nex_animation/backend/app/services/preview.py

214 lines
11 KiB
Python
Raw Permalink 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.

"""镜头级「审片预览」:让评审听到的语音和成片一致。
语音来源跟随时间线设置 ``voice_source``:
- 走 TTS 的镜头 → 把模型自带音轨换成平台配音(画面直接 copy,不重编码);
- 走模型原声的镜头(原生音画同步)→ 预览就是镜头产物本身,不再多此一举换声。
否则「预览里是对的、成片里是错的」(或反过来),评审结论全废。
MiniMax H3 这类音画同出的模型,成片里自带一条语音轨,音色由模型即兴决定,
与项目登记的克隆声线无关(实测会把六十岁老船夫念成年轻女声,基频 110Hz → 224Hz)。
评审在分镜/审核页只能听到这条模型原声,就会把「配音」误判成错的;
真正的成片语音(IndexTTS2 声线)过去只有在导出成片后才听得到。
平台能力补齐:用 ffmpeg 把镜头视频轨(直接 copy,不重编码、不损画质)
与镜头配音重新封装成预览,配音入点与时间线一致(lead_padding_ms),
尾部空档补静音,保证音轨与画面等长。
"""
from __future__ import annotations
import hashlib
import secrets
from sqlalchemy.orm import Session
from app.constants import AssetKind
from app.models import Asset, Project, Shot, Timeline
from app.services import events, storage
from app.services import ffmpeg as ff
from app.services.timeline import DEFAULT_LEAD_MS, uses_native_voice
KIND = AssetKind.PREVIEW
# 镜头产物是否自带音轨只需要探测一次(按文件指纹缓存),
# 否则分镜列表每个镜头都要多跑一次 ffprobe,列表页会明显变慢。
_audio_cache: dict[str, bool] = {}
def has_native_audio(path: str | None) -> bool:
if not path or not storage.exists(path):
return False
key = _stamp(path)
cached = _audio_cache.get(key)
if cached is not None:
return cached
try:
flag = bool(ff.probe(storage.to_absolute(path)).has_audio)
except ff.FFmpegError:
flag = False
if len(_audio_cache) > 512:
_audio_cache.clear()
_audio_cache[key] = flag
return flag
def lead_ms_of(db: Session, project_id: int) -> int:
"""配音在镜头内的入点留白:跟随项目时间线设置,缺省用时间线默认值。"""
timeline = db.query(Timeline).filter(Timeline.project_id == project_id) \
.order_by(Timeline.id.desc()).first()
conf = (timeline.settings_json or {}) if timeline else {}
try:
return max(0, int(conf.get("lead_padding_ms", DEFAULT_LEAD_MS)))
except (TypeError, ValueError):
return DEFAULT_LEAD_MS
def _stamp(path: str | None) -> str:
if not path or not storage.exists(path):
return "0:0"
info = storage.to_absolute(path).stat()
return f"{int(info.st_mtime)}:{info.st_size}"
def timeline_conf(db: Session, project_id: int) -> dict:
"""当前时间线的语音/字幕设置;还没排过时间线时用出厂默认。"""
from app.services.timeline import DEFAULT_TIMELINE_SETTINGS
row = db.query(Timeline).filter(Timeline.project_id == project_id) \
.order_by(Timeline.id.desc()).first()
if row is None:
return dict(DEFAULT_TIMELINE_SETTINGS)
return {**DEFAULT_TIMELINE_SETTINGS, **(row.settings_json or {})}
def voice_source_of(db: Session, shot: Shot) -> str:
"""这条镜头的台词在成片里到底用谁的声音:native(模型原声)/ dub(平台配音)。"""
return "native" if uses_native_voice(shot, timeline_conf(db, shot.project_id)) else "dub"
def signature(db: Session, project_id: int, shot: Shot) -> str:
"""预览内容指纹:画面、配音、入点、语音来源任一变化都必须重新封装。"""
raw = (f"{_stamp(shot.output_path)}|{_stamp(shot.voiceover_path)}|{lead_ms_of(db, project_id)}"
f"|{voice_source_of(db, shot)}")
return hashlib.sha256(raw.encode()).hexdigest()[:16]
def _asset_for(db: Session, shot: Shot) -> Asset | None:
return db.query(Asset).filter(Asset.project_id == shot.project_id, Asset.owner_type == "shot",
Asset.owner_id == shot.id, Asset.kind == KIND) \
.order_by(Asset.id.desc()).first()
def preview_state(db: Session, shot: Shot) -> dict:
"""不改文件地回答「现在有没有可用的审片预览」,供镜头负载与前端渲染。"""
native = ff.available() and has_native_audio(shot.output_path)
asset = _asset_for(db, shot)
ready = bool(asset and storage.exists(asset.path)
and (asset.metadata_json or {}).get("signature") == signature(db, shot.project_id, shot))
return {
"preview_url": storage.url_for(asset.path) if ready else "",
"preview_path": asset.path if ready else "",
"preview_ready": ready,
"native_audio": native,
"voiceover_ready": bool(storage.exists(shot.voiceover_path)),
# 成片与预览的语音口径:native = 保留模型原声(音画同步),dub = 用平台配音
"voice_source": voice_source_of(db, shot),
}
def build_shot_preview(db: Session, project: Project, shot: Shot, *, force: bool = False,
actor: str = "studio", commit: bool = True) -> dict:
"""生成(或复用)镜头审片预览:画面 = 模型产物,声音 = 平台配音。"""
if not storage.exists(shot.output_path):
return {"ok": False, "shot_id": shot.id, "reason": "missing_output",
"message": "镜头还没有生成视频,无法封装审片预览"}
if voice_source_of(db, shot) == "native":
# 台词本来就用模型原声:预览无需换声,直接给镜头产物,避免「预览换了嗓子」
return {"ok": True, "shot_id": shot.id, "reused": True, "mode": "native",
"path": storage.to_relative(shot.output_path),
"url": storage.url_for(shot.output_path),
"message": "该镜头台词使用模型原声(原生音画同步),预览即镜头产物"}
sig = signature(db, project.id, shot)
existing = _asset_for(db, shot)
if (not force and existing and storage.exists(existing.path)
and (existing.metadata_json or {}).get("signature") == sig):
return {"ok": True, "shot_id": shot.id, "reused": True, "mode": "voiceover",
"path": existing.path, "url": storage.url_for(existing.path),
"duration_ms": existing.duration_ms or 0}
if not storage.exists(shot.voiceover_path):
# 没有配音就没有可替换的语音,直接用模型产物(并说明原因)
return {"ok": True, "shot_id": shot.id, "reused": True, "mode": "native",
"path": storage.to_relative(shot.output_path),
"url": storage.url_for(shot.output_path),
"message": "镜头尚无配音,预览使用模型自带音轨"}
if not ff.available():
return {"ok": False, "shot_id": shot.id, "reason": "ffmpeg_missing",
"message": "本机 ffmpeg 不可用,无法封装审片预览"}
lead = lead_ms_of(db, project.id)
folder = storage.project_dir(project.id, "preview")
out = folder / f"shot-{shot.order_index:02d}-preview_{secrets.token_hex(5)}.mp4"
video = storage.to_absolute(shot.output_path)
voice = storage.to_absolute(shot.voiceover_path)
# 预览长度以画面为准。apad 补的是「无限」静音,必须同时给出 whole_dur 并
# 用 -t 封口:apad + -c:v copy 的组合下 -shortest 不会触发,
# 封装进程会停在 100% CPU 一直编码,产物永远只有文件头 48 字节(实测挂死)。
try:
video_s = round(float(ff.probe(video).duration or 0), 3)
except ff.FFmpegError:
video_s = 0.0
if video_s <= 0:
return {"ok": False, "shot_id": shot.id, "reason": "output_unreadable",
"message": "镜头视频读不到时长,无法封装审片预览"}
delay = f"adelay=delays={lead}:all=1," if lead > 0 else ""
ff.run([ff.FFMPEG, "-hide_banner", "-loglevel", "error", "-y",
"-i", str(video), "-i", str(voice),
"-filter_complex", f"[1:a]{delay}apad=whole_dur={video_s}[previewa]",
"-map", "0:v:0", "-map", "[previewa]",
"-c:v", "copy", "-c:a", "aac", "-b:a", "160k",
"-t", f"{video_s:.3f}", "-movflags", "+faststart", str(out)], timeout=600)
info = ff.probe(out).as_dict()
duration_ms = int(round(float(info.get("duration") or 0) * 1000))
if duration_ms <= 0:
out.unlink(missing_ok=True)
return {"ok": False, "shot_id": shot.id, "reason": "preview_unreadable",
"message": "审片预览封装完成但读不到时长"}
rel = storage.to_relative(out)
asset = storage.register_asset(
db, project_id=project.id, owner_type="shot", owner_id=shot.id, kind=KIND,
path=rel, name=f"镜头 {shot.order_index:02d} 审片预览", source="compose",
info={"duration_ms": duration_ms, "size_bytes": out.stat().st_size, "mime": "video/mp4",
"width": info.get("width") or 0, "height": info.get("height") or 0},
metadata={"signature": sig, "lead_ms": lead, "audio_source": "voiceover",
"video_source": storage.to_relative(shot.output_path),
"voiceover_path": storage.to_relative(shot.voiceover_path)})
events.log_event(db, "preview.built", project_id=project.id, shot_id=shot.id,
message=f"镜头 {shot.order_index:02d} 审片预览封装完成(配音替换模型原声)",
entity_type="shot", entity_id=shot.id,
data={"path": rel, "duration_ms": duration_ms, "lead_ms": lead,
"voiceover": storage.to_relative(shot.voiceover_path)}, actor=actor)
if commit:
db.commit()
else:
db.flush()
return {"ok": True, "shot_id": shot.id, "reused": False, "mode": "voiceover", "path": rel,
"url": storage.url_for(rel), "duration_ms": duration_ms,
"version": asset.current_version, "lead_ms": lead}
def build_project_previews(db: Session, project: Project, *, shot_ids: list[int] | None = None,
force: bool = False, actor: str = "studio") -> dict:
"""批量封装:调度台与验收脚本一次把全片镜头的审片预览准备好。"""
wanted = set(shot_ids) if shot_ids else None
items = []
for shot in sorted(project.shots or [], key=lambda s: (s.order_index, s.id)):
if wanted is not None and shot.id not in wanted:
continue
items.append(build_shot_preview(db, project, shot, force=force, actor=actor))
db.commit()
return {"project_id": project.id, "count": len(items),
"built": sum(1 for i in items if i.get("ok") and not i.get("reused")),
"reused": sum(1 for i in items if i.get("ok") and i.get("reused")),
"failed": sum(1 for i in items if not i.get("ok")), "items": items}