nex_animation/backend/app/constants.py

330 lines
15 KiB
Python
Raw Blame History

This file contains ambiguous Unicode characters!

This file contains ambiguous Unicode characters that may be confused with others in your current locale. If your use case is intentional and legitimate, you can safely ignore this warning. Use the Escape button to highlight these characters.

"""平台领域常量:与业务流程图(图2)和模块规划图(图1)一一对应。"""
from __future__ import annotations
import re
class ProjectStatus:
DRAFT = "draft" # 已创建
STORYBOARD = "storyboard" # 分镜/脚本编辑中
BINDING = "binding" # 资产绑定中
GENERATING = "generating" # 生成中
REVIEWING = "reviewing" # 质检与人工审核
COMPOSING = "composing" # 时间线合成 / 混音
DELIVERED = "delivered" # 已导出成片
ARCHIVED = "archived"
ALL = [DRAFT, STORYBOARD, BINDING, GENERATING, REVIEWING, COMPOSING, DELIVERED, ARCHIVED]
class TaskKind:
GENERATE_SHOT = "generate_shot" # 调用视频模型
VOICEOVER = "voiceover" # 台词配音(TTS)
QC_SHOT = "qc_shot" # 自动质量检查
COMPOSE = "compose" # 按时间线合成 + 混音 + 导出
IMPORT = "import" # 分镜/脚本导入
ASSET = "asset" # 资产衍生(占位,供扩展)
# 平台资产库的 AI 生成:出定妆图 / 场景图 / 试听样本。
# 它与 ASSET 的区别只在「不属于任何项目」——正因如此才必须单独一种 kind,
# 否则按 kind 过滤的队列视图会把平台侧的生成混进某个项目的资产任务里。
PLATFORM_ASSET = "platform_asset"
ALL = [GENERATE_SHOT, VOICEOVER, QC_SHOT, COMPOSE, IMPORT, ASSET, PLATFORM_ASSET]
class TaskStatus:
PLANNED = "planned" # 已生成任务清单,尚未进入队列
QUEUED = "queued" # 队列中,等待限流放行
RUNNING = "running" # 执行中
RETRYING = "retrying" # 失败已记录,等待重试
WAITING_REVIEW = "waiting_review"
SUCCEEDED = "succeeded"
FAILED = "failed"
CANCELLED = "cancelled"
SKIPPED = "skipped" # 断点续跑时命中已有产物
ACTIVE = [QUEUED, RUNNING, RETRYING]
OPEN = [PLANNED, QUEUED, RUNNING, RETRYING, WAITING_REVIEW]
TERMINAL = [SUCCEEDED, FAILED, CANCELLED, SKIPPED]
ALL = [PLANNED, QUEUED, RUNNING, RETRYING, WAITING_REVIEW, SUCCEEDED, FAILED, CANCELLED, SKIPPED]
class ShotStatus:
DRAFT = "draft" # 已导入/新建,未就绪
READY = "ready" # 通过素材与参数检查
QUEUED = "queued"
RUNNING = "running"
RETRYING = "retrying"
GENERATED = "generated" # 生成成功、已下载保存,等待质检
QC_PASSED = "qc_passed" # 自动质检通过,待人工审核
APPROVED = "approved" # 人工审核合格
REJECTED = "rejected" # 人工审核不合格,需调整提示词或重新生成
FAILED = "failed"
ACTIVE = [QUEUED, RUNNING, RETRYING]
USABLE = [APPROVED]
ALL = [DRAFT, READY, QUEUED, RUNNING, RETRYING, GENERATED, QC_PASSED, APPROVED, REJECTED, FAILED]
class ReviewVerdict:
PENDING = "pending"
PASS = "pass"
FAIL = "fail"
class AssetKind:
PORTRAIT = "portrait" # 定妆图
TURNAROUND = "turnaround" # 三视图
COSTUME = "costume" # 服装
EXPRESSION = "expression" # 表情/情绪参考
VOICE_SAMPLE = "voice_sample" # 音色样本
REFERENCE = "reference" # 通用参考
SCENE_IMAGE = "scene_image" # 场景图
KEYFRAME = "keyframe" # 分镜关键帧
LAST_FRAME = "last_frame" # 尾帧(首尾帧模型)
REFERENCE_VIDEO = "reference_video"
REFERENCE_IMAGE = "reference_image"
BGM = "bgm" # 配乐
SFX = "sfx"
VOICEOVER = "voiceover" # 已生成的配音
PREVIEW = "preview_video" # 审片预览(模型画面 + 平台配音)
OUTPUT = "output" # 成片
HUMAN = "human"
SCENE = "scene"
SHOT = "shot"
PROJECT = "project"
OWNER_BY_KIND = {
PORTRAIT: HUMAN, TURNAROUND: HUMAN, COSTUME: HUMAN, EXPRESSION: HUMAN,
VOICE_SAMPLE: HUMAN, REFERENCE: HUMAN,
SCENE_IMAGE: SCENE,
KEYFRAME: SHOT, LAST_FRAME: SHOT, REFERENCE_VIDEO: SHOT, REFERENCE_IMAGE: SHOT, VOICEOVER: SHOT,
PREVIEW: SHOT,
BGM: PROJECT, SFX: PROJECT, OUTPUT: PROJECT,
}
IMAGE_KINDS = {PORTRAIT, TURNAROUND, COSTUME, EXPRESSION, SCENE_IMAGE, KEYFRAME, LAST_FRAME, REFERENCE_IMAGE, REFERENCE}
AUDIO_KINDS = {VOICE_SAMPLE, VOICEOVER, BGM, SFX}
VIDEO_KINDS = {REFERENCE_VIDEO, OUTPUT, PREVIEW}
LABELS = {
PORTRAIT: "定妆图", TURNAROUND: "三视图", COSTUME: "服装", EXPRESSION: "表情参考",
VOICE_SAMPLE: "音色样本", REFERENCE: "参考资料", SCENE_IMAGE: "场景图",
KEYFRAME: "关键帧", LAST_FRAME: "尾帧", REFERENCE_VIDEO: "参考视频",
REFERENCE_IMAGE: "参考图", BGM: "配乐", SFX: "音效", VOICEOVER: "配音", OUTPUT: "成片",
PREVIEW: "审片预览",
}
class ShotSize:
EXTREME_WIDE = "extreme_wide"
WIDE = "wide"
FULL = "full"
MEDIUM = "medium"
CLOSE_UP = "close_up"
EXTREME_CLOSE_UP = "extreme_close_up"
ALL = [EXTREME_WIDE, WIDE, FULL, MEDIUM, CLOSE_UP, EXTREME_CLOSE_UP]
LABELS = {EXTREME_WIDE: "大远景", WIDE: "远景", FULL: "全景", MEDIUM: "中景",
CLOSE_UP: "近景", EXTREME_CLOSE_UP: "特写"}
class CameraMovement:
STATIC = "static"
PUSH_IN = "push_in"
PULL_OUT = "pull_out"
PAN_LEFT = "pan_left"
PAN_RIGHT = "pan_right"
TILT_UP = "tilt_up"
TILT_DOWN = "tilt_down"
DOLLY_IN = "dolly_in"
DOLLY_OUT = "dolly_out"
TRACK = "track"
ORBIT = "orbit"
HANDHELD = "handheld"
CRANE = "crane"
ZOOM_IN = "zoom_in"
ZOOM_OUT = "zoom_out"
ALL = [STATIC, PUSH_IN, PULL_OUT, PAN_LEFT, PAN_RIGHT, TILT_UP, TILT_DOWN,
DOLLY_IN, DOLLY_OUT, TRACK, ORBIT, HANDHELD, CRANE, ZOOM_IN, ZOOM_OUT]
LABELS = {STATIC: "固定", PUSH_IN: "推镜", PULL_OUT: "拉镜", PAN_LEFT: "左摇", PAN_RIGHT: "右摇",
TILT_UP: "上摇", TILT_DOWN: "下摇", DOLLY_IN: "推轨", DOLLY_OUT: "拉轨", TRACK: "移镜",
ORBIT: "环绕", HANDHELD: "手持", CRANE: "升降", ZOOM_IN: "推焦", ZOOM_OUT: "拉焦"}
class CameraAngle:
EYE_LEVEL = "eye_level"
HIGH = "high"
LOW = "low"
BIRD_EYE = "bird_eye"
DUTCH = "dutch"
OVER_SHOULDER = "over_shoulder"
ALL = [EYE_LEVEL, HIGH, LOW, BIRD_EYE, DUTCH, OVER_SHOULDER]
LABELS = {EYE_LEVEL: "平视", HIGH: "俯拍", LOW: "仰拍", BIRD_EYE: "鸟瞰", DUTCH: "倾斜", OVER_SHOULDER: "过肩"}
class Transition:
CUT = "cut"
DISSOLVE = "dissolve"
FADE_IN = "fade_in"
FADE_OUT = "fade_out"
FADE_THROUGH_BLACK = "fade_through_black"
WIPE = "wipe"
ALL = [CUT, DISSOLVE, FADE_IN, FADE_OUT, FADE_THROUGH_BLACK, WIPE]
LABELS = {CUT: "直切", DISSOLVE: "叠化", FADE_IN: "淡入", FADE_OUT: "淡出",
FADE_THROUGH_BLACK: "黑场过渡", WIPE: "擦除"}
class GenerationMode:
"""视频生成模式:镜头 mode 字段与模型 capabilities.modes 共用同一套取值。"""
T2V = "t2v" # 纯文生视频(不需要关键帧)
I2V = "i2v" # 图生视频(首帧 = 关键帧)
FL = "fl" # 首尾帧
ALL = [T2V, I2V, FL]
LABELS = {T2V: "文生视频", I2V: "图生视频(关键帧)", FL: "首尾帧"}
# ---------------------------------------------------------------- 声音话题词
# MiniMax H3 这类「音画一次成片」的联合音视频模型,把整段提示词同时当成
# **画面脚本**和**声音脚本**:凡是句子里出现「说 / 话 / 台词 / 为准 / 音画不同步」
# 这类**关于说话与声音的句子**,它就去念给你听。项目 11 实测(ASR 逐段对齐):
# · 台词镜头开头 0-2.8s 的乱语 = 提示词里那句「……以定妆图/三视图为准,不得按文字另行改写衣着」;
# · 写「这一段没有人说话」,成片末尾真的冒出一句「没有人说话」;
# · 纯视觉 caption(同样的画面、同样的参考图)ASR 几乎听不到人声。
# 结论:**给这类模型的提示词里不许出现任何关于声音/指令的句子**,
# 想要什么环境声,只写「听得见的名词短语」(键盘敲击声、房间底噪)。
# 「嘴/唇/口」是发声器官词:shot 36 的 caption 里唯一提到嘴的那句
# (「俊杰正面朝向镜头,嘴部张开,是画面的主体;紫小慧…嘴部闭合」)被逐字念了出来——
# 词级 ASR 8+7 个音节、中间 1.3s 处一个停顿,和这句的两个短句一一对应,
# 台词直到 2.68s 才开始;shot 35 的 caption 一个「嘴」字都没有,开头就是干净的。
# 所以「谁在张嘴」只能画在首帧里,不能写进给 H3 的文字。
AUDIO_TOPIC_MARKERS = ("说话", "说", "讲", "唱", "念", "吼", "喊", "提问", "问答",
"嘴", "唇", "张口", "开口", "抿嘴",
"台词", "对白", "旁白", "独白", "配音", "口型", "音色", "嗓音", "语气",
"人声", "声音", "噪音", "杂音", "静音", "无声", "沉默", "音画", "音效",
"音乐", "背景音", "音量", "音频", "声线", "字幕", "同步", "对话",
"bgm", "audio", "sound", "voice", "speech", "music", "narration")
# 提示词里的「元指令」词:模型会把它当成有人在下指令,同样容易被念出来。
META_INSTRUCTION_MARKERS = ("必须", "不得", "不要", "不要出现", "不准", "不能", "严禁", "禁止",
"为准", "以免", "避免", "防止", "要求", "修正", "上一轮", "上一版",
"质检", "提示词", "严格", "应当", "应该", "需要注意的是", "返工",
"画面中", "镜头内", "保持造型", "一致性", "逐件", "自行改写", "另行改写")
def is_audio_topic(text: str) -> bool:
"""这句话在讲「声音/说话」本身(而不是画面里看得见的内容)。"""
probe = str(text or "").strip().lower()
if not probe:
return False
return any(marker in probe for marker in AUDIO_TOPIC_MARKERS)
def is_meta_instruction(text: str) -> bool:
"""这句话是「给系统下的指令/约束」,不是画面内容描述。"""
probe = str(text or "").strip()
if not probe:
return False
return any(marker in probe for marker in META_INSTRUCTION_MARKERS)
def split_phrases(text: str) -> list[str]:
"""按标点切成短句/短语(保留顺序,丢弃空串)。"""
return [piece.strip() for piece in re.split(r"[,、。;;!!??\n,]+", str(text or "")) if piece.strip()]
def drop_topic(text: str, *, audio: bool = True, meta: bool = True) -> str:
"""逐句过滤:删掉「关于声音」和「下指令」的短语,剩下的才是能安全喂给联合音视频模型的内容。"""
kept = []
for piece in split_phrases(text):
if audio and is_audio_topic(piece):
continue
if meta and is_meta_instruction(piece):
continue
kept.append(piece)
return ",".join(kept)
class AudioTrackKind:
VOICEOVER = "voiceover"
BGM = "bgm"
SFX = "sfx"
NATIVE = "native" # 模型原生音画同步音轨(随镜头片段串接,不是独立音轨)
RESOLUTIONS = {"360p": (640, 360), "480p": (854, 480), "720p": (1280, 720),
"1080p": (1920, 1080), "2k": (2560, 1440), "4k": (3840, 2160)}
ASPECT_RATIOS = {"16:9": 16 / 9, "9:16": 9 / 16, "1:1": 1.0, "4:3": 4 / 3, "3:4": 3 / 4, "21:9": 21 / 9}
def canvas_for(resolution: str, aspect_ratio: str) -> tuple[int, int]:
"""按分辨率与画幅计算输出宽高(偶数,便于编码)。
档位号的含义是**横向档位的短边(高度)**,竖屏必须沿用同一条规则:
720p 9:16 是 720x1280,不是 404x720。竖屏分支曾经把档位号当长边用
(h=720、w=404),于是时间线母版、关键帧、喂给视频模型的首帧
全部只有 404x720——模型原生出的 736x1312 被砍掉 3.2 倍面积,
再放大回 720x1280 交付,成片的有效细节只剩四分之一。
导出预设 `vertical_720p=720x1280` 一直是按短边算的,修完之后两边才对得上。
"""
base_w, base_h = RESOLUTIONS.get(resolution, RESOLUTIONS["1080p"])
ratio = ASPECT_RATIOS.get(aspect_ratio, 16 / 9)
if ratio >= 1:
w, h = base_w, base_w / ratio # 横屏:长边 = 档位宽度
else:
w, h = base_h, base_h / ratio # 竖屏:短边 = 档位高度
w, h = int(w // 2 * 2), int(h // 2 * 2)
return max(2, w), max(2, h)
def label(mapping: dict[str, str], key: str | None, default: str = "") -> str:
if not key:
return default
return mapping.get(key, key)
# ---------------------------------------------------------- 镜头语言字段归一
# 每个字段只认自己那一组的枚举。曾经四个字段共用一张白名单,别的组的值就能直接
# 过关——project 11 的 S03「景别」里存的就是机位 over_shoulder:前端下拉框显示
# 不出来,提示词里还漏出一个英文原词。联合音视频模型(MiniMax H3)会把提示词里的
# 每句话都当成可能出声的台词,脏值等于凭空多一段杂音。
GROUP_VALUES = {"shot_size": ShotSize.ALL, "camera_angle": CameraAngle.ALL,
"camera_movement": CameraMovement.ALL, "transition": Transition.ALL}
GROUP_DEFAULTS = {"shot_size": ShotSize.MEDIUM, "camera_angle": CameraAngle.EYE_LEVEL,
"camera_movement": CameraMovement.STATIC, "transition": Transition.CUT}
CAMERA_GROUPS = ("shot_size", "camera_angle", "camera_movement")
def normalize_camera_fields(shot_size: str = "", camera_angle: str = "",
camera_movement: str = "") -> tuple[str, str, str]:
"""清洗镜头语言三件套,并把「写错列」的值搬回正确的字段。
大模型导出的分镜里「过肩」这类机位值经常被塞进「景别」列。只要该值确实属于
另一组、而那一组正好空着,就搬过去;否则交给调用方的同义词表折算,最后落回
默认值。既不丢编导的原意,也不会让脏值漏进提示词。
返回三元组,识别不出且无同义词表时给出空串(由调用方决定默认值)。
"""
raws = dict(zip(CAMERA_GROUPS, ((shot_size or "").strip(), (camera_angle or "").strip(),
(camera_movement or "").strip())))
vals = {g: (v if v in GROUP_VALUES[g] else "") for g, v in raws.items()}
for group, value in list(raws.items()):
if not value or vals[group]:
continue # 本来就写在正确的列上
for other in CAMERA_GROUPS:
if other == group or value not in GROUP_VALUES[other]:
continue
# 目标列空着、或填的只是默认值(eye_level / static 多半不是刻意选的)才搬:
# 「景别=过肩 + 机位=平视」就是典型的填错列,默认值让位给有信息量的值。
if vals[other] and vals[other] != GROUP_DEFAULTS[other]:
continue
vals[other] = value
break
# 认不出归属的原值一律留空,交给调用方(同义词表 → 默认值);
# 已经被搬到正确列的值只在那一列出现一次,不会同时污染两列。
return tuple(vals[g] for g in CAMERA_GROUPS)