒的时长，**确保加总等于总时长**。
|
||||
2. **定镜** — 读画幅（`aspect_ratio` 默认 9:16）与总时长（15/30/45/60，未指定默认 30）；
|
||||
镜数 = 总时长 / 15，**每镜 duration 一律填 15**。
|
||||
3. **定调（tone）** — 依据视频结构 + 品类话术 + 平台调性，选定 `tone`；②③ 模式尊重用户已表达的倾向。
|
||||
4. **抽取/创建 entities** — 识别脚本需要的角色 / 场景 / 商品；为每个 entity 写一份**全脚本共用**的 `visual_prompt`（保证多镜同一角色同一张脸）；可选写 `voice_ref` 锁音色。
|
||||
5. **按套路填结构** — 优先用视频结构套路里的骨架给每个 segment 分配 `role`；
|
||||
套路没覆盖到的用「role 按镜数分配」表兜底；钩子镜套用该套路指定的钩子写法。
|
||||
6. **写 narration / visual / 商品露出** — 每镜旁白 ≤ `duration × 3.5` 字、口语化、过红线；
|
||||
6. **写 narration / visual / 商品露出** — 口播 15 秒镜旁白 38–47 字（下限 36）、口语化、过红线；
|
||||
**visual 按秒拆成 3–5 条分镜**；必须把输入给出的卖点原词和商品描述细节写进旁白或对白，不要每句只重复商品名；
|
||||
发声方式按表现形式套路来；每镜规划自然的 `product_exposure`。
|
||||
7. **连引用** — 填 `entity_refs` 与 `speaker`，确认每个 entity 都被引用、id 都合法。
|
||||
8. **自检** — 跑 `checklist.md`，过了再输出。
|
||||
|
||||
@@ -2,6 +2,7 @@
|
||||
|
||||
> 全自动 / 一句话模式读。根据商品所属品类，调整语气与卖点侧重。
|
||||
> 不确定品类时，按商品功能就近归类；都不沾就用"通用"原则（讲场景痛点 + 真实使用感）。
|
||||
> **示例旁白只学语气，禁止把例句里的假商品/假卖点抄进正文。** 正文必须用输入给的商品名和卖点原词。
|
||||
|
||||
每个品类给：**语气** / **卖点侧重** / **常用露出** / **违规雷区** / **示例旁白**。
|
||||
|
||||
|
||||
@@ -13,11 +13,11 @@
|
||||
- [ ] `presentation_format` ∈ `{口播, 短剧, Vlog}`，且与输入一致（输入没给则自选并如实填写）。
|
||||
- [ ] `video_structure` ∈ `{痛点解决, 前后对比, 测评验证, 场景种草}`，同上。
|
||||
- [ ] **组合合法**：不是「短剧 × 测评验证」（唯一禁用组合，见 `playbooks/combo-matrix.md`）。
|
||||
- [ ] `total_duration` 是 **5–60 之间 5 的倍数**。
|
||||
- [ ] **每个 `segment.duration` 是 4–15 之间的整数**（15 是下游出片硬上限，越界直接出不了片）。
|
||||
- [ ] `total_duration` 是 **15 / 30 / 45 / 60**。
|
||||
- [ ] **每个 `segment.duration` 必须是 15**。
|
||||
- [ ] **`sum(segments[].duration) == total_duration`**，逐个加一遍，差一秒都不行。
|
||||
- [ ] `segment_count == len(segments)`；`index` 从 0 连续递增无跳号。
|
||||
- [ ] 镜数落在 `ceil(total_duration/15)` 到 `floor(total_duration/4)` 区间内。
|
||||
- [ ] 镜数 **等于** `total_duration / 15`。
|
||||
- [ ] 每个 `segment.role` ∈ `{钩子, 痛点, 卖点, CTA}`。
|
||||
- [ ] 首镜 `role == 钩子`；末镜收 CTA（独立 CTA 镜，或末镜旁白末尾含明确行动指令）。
|
||||
- [ ] `entities` 每项 `type` ∈ `{character, scene, product}`，`id` 全局唯一。
|
||||
@@ -46,10 +46,11 @@
|
||||
|
||||
## C. 旁白红线扫描（逐镜）
|
||||
|
||||
- [ ] **每镜 `narration` ≤ `duration × 3.5` 字，且绝不超过 55 字**（逐镜按自己的秒数数一遍，别凭感觉。
|
||||
4 秒镜 ≤14 字、8 秒镜 ≤28 字、15 秒镜 ≤52 字，对照表见 `methodology.md`）。
|
||||
- [ ] 每镜 `visual` 含「景别/运镜 + 一个画面变化」，不是一句静态动作；能撑满这一镜的秒数
|
||||
（≥8 秒的镜约 40–70 字；4–6 秒的短镜可以更短，但景别必须写清）。
|
||||
- [ ] **每镜 `narration` ≤ `duration × 3.5` 字，且绝不超过 55 字**（15 秒镜 ≤52 字）。
|
||||
- [ ] **口播 15 秒镜旁白 ≥ 36 字**（写作目标 38–47 字，2–4 句）。禁止一句 20 字收工；口播禁止整镜无声。
|
||||
- [ ] **秒级分镜**：每镜 `visual` 至少 3 条 `0-3s：景别，动作`，从 0 接到 15s；15 秒内至少切两次景别。
|
||||
- [ ] **商品原词**：旁白/对白出现了输入给出的至少一个勾选卖点原词；商品名全片点名 1–2 次即可，其余镜写描述/卖点细节，没有用空话替换。
|
||||
- [ ] 每镜 `visual` 能撑满 15 秒（至少 72 字），写清手/商品/容器的空间关系和真实用法，不是一句静态动作。
|
||||
- [ ] 口语化，无书面腔/AI 腔（"综上""不仅…而且""值得一提"等已清除）。
|
||||
- [ ] **违规词扫描**：无医疗功效词（治疗/根治/抗癌/消炎/排毒/速效…）。
|
||||
- [ ] **绝对化用语扫描**：无 最/第一/唯一/100%/国家级/永久/绝对/史上 等。
|
||||
|
||||
@@ -24,30 +24,18 @@
|
||||
|
||||
## 二、时长 → 镜数 → 黄金结构映射（关键）
|
||||
|
||||
### 2.1 时长规则（不再是固定四档，也不再每镜等长）
|
||||
### 2.1 时长规则（每镜固定 15 秒）
|
||||
|
||||
- **总时长**由输入给定：**5–60 秒之间、5 秒一档**（5/10/15/…/60）。未指定时默认 30。
|
||||
- **单镜时长 4–15 秒**，这是下游出片模型的硬上限，**任何一镜都不许超过 15 秒、不许低于 4 秒**。
|
||||
- **每镜时长可以不一样。** 该长的镜（讲卖点、举证）就给足，该短的镜（空镜、转场、一句钩子）就压短。
|
||||
别再机械均分。
|
||||
- **总时长**由输入给定：**15 / 30 / 45 / 60**。未指定时默认 30。
|
||||
- **每镜时长固定 15 秒**，这既是主流程产品规则，也是下游出片模型的硬上限。
|
||||
- 镜数直接由总时长决定：15 秒 → 1 镜，30 秒 → 2 镜，45 秒 → 3 镜，60 秒 → 4 镜。
|
||||
- **不要**写成 8/10/12 这种不等长，也**不要**自行加减镜。15 秒要靠画面变化和一句完整意思撑满，不是靠切碎镜头。
|
||||
- 所有 `segments[].duration` 相加**必须精确等于** `total_duration`。
|
||||
|
||||
### 2.2 镜数怎么定
|
||||
|
||||
镜数不是算出来的，是**按表现形式的节奏定**的。先取该形式的推荐单镜时长，再用总时长去除：
|
||||
|
||||
| 表现形式 | 推荐单镜时长 | 理由 |
|
||||
| ---- | ---- | ---- |
|
||||
| **口播** | 10–15 秒 | 说完一个完整意思需要时间，切太碎话会断 |
|
||||
| **短剧** | 6–10 秒 | 靠切换制造节奏，镜头长了拖沓 |
|
||||
| **Vlog** | 5–8 秒 | 碎片拼贴感，镜头短、数量多 |
|
||||
|
||||
`segment_count ≈ total_duration / 推荐单镜时长`，向最接近的整数取整，
|
||||
然后**必须**落在 `ceil(total_duration/15)` 到 `floor(total_duration/4)` 这个区间内。
|
||||
|
||||
例：口播 30 秒 → 30/12 ≈ 2.5 → 3 镜（如 12+10+8）。
|
||||
短剧 45 秒 → 45/8 ≈ 5.6 → 6 镜（如 8+7+9+7+8+6）。
|
||||
Vlog 30 秒 → 30/6.5 ≈ 4.6 → 5 镜（如 6+6+7+6+5）。
|
||||
镜数**不是按表现形式估算的**，就是 `total_duration / 15`。
|
||||
表现形式只影响这一镜里怎么拍（口播把 15 秒说完一件事；短剧在 15 秒里完成一次冲突/对白交换；Vlog 用镜内空镜+细节撑满），**不改变镜数**。
|
||||
|
||||
### 2.3 role 按镜数分配
|
||||
|
||||
@@ -102,22 +90,23 @@
|
||||
## 五、旁白红线（硬规则，违反即不合格）
|
||||
|
||||
- **口语化**：像真人对着镜头唠嗑，不准书面腔 / 不准 AI 腔（"综上所述""不仅…而且""值得一提的是"全禁）。
|
||||
- **`narration` 字数按这一镜的秒数算**：出片模型在镜内直接发声，可懂语速上限约 **3.5 字 / 秒**。
|
||||
**每镜 `narration` ≤ `duration × 3.5` 字，且任何情况下不超过 55 字。**
|
||||
对照表（写作目标再留 10% 缓冲，宁可短不要卡满）：
|
||||
- **必须用商品原词**：输入给了商品名、品牌、卖点，就用这些词，不许改成「补水/好用/值得买」这种空卖点。品类示例句只是语气参考。
|
||||
- **`narration` 字数按这一镜的秒数算**：出片模型在镜内直接发声。
|
||||
可懂语速上限约 **3.5 字 / 秒**，口播下限约 **2.4 字 / 秒**（再短镜头空一半）。
|
||||
**每镜上限 `duration × 3.5` 字，且不超过 55 字。口播 15 秒镜还有下限 36 字。**
|
||||
|
||||
| 单镜时长 | 字数上限 | 写作目标 |
|
||||
| ---- | ---- | ---- |
|
||||
| 4 秒 | 14 字 | ≤ 12 字 |
|
||||
| 5 秒 | 17 字 | ≤ 15 字 |
|
||||
| 6 秒 | 21 字 | ≤ 19 字 |
|
||||
| 8 秒 | 28 字 | ≤ 25 字 |
|
||||
| 10 秒 | 35 字 | ≤ 31 字 |
|
||||
| 12 秒 | 42 字 | ≤ 38 字 |
|
||||
| 15 秒 | 52 字 | ≤ 47 字 |
|
||||
| 单镜时长 | 下限 | 字数上限 | 写作目标 |
|
||||
| ---- | ---- | ---- | ---- |
|
||||
| 4 秒 | — | 14 字 | ≤ 12 字 |
|
||||
| 5 秒 | — | 17 字 | ≤ 15 字 |
|
||||
| 6 秒 | — | 21 字 | ≤ 19 字 |
|
||||
| 8 秒 | — | 28 字 | ≤ 25 字 |
|
||||
| 10 秒 | — | 35 字 | ≤ 31 字 |
|
||||
| 12 秒 | — | 42 字 | ≤ 38 字 |
|
||||
| 15 秒 | **36 字** | 52 字 | **38–47 字，2–4 句** |
|
||||
|
||||
**短镜不要硬塞长句** —— 装不下就把这句拆到下一镜，或者干脆让这镜没有旁白（纯画面）。
|
||||
总时长很短、一镜要装多功能时尤其要狠删，先保 CTA 不被砍。
|
||||
主流程每镜都是 15 秒：**要说满，不要宁可短。** 一句 20 字撑不满镜头。
|
||||
留白靠句间停顿和画面，不靠少写字。只有 Vlog 才允许个别镜纯画面无声。
|
||||
- **不浮夸、不空喊**：卖点要给具体理由，不堆形容词。
|
||||
- **违规词禁令**（电商广告法红线，一律不写）：
|
||||
- **医疗功效类**：治疗 / 根治 / 疗效 / 抗癌 / 消炎 / 杀菌(无证) / 排毒 / 速效 / 抑制 ××病 …
|
||||
@@ -126,3 +115,22 @@
|
||||
- 替换策略：用"帮助""更""不少人反馈""上脸清爽"等柔性、主观化表达替代。
|
||||
|
||||
> 自检阶段必须逐镜扫一遍违规词，命中即改写。详见 `checklist.md`。
|
||||
|
||||
---
|
||||
|
||||
## 六、15 秒场内的秒级分镜（硬规则）
|
||||
|
||||
主流程一场就是 15 秒。15 秒里**必须再切 3–5 刀**，写进 `visual`，下游故事板和视频只认这段。
|
||||
|
||||
```
|
||||
0-3s：近景，人看镜头或看向场景，先立住环境和情绪
|
||||
3-8s：手部入画，开始使用商品（写清容器、材质、液体/膏体状态）
|
||||
8-12s：商品特写，用法过程可见
|
||||
12-15s：拉回人物，使用后的反应
|
||||
```
|
||||
|
||||
- 第一条从 **0 秒**起，最后一条接到 **15s**。
|
||||
- 每一条写：景别 + 谁 + 手从哪来 + 商品/容器什么样 + 这一拍在变什么。
|
||||
- 15 秒内至少切 **两次景别**（近景 / 特写 / 手持跟拍…）。
|
||||
- 商品用法必须真实：茶要热水和蒸汽，不要把茶包丢进冷白开；手不要悬浮。
|
||||
- 允许另给 `beats:[{start,end,shot_size,action}]`，但 `visual` 仍要能折成上面这种多行文本。
|
||||
|
||||
@@ -48,6 +48,8 @@
|
||||
最弱的一格，因为口播是「对镜说话」，天然打断场景沉浸。
|
||||
- **必须弱化对镜**：人在场景里自然活动，偶尔看向镜头，不要全程盯着
|
||||
- 台词改成自言自语的语气，不要介绍腔
|
||||
- **每镜仍要说满 36–47 字**：写具体时刻、动作、使用感，不要一句商品名就收
|
||||
- 商品名全片点名 1–2 次，其余镜用卖点原词和场景细节
|
||||
- 如果商品本身氛围属性强（香氛、饮品），优先考虑改用 Vlog
|
||||
|
||||
### 短剧 × 痛点解决 ★ 最强组合
|
||||
@@ -100,9 +102,9 @@
|
||||
|
||||
| 结构 | 最短可用 | 原因 |
|
||||
| ---- | ---- | ---- |
|
||||
| 痛点解决 | 15 秒 | 痛点一句、卖点一句、CTA 一句，勉强够 |
|
||||
| 前后对比 | 10 秒 | 前后两镜就能成立，最抗压缩 |
|
||||
| 测评验证 | 20 秒 | 举证需要时间，低于 20 秒证据不足 |
|
||||
| 场景种草 | 20 秒 | 氛围需要铺陈，太短像随机片段 |
|
||||
| 痛点解决 | 15 秒 | 1 镜也能压进钩子+卖点+CTA |
|
||||
| 前后对比 | 15 秒 | 前后对比至少要撑满一镜 |
|
||||
| 测评验证 | 30 秒 | 举证需要时间，1 镜证据不足 |
|
||||
| 场景种草 | 30 秒 | 氛围需要铺陈，太短像随机片段 |
|
||||
|
||||
低于该结构的最短时长时，优先保住「钩子 + 结构核心 + CTA」三件，其余全砍。
|
||||
|
||||
@@ -51,11 +51,11 @@
|
||||
|
||||
## 五、节奏与单镜时长
|
||||
|
||||
- **单镜 6–10 秒**（短剧靠切换制造节奏，镜头长了就拖）
|
||||
- 一镜一个动作或一次对话交换，**别在一镜里塞完整对话**
|
||||
- **单镜固定 15 秒**（短剧用镜内动作和对白交换撑满这 15 秒，不要靠切很多短镜）
|
||||
- 一镜一次冲突或一次对话交换，**别在一镜里塞完整三幕**
|
||||
- 有对白的镜，每句 `line` ≤ 25 字（两个人说话，总量还是受 55 字限制）
|
||||
|
||||
**推荐总时长 45 秒。** 短剧需要装下三幕，低于 30 秒会讲不完整；
|
||||
**推荐总时长 45 秒（3 镜）。** 短剧需要装下三幕，低于 30 秒会讲不完整；
|
||||
超过 60 秒在电商场景下性价比开始下降。
|
||||
|
||||
---
|
||||
|
||||
@@ -49,11 +49,12 @@
|
||||
|
||||
## 四、节奏与单镜时长
|
||||
|
||||
- **单镜 10–15 秒**（口播里说完一个完整意思需要时间，切太碎会让话说不完整）
|
||||
- 一句话讲一件事，**一镜讲一到两句话**
|
||||
- 语速按 **3.5 字 / 秒** 估：12 秒的镜 ≈ 42 字，别超
|
||||
- **单镜固定 15 秒**（口播里说完一个完整意思需要时间；15 秒靠景别切换和手部动作撑满，不要切碎）
|
||||
- 一句话讲一件事，**一镜 2–4 句短句**，禁止一句收工
|
||||
- **visual 按秒拆 3–5 刀**，写成 `0-3s：近景……` / `3-8s：手持……` / `8-12s：特写……` / `12-15s：拉回……`
|
||||
- 语速按 **2.4–3.5 字 / 秒** 估：15 秒镜 **36–52 字**，写作目标 **38–47 字**
|
||||
|
||||
**推荐总时长 30 秒。** 口播超过 45 秒完播率掉得很快，除非是测评验证这种
|
||||
**推荐总时长 30 秒（2 镜）。** 口播超过 45 秒完播率掉得很快，除非是测评验证这种
|
||||
需要举证的结构。
|
||||
|
||||
---
|
||||
|
||||
@@ -55,11 +55,11 @@ Vlog 的镜头必须有**手持感和生活质感**，这是它的可信度来
|
||||
|
||||
## 五、节奏与单镜时长
|
||||
|
||||
- **单镜 5–8 秒**（Vlog 是碎片拼贴，镜头短、数量多）
|
||||
- 允许有 4–5 秒的纯空镜，用来换时间段
|
||||
- **单镜固定 15 秒**（Vlog 的碎片感靠镜内空镜、手部细节、光线变化，不靠把镜子切到 5 秒）
|
||||
- 允许镜内有一段纯画面（旁白留空或只留半句）
|
||||
- 不需要每镜都说话，**画面自己会讲**
|
||||
|
||||
**推荐总时长 30 秒。** Vlog 靠氛围，时长太短装不下时间感（低于 20 秒会像随机片段），
|
||||
**推荐总时长 30 秒（2 镜）。** Vlog 靠氛围，时长太短装不下时间感（低于 30 秒会像随机片段），
|
||||
太长又超出电商场景的耐心。
|
||||
|
||||
---
|
||||
|
||||
@@ -40,12 +40,14 @@
|
||||
|
||||
场景种草的说服力来自细节密度。每条脚本至少铺三样：
|
||||
|
||||
- **光**：晨光 / 夕照 / 台灯 / 烛光 —— 光决定情绪，必须明确写进 `visual`
|
||||
- **光**：晨光 / 夕照 / 台灯 / 烛光 —— 光决定情绪，必须写进每一条秒级分镜
|
||||
- **声音**：水沸 / 键盘 / 雨 / 风 —— 写进 `visual` 提示环境音
|
||||
- **触感**：蒸汽、毛毯的绒、杯壁的温度 —— 让画面有温度
|
||||
- **时间**：具体到时刻（清晨六点 / 下午三点 / 睡前）
|
||||
- **陪衬物**：书、猫、外卖盒、耳机 —— 让场景像真的有人生活
|
||||
|
||||
`visual` 不要写成一段氛围散文，按秒切开：先立场景，再让商品被使用着入场。
|
||||
|
||||
---
|
||||
|
||||
## 四、商品怎么出场
|
||||
@@ -77,10 +79,12 @@
|
||||
|
||||
## 六、台词分寸
|
||||
|
||||
- 语气是**自述或旁白**，轻、慢、留白
|
||||
- 语气是**自述或旁白**，轻、慢；留白靠**句间停顿和画面**，不靠少写字
|
||||
- **不讲参数、不讲价格、不讲功效** —— 这些会立刻把观众从场景里踢出来
|
||||
- 可以整镜无声，让环境音和画面说话
|
||||
- **口播 × 场景种草禁止整镜无声**：每镜仍要说满 36–47 字（2–4 句自述），像在场景里自言自语
|
||||
- **Vlog** 才可以个别镜无声，让环境音和画面说话
|
||||
- CTA 要软：「链接放下面了」而不是「快去抢」
|
||||
- 商品名全片点名 1–2 次，其余镜写时刻、动作、使用感、卖点原词
|
||||
|
||||
---
|
||||
|
||||
|
||||
@@ -374,6 +374,7 @@ export const api = {
|
||||
prompt?: string;
|
||||
model_config_id?: string;
|
||||
selling_point_ids?: string[];
|
||||
persona?: string;
|
||||
base_version_id?: string;
|
||||
aspect_ratio?: string;
|
||||
total_duration?: number;
|
||||
|
||||
@@ -81,16 +81,18 @@
|
||||
--honey-bg: rgba(236, 183, 48, .08);
|
||||
--honey-bd: rgba(236, 183, 48, .20);
|
||||
|
||||
/* ===== Heat · 单 hue + 8 档 alpha · 丁香蓝(对齐 YZ 主 CTA) ===== */
|
||||
--heat: #A5B4FC;
|
||||
--heat-hover: #8ea0f8;
|
||||
--heat-90: rgba(165, 180, 252, .90);
|
||||
--heat-40: rgba(165, 180, 252, .40);
|
||||
--heat-20: rgba(165, 180, 252, .20);
|
||||
--heat-16: rgba(165, 180, 252, .16);
|
||||
--heat-12: rgba(165, 180, 252, .12);
|
||||
--heat-8: rgba(165, 180, 252, .08);
|
||||
--heat-4: rgba(165, 180, 252, .04);
|
||||
/* ===== Heat · 单 hue + 8 档 alpha · 克莱因蓝(对齐 YZ 主 CTA,别用浅紫以免像禁用) ===== */
|
||||
--klein: #002fa7;
|
||||
--klein-hover: #002680;
|
||||
--heat: var(--klein);
|
||||
--heat-hover: var(--klein-hover);
|
||||
--heat-90: rgba(0, 47, 167, .90);
|
||||
--heat-40: rgba(0, 47, 167, .40);
|
||||
--heat-20: rgba(0, 47, 167, .20);
|
||||
--heat-16: rgba(0, 47, 167, .16);
|
||||
--heat-12: rgba(0, 47, 167, .12);
|
||||
--heat-8: rgba(0, 47, 167, .08);
|
||||
--heat-4: rgba(0, 47, 167, .04);
|
||||
|
||||
/* ===== Black-alpha 阶梯 (20 档) ===== */
|
||||
--black-alpha-1: rgba(0, 0, 0, .01);
|
||||
@@ -133,13 +135,13 @@
|
||||
|
||||
/* ===== Shadows ===== */
|
||||
--shadow-cta:
|
||||
0 1px 2px rgba(129, 140, 248, .16),
|
||||
0 6px 16px rgba(165, 180, 252, .28);
|
||||
0 1px 2px rgba(0, 47, 167, .16),
|
||||
0 6px 16px rgba(0, 47, 167, .24);
|
||||
--shadow-cta-hover:
|
||||
0 2px 4px rgba(129, 140, 248, .20),
|
||||
0 10px 22px rgba(165, 180, 252, .34);
|
||||
0 2px 4px rgba(0, 47, 167, .20),
|
||||
0 10px 22px rgba(0, 47, 167, .28);
|
||||
--shadow-cta-active:
|
||||
0 1px 2px rgba(129, 140, 248, .18);
|
||||
0 1px 2px rgba(0, 47, 167, .18);
|
||||
--shadow-floating: 0 8px 28px rgba(17, 19, 24, .06);
|
||||
}
|
||||
|
||||
@@ -220,11 +222,11 @@ img, svg, video { display: block; max-width: 100%; }
|
||||
/* ─── App shell ─── */
|
||||
.app {
|
||||
display: grid;
|
||||
grid-template-columns: var(--sidebar-width) 1fr;
|
||||
grid-template-columns: var(--sidebar-width) minmax(0, 1fr);
|
||||
min-height: 100vh;
|
||||
transition: grid-template-columns var(--t-base);
|
||||
}
|
||||
body.sidebar-collapsed .app { grid-template-columns: 96px 1fr; }
|
||||
body.sidebar-collapsed .app { grid-template-columns: 96px minmax(0, 1fr); }
|
||||
|
||||
/* ─── Sidebar ─── */
|
||||
aside.sidebar {
|
||||
@@ -758,7 +760,17 @@ body.sidebar-collapsed .user .em,
|
||||
body.sidebar-collapsed .user::after { display: none; }
|
||||
|
||||
/* ─── Main + grid background ─── */
|
||||
main { position: relative; background: #fff; min-width: 0; }
|
||||
main { position: relative; background: #fff; min-width: 0; overflow-x: hidden; }
|
||||
.app:has(.pipeline-page) {
|
||||
height: 100vh;
|
||||
overflow: hidden;
|
||||
}
|
||||
main:has(.pipeline-page) {
|
||||
height: 100vh;
|
||||
overflow: hidden;
|
||||
display: flex;
|
||||
flex-direction: column;
|
||||
}
|
||||
.grid-bg {
|
||||
position: absolute;
|
||||
inset: 0;
|
||||
@@ -817,12 +829,13 @@ main { position: relative; background: #fff; min-width: 0; }
|
||||
/* ─── Topbar ─── */
|
||||
.topbar {
|
||||
display: flex; align-items: center; gap: 16px;
|
||||
padding: 0 24px;
|
||||
padding: 0 clamp(36px, 3.2vw, 68px);
|
||||
border-bottom: 1px solid rgba(27, 32, 40, 0.08);
|
||||
background: #fff;
|
||||
position: sticky; top: 0; z-index: 50;
|
||||
height: 116px;
|
||||
min-height: 116px;
|
||||
min-width: 0;
|
||||
box-sizing: border-box;
|
||||
flex-wrap: nowrap;
|
||||
}
|
||||
@@ -926,7 +939,7 @@ main { position: relative; background: #fff; min-width: 0; }
|
||||
color: var(--black-alpha-40);
|
||||
letter-spacing: .02em;
|
||||
}
|
||||
.topbar .right { margin-left: auto; display: flex; align-items: center; gap: 10px; }
|
||||
.topbar .right { margin-left: auto; display: flex; align-items: center; gap: 10px; min-width: 0; flex: 0 1 auto; }
|
||||
.balance-chip {
|
||||
display: inline-flex; align-items: center; gap: 7px;
|
||||
height: 40px;
|
||||
@@ -1007,6 +1020,14 @@ main { position: relative; background: #fff; min-width: 0; }
|
||||
min-height: calc(100vh - 116px);
|
||||
animation: yz-page-enter 280ms cubic-bezier(0.22, 1, 0.36, 1);
|
||||
}
|
||||
.content:has(.pipeline-page) {
|
||||
padding: 0;
|
||||
flex: 1 1 auto;
|
||||
height: auto;
|
||||
min-height: 0;
|
||||
overflow: hidden;
|
||||
animation: none;
|
||||
}
|
||||
@keyframes yz-page-enter {
|
||||
from { opacity: 0; transform: translateY(8px); }
|
||||
to { opacity: 1; transform: translateY(0); }
|
||||
|
||||
@@ -7,17 +7,14 @@
|
||||
width: 100%;
|
||||
max-width: none;
|
||||
box-sizing: border-box;
|
||||
margin: -24px -28px -60px;
|
||||
margin: 0;
|
||||
padding: 28px clamp(40px, 3.2vw, 68px) 32px;
|
||||
height: calc(100vh - 116px);
|
||||
height: 100%;
|
||||
min-height: 0;
|
||||
display: flex;
|
||||
flex-direction: column;
|
||||
overflow: hidden;
|
||||
}
|
||||
@media (max-width: 1100px) {
|
||||
.pipeline-page { margin: -28px -24px -48px; }
|
||||
}
|
||||
|
||||
.pipeline-page .pl-head {
|
||||
min-height: 54px;
|
||||
@@ -1223,9 +1220,10 @@
|
||||
overflow: hidden;
|
||||
display: flex;
|
||||
flex-direction: column;
|
||||
border: 1px solid var(--border-faint);
|
||||
border: 1px solid rgba(0, 47, 167, 0.13);
|
||||
border-radius: 12px;
|
||||
background: var(--surface);
|
||||
box-shadow: 0 12px 28px rgba(20, 27, 38, 0.07);
|
||||
cursor: pointer;
|
||||
}
|
||||
.video-card:hover { border-color: rgba(0, 47, 167, 0.28); }
|
||||
@@ -1235,7 +1233,55 @@
|
||||
aspect-ratio: 9 / 16;
|
||||
max-height: 280px;
|
||||
overflow: hidden;
|
||||
background: #e5e6e9;
|
||||
background: #fff;
|
||||
}
|
||||
.video-thumb.pending::before {
|
||||
content: "";
|
||||
position: absolute;
|
||||
inset: 0;
|
||||
pointer-events: none;
|
||||
background-color: #fff;
|
||||
background-image:
|
||||
radial-gradient(ellipse 4px 1px at center, rgba(0, 47, 167, 0.28) 98%, transparent),
|
||||
radial-gradient(ellipse 1px 4px at center, rgba(0, 47, 167, 0.28) 98%, transparent);
|
||||
background-position: center;
|
||||
background-size: 10px 10px;
|
||||
-webkit-mask-image: radial-gradient(ellipse 42% 28% at center, #000 0 18%, rgba(0, 0, 0, 0.78) 40%, rgba(0, 0, 0, 0.24) 72%, transparent 100%);
|
||||
mask-image: radial-gradient(ellipse 42% 28% at center, #000 0 18%, rgba(0, 0, 0, 0.78) 40%, rgba(0, 0, 0, 0.24) 72%, transparent 100%);
|
||||
}
|
||||
.video-thumb.pending::after,
|
||||
.video-thumb.generating::after {
|
||||
content: "待生成";
|
||||
position: absolute;
|
||||
top: 50%;
|
||||
left: 50%;
|
||||
z-index: 2;
|
||||
pointer-events: none;
|
||||
min-width: 72px;
|
||||
height: 32px;
|
||||
display: grid;
|
||||
place-items: center;
|
||||
padding: 0 14px;
|
||||
border: 1px solid rgba(0, 47, 167, 0.26);
|
||||
border-radius: 8px;
|
||||
color: var(--klein);
|
||||
background: #fff;
|
||||
box-shadow: 0 8px 22px rgba(0, 47, 167, 0.11);
|
||||
font-size: 12px;
|
||||
font-weight: 600;
|
||||
letter-spacing: 0.08em;
|
||||
transform: translate(-50%, -50%);
|
||||
}
|
||||
.video-thumb.generating::after { content: "生成中"; }
|
||||
.video-thumb .video-loader {
|
||||
position: absolute;
|
||||
inset: 0;
|
||||
z-index: 1;
|
||||
overflow: hidden;
|
||||
background:
|
||||
linear-gradient(120deg, transparent 20%, rgba(255, 255, 255, 0.88) 43%, transparent 66%) 0 0 / 220% 100%,
|
||||
linear-gradient(145deg, #eaf1fb, #f8fbff 48%, #e9eef7);
|
||||
animation: videoShimmer 2.2s linear infinite;
|
||||
}
|
||||
.video-thumb video {
|
||||
position: absolute;
|
||||
@@ -1531,8 +1577,265 @@
|
||||
.vd-history-thumb.tri .placeholder { aspect-ratio: 16/9; }
|
||||
}
|
||||
|
||||
/* 弹窗 portal 到 body,选择器必须顶层,否则会被工作区 overflow 裁掉顶栏 */
|
||||
.asset-modal-bg {
|
||||
--klein: #002fa7;
|
||||
position: fixed;
|
||||
inset: 0;
|
||||
z-index: 1800;
|
||||
display: flex;
|
||||
align-items: center;
|
||||
justify-content: center;
|
||||
padding: 32px 24px;
|
||||
background: rgba(21, 20, 15, 0.42);
|
||||
}
|
||||
.asset-modal {
|
||||
background: var(--surface);
|
||||
border: 1px solid var(--border-faint);
|
||||
border-radius: 12px;
|
||||
width: min(960px, 100%);
|
||||
max-height: calc(100vh - 48px);
|
||||
overflow: hidden;
|
||||
display: flex;
|
||||
flex-direction: column;
|
||||
box-shadow: 0 16px 48px rgba(0, 0, 0, 0.18);
|
||||
}
|
||||
.asset-modal-h {
|
||||
display: flex;
|
||||
align-items: center;
|
||||
gap: 10px;
|
||||
min-height: 56px;
|
||||
padding: 12px 20px;
|
||||
border-bottom: 1px solid var(--border-faint);
|
||||
flex-shrink: 0;
|
||||
}
|
||||
.asset-modal-h h2 {
|
||||
margin: 0;
|
||||
font-size: 16px;
|
||||
font-weight: 600;
|
||||
color: var(--accent-black);
|
||||
}
|
||||
.asset-modal-h .ad-tag,
|
||||
.asset-modal-h .mono {
|
||||
font-family: var(--font-mono);
|
||||
font-size: 12px;
|
||||
color: var(--black-alpha-48);
|
||||
letter-spacing: .02em;
|
||||
}
|
||||
.asset-modal-h .x {
|
||||
width: 30px;
|
||||
height: 30px;
|
||||
display: grid;
|
||||
place-items: center;
|
||||
background: transparent;
|
||||
border: 0;
|
||||
cursor: pointer;
|
||||
color: var(--black-alpha-56);
|
||||
border-radius: var(--r-sm);
|
||||
margin-left: auto;
|
||||
}
|
||||
.asset-modal-h .x:hover { background: var(--black-alpha-08); color: var(--accent-black); }
|
||||
.asset-modal-body { padding: 20px 24px 24px; overflow-y: auto; flex: 1; min-height: 0; }
|
||||
.asset-modal-f {
|
||||
padding: 14px 20px;
|
||||
border-top: 1px solid var(--border-faint);
|
||||
display: flex;
|
||||
align-items: center;
|
||||
gap: 8px;
|
||||
flex-shrink: 0;
|
||||
}
|
||||
.asset-detail-grid { display: grid; grid-template-columns: 280px minmax(0, 1fr); gap: 24px; align-items: start; }
|
||||
.asset-detail-lead { display: flex; flex-direction: column; gap: 10px; }
|
||||
.asset-detail-lead .ad-lead-wrap { position: relative; }
|
||||
.asset-detail-lead .placeholder.ad-lead-img { aspect-ratio: 9/16; border-radius: var(--r-md); }
|
||||
.ad-zoom-btn {
|
||||
position: absolute;
|
||||
right: 8px;
|
||||
bottom: 8px;
|
||||
width: 32px;
|
||||
height: 32px;
|
||||
padding: 0;
|
||||
background: rgba(21, 20, 15, .7);
|
||||
color: #fff;
|
||||
border: 0;
|
||||
border-radius: var(--r-sm);
|
||||
display: grid;
|
||||
place-items: center;
|
||||
cursor: pointer;
|
||||
opacity: 0;
|
||||
transition: opacity var(--t-base), background var(--t-base);
|
||||
z-index: 3;
|
||||
}
|
||||
.ad-zoom-btn:hover { background: rgba(21, 20, 15, .92); }
|
||||
.ad-zoom-btn svg { width: 14px; height: 14px; }
|
||||
.asset-detail-lead .ad-lead-wrap:hover .ad-zoom-btn,
|
||||
.asset-detail-tri-row .ad-tri-wrap:hover .ad-zoom-btn { opacity: 1; }
|
||||
.asset-detail-tri-row .ad-tri-wrap { position: relative; }
|
||||
.asset-detail-lead .ad-thumbs { display: flex; gap: 8px; flex-wrap: wrap; }
|
||||
.asset-detail-lead .ad-thumbs .thumb {
|
||||
flex: 0 0 56px;
|
||||
width: 56px;
|
||||
aspect-ratio: 3/4;
|
||||
border-radius: var(--r-sm);
|
||||
border: 1px solid var(--border-faint);
|
||||
cursor: pointer;
|
||||
overflow: hidden;
|
||||
transition: border-color var(--t-base);
|
||||
}
|
||||
.asset-detail-lead .ad-thumbs .thumb:hover { border-color: var(--heat-40); }
|
||||
.asset-detail-lead .ad-thumbs .thumb.active { border-color: var(--heat); border-width: 2px; }
|
||||
.asset-detail-right .ad-section + .ad-section { margin-top: 18px; }
|
||||
.asset-detail-section-h {
|
||||
display: flex;
|
||||
align-items: center;
|
||||
gap: 8px;
|
||||
font-size: 13px;
|
||||
font-weight: 600;
|
||||
color: var(--accent-black);
|
||||
margin-bottom: 10px;
|
||||
}
|
||||
.asset-detail-section-h .ic { width: 14px; height: 14px; color: var(--klein); flex-shrink: 0; display: grid; place-items: center; }
|
||||
.asset-detail-section-h .ic svg { width: 14px; height: 14px; }
|
||||
.asset-detail-section-h .ad-ratio-chip {
|
||||
margin-left: auto;
|
||||
font-family: var(--font-mono);
|
||||
font-size: 12px;
|
||||
padding: 2px 8px;
|
||||
border-radius: var(--r-sm);
|
||||
background: var(--background-lighter);
|
||||
border: 1px solid var(--border-faint);
|
||||
color: var(--black-alpha-56);
|
||||
letter-spacing: .02em;
|
||||
}
|
||||
.asset-detail-section-h .ad-icon-btn {
|
||||
width: 28px;
|
||||
height: 28px;
|
||||
display: grid;
|
||||
place-items: center;
|
||||
background: var(--surface);
|
||||
border: 1px solid var(--border-faint);
|
||||
border-radius: var(--r-sm);
|
||||
color: var(--black-alpha-56);
|
||||
cursor: pointer;
|
||||
transition: background var(--t-base), color var(--t-base), border-color var(--t-base);
|
||||
}
|
||||
.asset-detail-section-h .ad-icon-btn:hover { background: var(--klein); color: #fff; border-color: var(--klein); }
|
||||
.asset-detail-section-h .ad-icon-btn svg { width: 12px; height: 12px; }
|
||||
.asset-detail-section-h .ad-icon-gap { margin-left: 8px; }
|
||||
.asset-detail-tri-row .placeholder { aspect-ratio: 16 / 9; border-radius: var(--r-md); }
|
||||
.md-view-versions { display: flex; gap: 6px; overflow-x: auto; padding: 2px; margin-top: 8px; }
|
||||
.md-view-versions .v-thumb {
|
||||
flex: 0 0 auto;
|
||||
width: 72px;
|
||||
aspect-ratio: 16/9;
|
||||
background: var(--background-lighter);
|
||||
border: 1px solid var(--border-faint);
|
||||
border-radius: var(--r-sm);
|
||||
position: relative;
|
||||
cursor: pointer;
|
||||
display: grid;
|
||||
place-items: center;
|
||||
overflow: hidden;
|
||||
transition: border-color var(--t-base);
|
||||
}
|
||||
.md-view-versions .v-thumb:hover { border-color: var(--heat-40); }
|
||||
.md-view-versions .v-thumb.active { border-color: var(--heat); border-width: 2px; box-shadow: 0 0 0 2px var(--heat-12); }
|
||||
.md-view-versions .v-thumb .v { font-family: var(--font-mono); font-size: 10px; color: var(--black-alpha-56); letter-spacing: .02em; }
|
||||
.md-view-versions .v-thumb.active .v { color: var(--heat); font-weight: 600; }
|
||||
.asset-detail-tip {
|
||||
margin-top: 10px;
|
||||
padding: 10px 12px;
|
||||
background: var(--heat-12);
|
||||
border: 1px solid var(--heat-20);
|
||||
border-radius: var(--r-sm);
|
||||
font-size: 12px;
|
||||
color: var(--accent-black);
|
||||
display: flex;
|
||||
align-items: center;
|
||||
gap: 8px;
|
||||
line-height: 1.5;
|
||||
}
|
||||
.asset-detail-tip svg { width: 14px; height: 14px; color: var(--heat); flex-shrink: 0; }
|
||||
.asset-detail-tip .ai-gen-btn {
|
||||
margin-left: auto;
|
||||
height: 26px;
|
||||
padding: 0 10px;
|
||||
background: var(--heat);
|
||||
color: var(--accent-white);
|
||||
border: 1px solid var(--heat);
|
||||
border-radius: var(--r-sm);
|
||||
font-size: 12px;
|
||||
cursor: pointer;
|
||||
font-family: inherit;
|
||||
flex-shrink: 0;
|
||||
}
|
||||
.asset-modal-f .ad-foot-stats { display: flex; gap: 6px; margin-right: auto; }
|
||||
.asset-modal-f .ad-stat-btn {
|
||||
height: 32px;
|
||||
padding: 0 12px;
|
||||
display: inline-flex;
|
||||
align-items: center;
|
||||
gap: 6px;
|
||||
background: transparent;
|
||||
border: 1px solid var(--border-faint);
|
||||
border-radius: var(--r-sm);
|
||||
color: var(--black-alpha-72);
|
||||
font-size: 12px;
|
||||
font-family: inherit;
|
||||
cursor: pointer;
|
||||
transition: background var(--t-base), color var(--t-base), border-color var(--t-base);
|
||||
}
|
||||
.asset-modal-f .ad-stat-btn:hover { background: var(--background-lighter); color: var(--accent-black); }
|
||||
.asset-modal-f .ad-stat-btn svg { width: 13px; height: 13px; }
|
||||
.ad-detail-prompt {
|
||||
width: 100%;
|
||||
font-family: var(--font-mono);
|
||||
font-size: 12px;
|
||||
line-height: 1.6;
|
||||
color: var(--black-alpha-72);
|
||||
background: var(--background-base);
|
||||
border: 1px solid var(--border-faint);
|
||||
border-radius: var(--r-sm);
|
||||
padding: 10px 12px;
|
||||
outline: none;
|
||||
resize: none;
|
||||
height: 88px;
|
||||
transition: border-color var(--t-base), background var(--t-base), box-shadow var(--t-base);
|
||||
}
|
||||
.ad-thumbs .thumb.has-mock-media,
|
||||
.md-view-versions .v-thumb.has-mock-media,
|
||||
.prod-preview-history .h-thumb.has-mock-media {
|
||||
background-image: var(--mock-media-url);
|
||||
background-size: cover;
|
||||
background-position: center;
|
||||
background-repeat: no-repeat;
|
||||
}
|
||||
.ad-detail-prompt:hover { border-color: var(--heat-20); }
|
||||
.ad-detail-prompt:focus { border-color: var(--heat); background: var(--surface); color: var(--accent-black); box-shadow: 0 0 0 3px var(--heat-12); }
|
||||
.vd-main-wrap { display: flex; gap: 18px; align-items: flex-start; }
|
||||
.vd-main { flex: 0 0 280px; aspect-ratio: 9/16; max-height: 460px; background: #000; border-radius: var(--r-md); overflow: hidden; position: relative; }
|
||||
.vd-main video { position: absolute; inset: 0; width: 100%; height: 100%; object-fit: contain; }
|
||||
.vd-info { flex: 1; min-width: 0; }
|
||||
.vd-section-h { display: flex; align-items: center; gap: 8px; font-size: 13px; font-weight: 600; color: var(--accent-black); margin-bottom: 10px; }
|
||||
.vd-kv { display: grid; grid-template-columns: 72px 1fr; gap: 6px 10px; font-size: 13px; }
|
||||
.vd-kv .k { font-family: var(--font-mono); font-size: 12px; color: var(--black-alpha-48); letter-spacing: .03em; padding-top: 1px; }
|
||||
.vd-kv .v { color: var(--accent-black); min-width: 0; overflow-wrap: anywhere; }
|
||||
.vd-history-h { font-family: var(--font-mono); font-size: 12px; color: var(--black-alpha-48); letter-spacing: .06em; text-transform: uppercase; margin-bottom: 10px; }
|
||||
.vd-history-row { display: flex; gap: 8px; overflow-x: auto; padding-bottom: 6px; scrollbar-width: thin; }
|
||||
.vd-history-thumb { flex: 0 0 64px; min-width: 64px; display: flex; flex-direction: column; gap: 4px; padding: 4px; border: 1px solid var(--border-faint); border-radius: var(--r-sm); background: var(--surface); cursor: pointer; transition: border-color var(--t-base); position: relative; }
|
||||
.vd-history-thumb:hover { border-color: var(--heat); }
|
||||
.vd-history-thumb.current { border-color: var(--heat); background: var(--heat-12); }
|
||||
.vd-history-thumb .placeholder { aspect-ratio: 9/16; }
|
||||
.vd-history-thumb .ts { font-family: var(--font-mono); font-size: 12px; color: var(--black-alpha-48); text-align: center; }
|
||||
.vd-prompt-field { margin-top: 16px; }
|
||||
.vd-prompt-head { display: flex; align-items: baseline; justify-content: space-between; gap: 12px; margin-bottom: 6px; }
|
||||
.vd-prompt-head .label { font-family: var(--font-mono); font-size: 12px; font-weight: 500; color: var(--black-alpha-56); letter-spacing: .04em; }
|
||||
.vd-prompt-edit { background: var(--background-base); border: 1px solid var(--border-faint); border-radius: var(--r-md); padding: 12px 14px; font-family: var(--font-mono); font-size: 12px; line-height: 1.7; color: var(--accent-black); white-space: pre-wrap; min-height: 96px; outline: none; letter-spacing: .01em; cursor: text; }
|
||||
.vd-modal-actions { margin-left: auto; display: flex; align-items: center; gap: 8px; }
|
||||
|
||||
/* 流程步骤4 · 基础资产 loading 转圈(顶层 @keyframes 才注册) */
|
||||
@keyframes assetSpin { to { transform: rotate(360deg); } }
|
||||
@keyframes videoShimmer { to { background-position: -220% 0, 0 0; } }
|
||||
|
||||
/* 进度流逐条滚入(顶层 @keyframes 才注册) */
|
||||
@keyframes psRowIn { from { opacity: 0; transform: translateY(4px); } to { opacity: 1; transform: translateY(0); } }
|
||||
@@ -1646,7 +1949,7 @@
|
||||
.ref-gate-mask {
|
||||
position: fixed;
|
||||
inset: 0;
|
||||
z-index: 60;
|
||||
z-index: 1800;
|
||||
display: flex;
|
||||
align-items: center;
|
||||
justify-content: center;
|
||||
|
||||
@@ -1,4 +1,5 @@
|
||||
import { Fragment, memo, useCallback, useDeferredValue, useEffect, useMemo, useRef, useState } from "react";
|
||||
import { createPortal } from "react-dom";
|
||||
import type { ChangeEvent, CSSProperties, PointerEvent as ReactPointerEvent } from "react";
|
||||
import { ArrowLeft, ArrowRight, Check, ChevronRight, Image, Info, LayoutList, Play, RefreshCw, Route, Sparkles, Upload, UsersRound, X } from "lucide-react";
|
||||
import { api, ApiError } from "../api";
|
||||
@@ -352,7 +353,9 @@ function PromptBox({ value, onChange, className = "prompt-box", id, ariaLabel, s
|
||||
|
||||
// 分镜「旁白 / 画面」可编辑框:非受控 textarea,落库走 onBlur。
|
||||
// 聚焦中绝不覆盖,避免提交后父组件用旧值把正在输入的字冲掉。
|
||||
function ShotArea({ value, placeholder, ariaLabel, onCommit }: { value: string; placeholder: string; ariaLabel: string; onCommit: (text: string) => void }) {
|
||||
function ShotArea({ value, placeholder, ariaLabel, onCommit, rows = 2 }: {
|
||||
value: string; placeholder: string; ariaLabel: string; onCommit: (text: string) => void; rows?: number;
|
||||
}) {
|
||||
const ref = useRef<HTMLTextAreaElement | null>(null);
|
||||
useEffect(() => {
|
||||
const el = ref.current;
|
||||
@@ -366,7 +369,7 @@ function ShotArea({ value, placeholder, ariaLabel, onCommit }: { value: string;
|
||||
defaultValue={value}
|
||||
placeholder={placeholder}
|
||||
aria-label={ariaLabel}
|
||||
rows={2}
|
||||
rows={rows}
|
||||
onBlur={(e) => {
|
||||
const text = e.currentTarget.value.trim();
|
||||
if (text !== value) onCommit(text);
|
||||
@@ -404,7 +407,7 @@ function DraftShotCard({ draft, onCommit, onCancel }: {
|
||||
</label>
|
||||
<label className="shot-edit">
|
||||
<span>画面内容</span>
|
||||
<textarea className="shot-area" rows={2} placeholder="输入这一镜的画面描述…" value={visual}
|
||||
<textarea className="shot-area" rows={6} placeholder="0-3s：近景……（按秒写出每一镜）" value={visual}
|
||||
onChange={(e) => setVisual(e.target.value)}
|
||||
onKeyDown={(e) => { if (e.key === "Enter" && !e.shiftKey && !e.nativeEvent.isComposing) { e.preventDefault(); commit(); } else if (e.key === "Escape") { onCancel?.(draft.id); } }} />
|
||||
</label>
|
||||
@@ -599,7 +602,7 @@ export function PipelinePage(props: {
|
||||
try { return new Set<string>(JSON.parse(localStorage.getItem(`airshelf:pending-del:${project.id}`) || "[]")); } catch { return new Set(); }
|
||||
});
|
||||
const shots = [...(currentScript?.segments ?? [])].sort((a, b) => a.sort_order - b.sort_order).filter((s) => !pendingDeletes.has(s.id));
|
||||
// 单镜时长自二期起是变长的(4–15s),缺值时按上限兜底,跟后端 script_agent 一致
|
||||
// 单镜时长主流程固定 15s;缺值时按上限兜底,跟后端 script_agent 一致
|
||||
const shotSeconds = (s: { duration_seconds?: number | null }) => s.duration_seconds || SEGMENT_DURATION_MAX;
|
||||
const scriptTotalSeconds = shots.reduce((sum, s) => sum + shotSeconds(s), 0);
|
||||
// 对白 speaker(entity id)→ 角色名;null = 旁白。entities 在 ScriptVersion.metadata
|
||||
@@ -1534,9 +1537,11 @@ export function PipelinePage(props: {
|
||||
mode: agentMode,
|
||||
prompt,
|
||||
model_config_id: activeScriptModelId || undefined,
|
||||
selling_point_ids: Array.isArray(wizard?.selling_point_ids) ? wizard.selling_point_ids : undefined,
|
||||
persona: setupPersona,
|
||||
base_version_id: baseVersionId,
|
||||
aspect_ratio: "9:16",
|
||||
// 设定卡的三个参数一路传到后端:后端据此挑套路提示词 + 按节奏切镜(不再每镜写死 15 秒)
|
||||
// 设定卡参数一路传到后端:每镜固定 15 秒,总时长 15/30/45/60
|
||||
total_duration: setupDuration,
|
||||
presentation_format: setupFormat,
|
||||
video_structure: setupStructure,
|
||||
@@ -1625,8 +1630,7 @@ export function PipelinePage(props: {
|
||||
pushMsg("ai", summaryText || "镜头脚本已生成,左侧已刷新。可继续输入修改意见,或点「确认脚本」进入下一步。");
|
||||
}
|
||||
}
|
||||
// 1.5 · 确认设定后真正发起生成。表现形式/结构/时长走结构化参数(后端挑套路),
|
||||
// 人物没有对应的后端字段,仍并进 prompt 让模型读。
|
||||
// 1.5 · 确认设定后真正发起生成。形式/结构/时长/人物/勾选卖点都走结构化参数,不再塞一句空主题。
|
||||
async function runScriptWithSetup() {
|
||||
const format = coercePresentationFormat(setupFormat);
|
||||
const structure = coerceVideoStructure(setupStructure);
|
||||
@@ -1684,12 +1688,12 @@ export function PipelinePage(props: {
|
||||
// 5.2 换商品重跑:新建项目时选了套路模板 → 把模板的镜序/节奏/转化写法并进 prompt,
|
||||
// 但明确要求文案按当前商品重写(模板里不含旧文案,转化写法可能带旧品牌词,这里再挡一道)。
|
||||
const tplLine = templateOutline
|
||||
? `\n请照这个套路重出一版:\n${templateOutline}\n`
|
||||
+ `镜数、每镜作用与时长按上面来,文案全部按我当前的商品重新写,`
|
||||
? `请照这个套路重出一版:\n${templateOutline}\n`
|
||||
+ `镜数、每镜作用按上面来(每镜 15 秒),文案全部按我当前的商品重新写,`
|
||||
+ `参考写法里出现的旧商品、品牌、价格一律不要保留。`
|
||||
: "";
|
||||
await runScriptGeneration(
|
||||
`目标人群:${personaLabel}。突出商品卖点,节奏紧凑,适合短视频投放${tplLine}`,
|
||||
tplLine,
|
||||
templateName ? `套用模板「${templateName}」:${combo}` : `${sourceLabel}:${combo}`,
|
||||
"ai",
|
||||
);
|
||||
@@ -3004,8 +3008,9 @@ export function PipelinePage(props: {
|
||||
<span>画面内容</span>
|
||||
<ShotArea
|
||||
value={visualShown}
|
||||
placeholder="点击编辑画面内容"
|
||||
placeholder="0-3s：近景……（按秒写出每一镜）"
|
||||
ariaLabel={`场景 ${index + 1} 画面内容`}
|
||||
rows={6}
|
||||
onCommit={(text) => { void onUpdateShot({ segment_id: shot.id, visual_prompt: text }); }}
|
||||
/>
|
||||
</label>
|
||||
@@ -3102,7 +3107,7 @@ export function PipelinePage(props: {
|
||||
</select>
|
||||
</label>
|
||||
<div className="setup-rec">根据商品目标人群推荐</div>
|
||||
{/* 1.9 时长自由化:5–60 秒、5 秒一档;每镜切多长由后端按表现形式的节奏决定 */}
|
||||
{/* 时长:15/30/45/60;每镜固定 15 秒 */}
|
||||
<label className="setup-field">
|
||||
<span className="sf-k">时长</span>
|
||||
<select className="setup-select" value={setupDuration} onChange={(e) => { setupTouched.current = true; setSetupDuration(clampDuration(Number(e.target.value))); }}>
|
||||
@@ -3453,7 +3458,7 @@ export function PipelinePage(props: {
|
||||
return (
|
||||
<div className="as-gen" data-asset-kind="scene" key={dk}>
|
||||
<div className={`as-gen-preview${busy ? " generating" : " pending"}`}>
|
||||
{busy ? <span className="asset-card-loading"><span className="asset-spinner" aria-hidden="true"></span><span>生成中 · 约 8s</span></span> : null}
|
||||
{busy ? <span className="asset-card-loading"><span className="asset-spinner" aria-hidden="true"></span><span>生成中</span></span> : null}
|
||||
</div>
|
||||
<div className="as-gen-meta">
|
||||
<input type="text" value={draft.title} placeholder="场景图待生成" aria-label="场景名" disabled={busy} data-stop
|
||||
@@ -3498,7 +3503,7 @@ export function PipelinePage(props: {
|
||||
onClick={() => openSeedDetail(kind, tag, promptValue)}
|
||||
onKeyDown={(e) => { if (e.key === "Enter" || e.key === " ") { e.preventDefault(); openSeedDetail(kind, tag, promptValue); } }}
|
||||
>
|
||||
{busy ? <span className="asset-card-loading"><span className="asset-spinner" aria-hidden="true"></span><span>生成中 · 约 8s</span></span> : null}
|
||||
{busy ? <span className="asset-card-loading"><span className="asset-spinner" aria-hidden="true"></span><span>生成中</span></span> : null}
|
||||
</div>
|
||||
<div className="as-gen-meta">
|
||||
<h4 onClick={() => openSeedDetail(kind, tag, promptValue)}>{tag || (kind === "person" ? "角色图待生成" : "场景图待生成")}</h4>
|
||||
@@ -3823,11 +3828,17 @@ export function PipelinePage(props: {
|
||||
const showBusy = busy || pending;
|
||||
const verCount = seg.versions?.length ?? 0;
|
||||
return (
|
||||
<div className="video-card" key={seg.id} data-video-id={seg.id}>
|
||||
<div className="video-thumb" role="button" tabIndex={0} title="点击查看详情" onClick={() => openVideoDetail(seg.id)} onKeyDown={(event) => { if (event.key === "Enter" || event.key === " ") { event.preventDefault(); openVideoDetail(seg.id); } }}>
|
||||
{url
|
||||
? <video src={`${url}#t=0.1`} muted playsInline preload="metadata" />
|
||||
: <span className="ph-frame">场 {seg.sort_order + 1}</span>}
|
||||
<div className={`video-card${url ? " is-ready" : showBusy ? " is-generating" : " is-pending"}`} key={seg.id} data-video-id={seg.id}>
|
||||
<div
|
||||
className={`video-thumb${url ? " ready" : ""}${showBusy ? " generating" : ""}${!url && !showBusy ? " pending" : ""}`}
|
||||
role="button"
|
||||
tabIndex={0}
|
||||
title="点击查看详情"
|
||||
aria-label={showBusy ? `场 ${seg.sort_order + 1} 生成中` : url ? `场 ${seg.sort_order + 1}` : `场 ${seg.sort_order + 1} 待生成`}
|
||||
onClick={() => openVideoDetail(seg.id)}
|
||||
onKeyDown={(event) => { if (event.key === "Enter" || event.key === " ") { event.preventDefault(); openVideoDetail(seg.id); } }}
|
||||
>
|
||||
{url ? <video src={`${url}#t=0.1`} muted playsInline preload="metadata" /> : null}
|
||||
{url && !showBusy && (
|
||||
<button
|
||||
className="btn-play"
|
||||
@@ -3838,7 +3849,7 @@ export function PipelinePage(props: {
|
||||
<Play size={16} fill="currentColor" />
|
||||
</button>
|
||||
)}
|
||||
{showBusy && <span className="sb-gen-veil" aria-hidden="true"><span className="spinner" /></span>}
|
||||
{showBusy ? <span className="video-loader" aria-hidden="true" /> : null}
|
||||
</div>
|
||||
<div className="body">
|
||||
<div className="video-card-head">
|
||||
@@ -4282,7 +4293,7 @@ export function PipelinePage(props: {
|
||||
const showBusy = busy || rerunPending.has(vdSeg.id);
|
||||
const hint = (shots[vdSeg.sort_order]?.visual_prompt || shots[vdSeg.sort_order]?.narration || "").trim();
|
||||
const rerunPrompt = vdPrompt.trim() || `${videoPrompt} 第 ${vdSeg.sort_order + 1} 段，时长 ${vdSeg.target_duration_seconds} 秒`;
|
||||
return (
|
||||
return createPortal(
|
||||
<div className="asset-modal-bg" role="dialog" aria-modal="true" aria-label="视频详情" onClick={(event) => { if (event.target === event.currentTarget) setVdSegId(null); }}>
|
||||
<div className="asset-modal">
|
||||
<div className="asset-modal-h">
|
||||
@@ -4351,7 +4362,8 @@ export function PipelinePage(props: {
|
||||
</div>
|
||||
</div>
|
||||
</div>
|
||||
</div>
|
||||
</div>,
|
||||
document.body
|
||||
);
|
||||
})()}
|
||||
{/* ── 流程步骤4/5 · 兜底弹窗拦截:参考图不齐时拦住生成,让用户去补 / 仍要生成(随他便) ── */}
|
||||
@@ -4359,7 +4371,7 @@ export function PipelinePage(props: {
|
||||
// 拆两类:noref=连参考图都没有;notri=有立绘但缺三视图。文案分别告诉用户「去哪里点什么」。
|
||||
const noRef = refGate.missing.filter((m) => m.reason === "noref");
|
||||
const noTri = refGate.missing.filter((m) => m.reason === "notri");
|
||||
return (
|
||||
return createPortal(
|
||||
<div className="ref-gate-mask" onClick={() => setRefGate(null)}>
|
||||
<div className="ref-gate-modal" role="dialog" aria-modal="true" onClick={(e) => e.stopPropagation()}>
|
||||
<div className="rg-title">资产还没齐,先补一下?</div>
|
||||
@@ -4398,11 +4410,12 @@ export function PipelinePage(props: {
|
||||
<button type="button" className="btn btn-ghost rg-force" onClick={() => { const go = refGate.proceed; setRefGate(null); go(); }}>仍要继续 <span className="rg-warn">可能跟角色对不上</span></button>
|
||||
</div>
|
||||
</div>
|
||||
</div>
|
||||
</div>,
|
||||
document.body
|
||||
);
|
||||
})()}
|
||||
{/* ── 流程步骤5 · 过审闸弹窗:含真人脸的人物/分镜未过审 → 列出哪一镜/哪个,先过审再生成(火山合规要求) ── */}
|
||||
{reviewGate && (
|
||||
{reviewGate && createPortal(
|
||||
<div className="ref-gate-mask" onClick={() => setReviewGate(null)}>
|
||||
<div className="ref-gate-modal" role="dialog" aria-modal="true" onClick={(e) => e.stopPropagation()}>
|
||||
<div className="rg-title">这些素材还没过审,先过审再生成视频</div>
|
||||
@@ -4440,7 +4453,8 @@ export function PipelinePage(props: {
|
||||
<button type="button" className="btn btn-ghost" onClick={() => setReviewGate(null)}>知道了</button>
|
||||
</div>
|
||||
</div>
|
||||
</div>
|
||||
</div>,
|
||||
document.body
|
||||
)}
|
||||
{/* ── 流程步骤4 · 人物/场景详情弹窗:立绘 + 三视图(人物)+ 提示词重获 + 版本历史 + 应用到当前项目 ── */}
|
||||
{adDetail && (() => {
|
||||
@@ -4494,12 +4508,12 @@ export function PipelinePage(props: {
|
||||
setAdTriId(null);
|
||||
}
|
||||
const zoomSvg = <svg viewBox="0 0 24 24" fill="none" stroke="currentColor" strokeWidth="1.5" strokeLinecap="round" strokeLinejoin="round"><path d="M3 8V3h5M16 3h5v5M21 16v5h-5M8 21H3v-5" /></svg>;
|
||||
return (
|
||||
return createPortal(
|
||||
<div className="asset-modal-bg" role="dialog" aria-modal="true" aria-label={`${KIND_LABEL[adDetail.kind]}详情`} onClick={(event) => { if (event.target === event.currentTarget) setAdDetail(null); }}>
|
||||
<div className="asset-modal">
|
||||
<div className="asset-modal-h">
|
||||
<h2>资产详情</h2>
|
||||
<span className="ad-tag">/ {KIND_LABEL[adDetail.kind]} · {entity.name}</span>
|
||||
<h2>{entity.name}</h2>
|
||||
<span className="ad-tag">/ {KIND_LABEL[adDetail.kind]}</span>
|
||||
<button className="x" type="button" aria-label="关闭" onClick={() => setAdDetail(null)}>
|
||||
<svg width="16" height="16" viewBox="0 0 24 24" fill="none" stroke="currentColor" strokeWidth="1.8" strokeLinecap="round" strokeLinejoin="round"><path d="M6 6l12 12M6 18L18 6" /></svg>
|
||||
</button>
|
||||
@@ -4629,11 +4643,12 @@ export function PipelinePage(props: {
|
||||
}}>应用当前角色</button>
|
||||
</div>
|
||||
</div>
|
||||
</div>
|
||||
</div>,
|
||||
document.body
|
||||
);
|
||||
})()}
|
||||
{/* 问题5(a):用已有素材作三视图——选商品已有图 / 上传本地图,attach 成商品采用版(不走平台 AI 生成、不计费) */}
|
||||
{assetPick && (
|
||||
{assetPick && createPortal(
|
||||
<div className="asset-modal-bg" role="dialog" aria-modal="true" aria-label={assetPick.title} onClick={(e) => { if (e.target === e.currentTarget && !assetPickBusy) setAssetPick(null); }}>
|
||||
<div className="asset-modal">
|
||||
<div className="asset-modal-h">
|
||||
@@ -4675,7 +4690,8 @@ export function PipelinePage(props: {
|
||||
)}
|
||||
</div>
|
||||
</div>
|
||||
</div>
|
||||
</div>,
|
||||
document.body
|
||||
)}
|
||||
{/* 流程步骤4 · 模特库：浏览 / 添加模特（AI 生成·本地上传）/ 替换回填 */}
|
||||
<ModelLibrary
|
||||
|
||||
@@ -79,7 +79,7 @@ export function ProjectWizardPage({ products, projects = [], preselectProductId,
|
||||
useEffect(() => {
|
||||
const p = products.find((item) => item.id === productId);
|
||||
const seed: Record<string, boolean> = {};
|
||||
(p?.selling_points || []).forEach((sp) => { seed[sp.title] = false; });
|
||||
(p?.selling_points || []).forEach((sp) => { seed[sp.id || sp.title] = false; });
|
||||
setPoints(seed);
|
||||
}, [productId, products]);
|
||||
|
||||
@@ -344,12 +344,15 @@ export function ProjectWizardPage({ products, projects = [], preselectProductId,
|
||||
<div className="nw-field">
|
||||
<span>关键卖点<small>可勾选要重点突出的</small></span>
|
||||
<div className="nw-pills">
|
||||
{Object.entries(points).map(([key, on]) => (
|
||||
{Object.entries(points).map(([key, on]) => {
|
||||
const label = product?.selling_points?.find((sp) => (sp.id || sp.title) === key)?.title || key;
|
||||
return (
|
||||
<button className={`nw-pill${on ? " on" : ""}`} type="button" key={key} aria-pressed={on} onClick={() => setPoints((prev) => ({ ...prev, [key]: !prev[key] }))}>
|
||||
{on && <Check size={13} />}
|
||||
<span>{key}</span>
|
||||
<span>{label}</span>
|
||||
</button>
|
||||
))}
|
||||
);
|
||||
})}
|
||||
</div>
|
||||
</div>
|
||||
)}
|
||||
|
||||
@@ -69,12 +69,12 @@ export function allowedStructures(format: PresentationFormat): VideoStructure[]
|
||||
return STRUCTURE_KEYS.filter((key) => !isForbidden(format, key));
|
||||
}
|
||||
|
||||
// ── 时长:总时长 5–60 秒、5 秒步进;单镜 4–15 秒由后端切,前端只管总时长 ──
|
||||
// ── 时长:总时长 15/30/45/60;主流程每镜固定 15 秒 ──
|
||||
export const SEGMENT_DURATION_MIN = 4;
|
||||
export const SEGMENT_DURATION_MAX = 15;
|
||||
export const TOTAL_DURATION_MIN = 5;
|
||||
export const TOTAL_DURATION_MIN = 15;
|
||||
export const TOTAL_DURATION_MAX = 60;
|
||||
export const TOTAL_DURATION_STEP = 5;
|
||||
export const TOTAL_DURATION_STEP = 15;
|
||||
export const DURATION_OPTIONS: number[] = Array.from(
|
||||
{ length: (TOTAL_DURATION_MAX - TOTAL_DURATION_MIN) / TOTAL_DURATION_STEP + 1 },
|
||||
(_, i) => TOTAL_DURATION_MIN + i * TOTAL_DURATION_STEP
|
||||
@@ -82,8 +82,8 @@ export const DURATION_OPTIONS: number[] = Array.from(
|
||||
|
||||
/** 表现形式推荐的默认总时长:口播短平快 / 短剧要装下三幕 / Vlog 要铺氛围。 */
|
||||
const FORMAT_DEFAULT_DURATION: Record<PresentationFormat, number> = { oral: 30, drama: 45, vlog: 30 };
|
||||
/** 各结构能压到的最短总时长 —— 低于它证据或氛围就不成立了。 */
|
||||
const STRUCTURE_MIN_DURATION: Record<VideoStructure, number> = { pain: 15, contrast: 10, review: 20, scene: 20 };
|
||||
/** 各结构能压到的最短总时长 —— 低于它证据或氛围就不成立了。须落在 15 秒步进上。 */
|
||||
const STRUCTURE_MIN_DURATION: Record<VideoStructure, number> = { pain: 15, contrast: 15, review: 30, scene: 30 };
|
||||
|
||||
export function recommendDuration(format: PresentationFormat, structure: VideoStructure): number {
|
||||
return Math.max(FORMAT_DEFAULT_DURATION[format], STRUCTURE_MIN_DURATION[structure]);
|
||||
|
||||
@@ -56,10 +56,10 @@
|
||||
--ink-2: #5A5F6A;
|
||||
--ink-3: #8B909A;
|
||||
--ink-4: #C5C9D2;
|
||||
--orange: #A5B4FC;
|
||||
--orange-soft: #E4E9FE;
|
||||
--orange-tint: #F0F3FF;
|
||||
--orange-hover: #8B9CF7;
|
||||
--orange: #002fa7;
|
||||
--orange-soft: rgba(0, 47, 167, .18);
|
||||
--orange-tint: rgba(0, 47, 167, .06);
|
||||
--orange-hover: #002680;
|
||||
--green: #3F6B3F;
|
||||
--green-bg: #EAF2EA;
|
||||
--green-bd: #D5E5D5;
|
||||
@@ -393,7 +393,7 @@ main { position: relative; overflow: hidden; background: #fff; }
|
||||
font-family: inherit;
|
||||
color: var(--ink);
|
||||
}
|
||||
.input:focus, .textarea:focus, .select:focus { border-color: var(--orange); box-shadow: 0 0 0 3px rgba(165, 180, 252, 0.18); }
|
||||
.input:focus, .textarea:focus, .select:focus { border-color: var(--orange); box-shadow: 0 0 0 3px rgba(0, 47, 167, 0.18); }
|
||||
.input::placeholder, .textarea::placeholder { color: var(--ink-3); }
|
||||
.textarea { height: auto; min-height: 80px; padding: 10px 12px; line-height: 1.55; resize: vertical; }
|
||||
.select { appearance: none; background-image: url("data:image/svg+xml;utf8,<svg xmlns='http://www.w3.org/2000/svg' width='12' height='12' viewBox='0 0 16 16' fill='none'><path d='M4 6l4 4 4-4' stroke='%239C988C' stroke-width='1.4'/></svg>"); background-repeat: no-repeat; background-position: right 10px center; padding-right: 30px; }
|
||||
@@ -714,8 +714,8 @@ table.t tbody tr:hover { background: var(--bg-soft); }
|
||||
--accent-crimson: #eb3424;
|
||||
--accent-forest: #42c366;
|
||||
--accent-honey: #ecb730;
|
||||
--heat: #A5B4FC;
|
||||
--heat-12: rgba(165, 180, 252, .12);
|
||||
--heat: #002fa7;
|
||||
--heat-12: rgba(0, 47, 167, .12);
|
||||
--black-alpha-32: rgba(38, 38, 38, .32);
|
||||
--black-alpha-48: rgba(38, 38, 38, .48);
|
||||
--black-alpha-56: rgba(38, 38, 38, .56);
|
||||
@@ -1020,7 +1020,7 @@ table.t tbody tr:hover { background: var(--bg-soft); }
|
||||
justify-content: center;
|
||||
gap: 8px;
|
||||
}
|
||||
.auth-exact-page .btn-cta:hover { box-shadow: 0 4px 14px rgba(165, 180, 252, .28); }
|
||||
.auth-exact-page .btn-cta:hover { box-shadow: 0 4px 14px rgba(0, 47, 167, .28); }
|
||||
.auth-exact-page .btn-cta:active { transform: translateY(1px); }
|
||||
.auth-exact-page .btn-cta:disabled { opacity: .82; cursor: wait; }
|
||||
.auth-exact-page .btn-cta .busy-copy {
|
||||
|
||||
Reference in New Issue
Block a user