817 lines
33 KiB
Python
817 lines
33 KiB
Python
"""KTV pipeline step handlers — 通过 llmage API 调用 GPU 服务。
|
||
|
||
配置方式（环境变量）：
|
||
KTV_API_BASE: llmage API 地址，默认 https://token.opencomputing.cn/llmage/v1
|
||
KTV_API_KEY: Bearer token
|
||
KTV_FILE_BASE: 文件服务地址，默认同 KTV_API_BASE 所在域
|
||
|
||
步骤→模型映射（KTV_STEP_MODELS 字典，可通过环境变量覆盖各项）:
|
||
demucs_separating → ky-demucs-separate
|
||
asr_transcribing → ky-asr-transcribe
|
||
subtitle_rendering → ky-subtitle-render
|
||
face_detecting → ky-face-detect
|
||
realesrgan_upscaling → ky-realesrgan-upscale
|
||
rvc_converting → ky-rvc-convert
|
||
video_evaluating → ky-video-eval-evaluate
|
||
songrate_evaluating → ky-songrate-evaluate
|
||
|
||
本地操作（ffmpeg/ffprobe）不走 API。
|
||
"""
|
||
|
||
import asyncio
|
||
import json
|
||
import os
|
||
import logging
|
||
import time
|
||
import aiohttp
|
||
|
||
from app.quality_gate import (
|
||
eval_music_quality, eval_character_design, eval_character_image,
|
||
eval_storyboard, eval_ktv_synthesis, eval_scene_video_quality,
|
||
)
|
||
from app.model_selector import handle_model_selecting
|
||
|
||
logger = logging.getLogger("pipeline.handlers.ktv")
|
||
|
||
# ── config ──────────────────────────────────────────────────────────
|
||
|
||
API_BASE = os.environ.get("KTV_API_BASE", "https://token.opencomputing.cn/llmage/v1")
|
||
API_KEY = os.environ.get("KTV_API_KEY", "")
|
||
FILE_BASE = os.environ.get("KTV_FILE_BASE", API_BASE.rsplit("/", 2)[0])
|
||
|
||
STEP_MODELS = {
|
||
"demucs_separating": os.environ.get("KTV_MODEL_DEMUCS", "ky-demucs-separate"),
|
||
"asr_transcribing": os.environ.get("KTV_MODEL_ASR", "ky-asr-transcribe"),
|
||
"subtitle_rendering": os.environ.get("KTV_MODEL_SUBTITLE", "ky-subtitle-render"),
|
||
"face_detecting": os.environ.get("KTV_MODEL_FACE", "ky-face-detect"),
|
||
"realesrgan_upscaling": os.environ.get("KTV_MODEL_REALESRGAN","ky-realesrgan-upscale"),
|
||
"rvc_converting": os.environ.get("KTV_MODEL_RVC", "ky-rvc-convert"),
|
||
"video_evaluating": os.environ.get("KTV_MODEL_VIDEO_EVAL","ky-video-eval-evaluate"),
|
||
"songrate_evaluating": os.environ.get("KTV_MODEL_SONGRATE", "ky-songrate-evaluate"),
|
||
}
|
||
|
||
LOCAL_WORK_DIR = "/data/pipeline/ktv"
|
||
GPU_HOST = None # 不再直连 GPU，保留变量兼容旧引用，新代码不应使用
|
||
|
||
|
||
def _task_dir(task_id: str) -> str:
|
||
d = os.path.join(LOCAL_WORK_DIR, task_id)
|
||
os.makedirs(d, exist_ok=True)
|
||
return d
|
||
|
||
|
||
# ── API client ─────────────────────────────────────────────────────
|
||
|
||
class KtvClient:
|
||
"""llmage pipeline API 客户端。"""
|
||
|
||
def __init__(self, base: str = "", key: str = ""):
|
||
self.base = base or API_BASE
|
||
self.key = key or API_KEY
|
||
self._session = None
|
||
|
||
async def _ensure_session(self):
|
||
if self._session is None:
|
||
self._session = aiohttp.ClientSession(
|
||
headers={"Authorization": f"Bearer {self.key}",
|
||
"Content-Type": "application/json"},
|
||
timeout=aiohttp.ClientTimeout(total=600),
|
||
)
|
||
|
||
async def close(self):
|
||
if self._session:
|
||
await self._session.close()
|
||
self._session = None
|
||
|
||
async def call(self, model: str, params: dict) -> dict:
|
||
"""同步调用，返回 GPU 响应 dict。"""
|
||
await self._ensure_session()
|
||
body = {"model": model, **params}
|
||
url = f"{self.base}/pipeline/submit"
|
||
async with self._session.post(url, json=body) as resp:
|
||
data = await resp.json()
|
||
if resp.status != 200:
|
||
raise RuntimeError(f"KTV API {model} HTTP {resp.status}: {data}")
|
||
if data.get("taskstatus") == "PENDING":
|
||
# 异步任务 → 轮询等待
|
||
taskid = data.get("taskid")
|
||
return await self._poll(taskid)
|
||
return data
|
||
|
||
async def _poll(self, taskid: str, interval: int = 5, max_wait: int = 600) -> dict:
|
||
"""轮询异步任务直到完成。"""
|
||
await self._ensure_session()
|
||
deadline = time.time() + max_wait
|
||
while time.time() < deadline:
|
||
await asyncio.sleep(interval)
|
||
url = f"{self.base}/tasks?taskid={taskid}"
|
||
async with self._session.get(url) as resp:
|
||
data = await resp.json()
|
||
ts = data.get("taskstatus", data.get("status", ""))
|
||
if ts in ("SUCCEEDED", "FAILED"):
|
||
return data
|
||
raise TimeoutError(f"Task {taskid} timeout after {max_wait}s")
|
||
|
||
|
||
# 全局单例，惰性初始化
|
||
_ktv_client: KtvClient | None = None
|
||
|
||
|
||
def get_ktv_client() -> KtvClient:
|
||
global _ktv_client
|
||
if _ktv_client is None:
|
||
_ktv_client = KtvClient()
|
||
return _ktv_client
|
||
|
||
|
||
# ── helpers ────────────────────────────────────────────────────────
|
||
|
||
async def _run_local(cmd: str, timeout: int = 300) -> tuple[str, str, int]:
|
||
proc = await asyncio.create_subprocess_shell(
|
||
cmd, stdout=asyncio.subprocess.PIPE, stderr=asyncio.subprocess.PIPE)
|
||
try:
|
||
stdout, stderr = await asyncio.wait_for(proc.communicate(), timeout=timeout)
|
||
return stdout.decode("utf-8", errors="replace"), stderr.decode("utf-8", errors="replace"), proc.returncode
|
||
except asyncio.TimeoutError:
|
||
proc.kill()
|
||
return "", "timeout", -1
|
||
|
||
|
||
async def _download(url: str, dest: str) -> None:
|
||
"""下载文件到本地。"""
|
||
async with aiohttp.ClientSession() as s:
|
||
async with s.get(url) as r:
|
||
r.raise_for_status()
|
||
with open(dest, "wb") as f:
|
||
async for chunk in r.content.iter_chunked(8192):
|
||
f.write(chunk)
|
||
|
||
|
||
# ── Media Preparation ──────────────────────────────────────────────
|
||
|
||
async def handle_audio_preparing(tenant_id, task_id, step_name, input_data, config):
|
||
work_dir = _task_dir(task_id)
|
||
params = input_data.get("task_params", {})
|
||
audio_url = params.get("audio_url", params.get("audio_path", ""))
|
||
if not audio_url:
|
||
raise ValueError("缺少 audio_url 或 audio_path 参数")
|
||
|
||
audio_path = os.path.join(work_dir, "original_audio.mp3")
|
||
if audio_url.startswith("http"):
|
||
await _download(audio_url, audio_path)
|
||
else:
|
||
await _run_local(f"cp '{audio_url}' '{audio_path}'")
|
||
|
||
stdout, _, rc = await _run_local(
|
||
f"ffprobe -v error -show_entries format=duration "
|
||
f"-of default=noprint_wrappers=1:nokey=1 '{audio_path}'")
|
||
duration = float(stdout.strip()) if rc == 0 else 0
|
||
|
||
return {"audio_path": audio_path, "duration": duration,
|
||
"format": os.path.splitext(audio_path)[1].lstrip(".")}
|
||
|
||
|
||
async def handle_video_preparing(tenant_id, task_id, step_name, input_data, config):
|
||
work_dir = _task_dir(task_id)
|
||
params = input_data.get("task_params", {})
|
||
video_url = params.get("video_url", params.get("video_path", ""))
|
||
if not video_url:
|
||
raise ValueError("缺少 video_url 或 video_path 参数")
|
||
|
||
video_path = os.path.join(work_dir, "original_video.mp4")
|
||
audio_path = os.path.join(work_dir, "original_audio.mp3")
|
||
if video_url.startswith("http"):
|
||
await _download(video_url, video_path)
|
||
else:
|
||
await _run_local(f"cp '{video_url}' '{video_path}'")
|
||
|
||
await _run_local(f"ffmpeg -y -i '{video_path}' -vn -acodec libmp3lame -q:a 2 '{audio_path}'")
|
||
stdout, _, rc = await _run_local(
|
||
f"ffprobe -v error -show_entries format=duration "
|
||
f"-of default=noprint_wrappers=1:nokey=1 '{video_path}'")
|
||
duration = float(stdout.strip()) if rc == 0 else 0
|
||
|
||
return {"video_path": video_path, "audio_path": audio_path, "duration": duration}
|
||
|
||
|
||
# ── Demucs ─────────────────────────────────────────────────────────
|
||
|
||
async def handle_demucs_separating(tenant_id, task_id, step_name, input_data, config):
|
||
"""通过 API 调 ky-demucs-separate。固定使用 separate_full 模式。"""
|
||
work_dir = _task_dir(task_id)
|
||
audio_path = None
|
||
for dep_output in input_data.values():
|
||
if isinstance(dep_output, dict):
|
||
audio_path = dep_output.get("audio_path")
|
||
if audio_path: break
|
||
if not audio_path:
|
||
raise ValueError("上游步骤未提供 audio_path")
|
||
|
||
client = get_ktv_client()
|
||
result = await client.call(STEP_MODELS["demucs_separating"],
|
||
{"audio_file": audio_path,
|
||
"task_type": "separate_full"})
|
||
|
||
if result.get("error"):
|
||
raise ValueError(f"Demucs 分离失败: {result['error']}")
|
||
|
||
# 下载结果文件
|
||
vocals_local = os.path.join(work_dir, "vocals.wav")
|
||
accompaniment_local = os.path.join(work_dir, "accompaniment.wav")
|
||
if result.get("vocals_url"):
|
||
await _download(result["vocals_url"], vocals_local)
|
||
if result.get("accompaniment_url"):
|
||
await _download(result["accompaniment_url"], accompaniment_local)
|
||
|
||
return {"vocals_path": vocals_local,
|
||
"no_vocals_path": accompaniment_local,
|
||
"accompaniment_path": accompaniment_local,
|
||
"mode": "separate_full",
|
||
"usage": result.get("usage", {})}
|
||
|
||
|
||
# ── Lyric Calibration ──────────────────────────────────────────────
|
||
|
||
async def handle_lyric_calibrating(tenant_id, task_id, step_name, input_data, config):
|
||
"""ASR 通过 API + LLM 校准。"""
|
||
work_dir = _task_dir(task_id)
|
||
params = input_data.get("task_params", {})
|
||
lyrics_text = params.get("lyrics", params.get("lyrics_text", ""))
|
||
|
||
# 找人声文件
|
||
vocals_path = None
|
||
for dep_output in input_data.values():
|
||
if isinstance(dep_output, dict):
|
||
vocals_path = dep_output.get("vocals_path")
|
||
if vocals_path: break
|
||
if not vocals_path:
|
||
raise ValueError("上游步骤未提供 vocals_path")
|
||
if not lyrics_text:
|
||
raise ValueError("缺少 lyrics 参数")
|
||
|
||
# 通过 API 做 ASR
|
||
client = get_ktv_client()
|
||
asr_result = await client.call(STEP_MODELS["asr_transcribing"],
|
||
{"audio_file": vocals_path})
|
||
|
||
asr_timings = asr_result.get("segments", []) if not asr_result.get("error") else []
|
||
if not asr_timings:
|
||
logger.warning(f"ASR 未返回时间戳，使用空列表: {asr_result.get('error','')}")
|
||
|
||
# LLM 校准
|
||
calibrated = await _llm_calibrate(lyrics_text, asr_timings)
|
||
|
||
calibrated_path = os.path.join(work_dir, "calibrated_lyrics.json")
|
||
with open(calibrated_path, "w", encoding="utf-8") as f:
|
||
json.dump(calibrated, f, ensure_ascii=False, indent=2)
|
||
|
||
return {"calibrated_lyrics_path": calibrated_path,
|
||
"calibrated_lyrics": calibrated,
|
||
"segment_count": len(calibrated)}
|
||
|
||
|
||
async def _llm_calibrate(lyrics_text: str, asr_timings: list) -> list:
|
||
prompt = f"""你是一个歌词时间轴校准专家。
|
||
|
||
原始歌词文本:
|
||
{lyrics_text}
|
||
|
||
ASR识别的时间戳（秒）:
|
||
{json.dumps(asr_timings, ensure_ascii=False)}
|
||
|
||
请将原始歌词的每一句与ASR时间戳对齐，输出JSON数组，每个元素包含:
|
||
- text: 歌词文本
|
||
- start: 开始时间（秒，浮点数）
|
||
- end: 结束时间（秒，浮点数）
|
||
|
||
要求:
|
||
1. 保持原始歌词的文本和顺序
|
||
2. 时间戳以ASR结果为基础进行微调
|
||
3. 确保时间不重叠，每句之间留适当间隔
|
||
4. 只输出JSON，不要其他内容"""
|
||
try:
|
||
from pipeline_service.llm_bridge import llm_call
|
||
result = await llm_call(prompt)
|
||
result = result.strip()
|
||
if result.startswith("```"):
|
||
result = result.split("\n", 1)[1].rsplit("```", 1)[0]
|
||
return json.loads(result)
|
||
except Exception as e:
|
||
logger.warning(f"LLM calibration failed, using ASR timings directly: {e}")
|
||
return asr_timings
|
||
|
||
|
||
# ── Subtitle ───────────────────────────────────────────────────────
|
||
|
||
async def handle_subtitle_rendering(tenant_id, task_id, step_name, input_data, config):
|
||
"""本地生成 ASS 字幕（轻量操作，不走 API）。"""
|
||
work_dir = _task_dir(task_id)
|
||
calibrated = None
|
||
for dep_output in input_data.values():
|
||
if isinstance(dep_output, dict):
|
||
calibrated = dep_output.get("calibrated_lyrics")
|
||
if not calibrated and dep_output.get("calibrated_lyrics_path"):
|
||
with open(dep_output["calibrated_lyrics_path"], "r") as f:
|
||
calibrated = json.load(f)
|
||
if calibrated: break
|
||
if not calibrated:
|
||
raise ValueError("上游步骤未提供 calibrated_lyrics")
|
||
|
||
ass_path = os.path.join(work_dir, "karaoke.ass")
|
||
_write_ass_file(ass_path, calibrated)
|
||
return {"ass_path": ass_path}
|
||
|
||
|
||
def _write_ass_file(path: str, segments: list):
|
||
header = """[Script Info]
|
||
Title: KTV Karaoke Subtitles
|
||
ScriptType: v4.00+
|
||
PlayResX: 1920
|
||
PlayResY: 1080
|
||
WrapStyle: 0
|
||
|
||
[V4+ Styles]
|
||
Format: Name, Fontname, Fontsize, PrimaryColour, SecondaryColour, OutlineColour, BackColour, Bold, Italic, Underline, StrikeOut, ScaleX, ScaleY, Spacing, Angle, BorderStyle, Outline, Shadow, Alignment, MarginL, MarginR, MarginV, Encoding
|
||
Style: KTV,Source Han Sans SC,72,&H00FFFFFF,&H0000FFFF,&H00000000,&H80000000,-1,0,0,0,100,100,1,0,1,3,1,2,40,40,60,1
|
||
|
||
[Events]
|
||
Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text
|
||
"""
|
||
with open(path, "w", encoding="utf-8") as f:
|
||
f.write(header)
|
||
for seg in segments:
|
||
start = _seconds_to_ass_time(seg["start"])
|
||
end = _seconds_to_ass_time(seg["end"])
|
||
text = seg["text"].replace("\n", "\\N")
|
||
dur_ms = int((seg["end"] - seg["start"]) * 100)
|
||
f.write(f"Dialogue: 0,{start},{end},KTV,,0,0,0,,{{\\k{dur_ms}}}{text}\n")
|
||
|
||
|
||
def _seconds_to_ass_time(seconds: float) -> str:
|
||
h = int(seconds // 3600)
|
||
m = int((seconds % 3600) // 60)
|
||
s = int(seconds % 60)
|
||
cs = int((seconds % 1) * 100)
|
||
return f"{h}:{m:02d}:{s:02d}.{cs:02d}"
|
||
|
||
|
||
async def handle_subtitle_exporting(tenant_id, task_id, step_name, input_data, config):
|
||
ass_path = None
|
||
for dep_output in input_data.values():
|
||
if isinstance(dep_output, dict):
|
||
ass_path = dep_output.get("ass_path")
|
||
if ass_path: break
|
||
if not ass_path or not os.path.exists(ass_path):
|
||
raise ValueError("上游步骤未提供有效的 ass_path")
|
||
return {"subtitle_path": ass_path, "format": "ass"}
|
||
|
||
|
||
# ── Lyric Gen / Eval ───────────────────────────────────────────────
|
||
|
||
async def handle_lyric_generating(tenant_id, task_id, step_name, input_data, config):
|
||
params = input_data.get("task_params", {})
|
||
topic = params.get("topic", params.get("outline", ""))
|
||
style = params.get("style", "流行")
|
||
language = params.get("language", "zh")
|
||
if not topic:
|
||
raise ValueError("缺少 topic/outline参数")
|
||
|
||
prompt = f"""请创作一首{style}风格的{language}歌词。
|
||
|
||
主题/大纲: {topic}
|
||
|
||
要求:
|
||
1. 包含完整结构: 主歌(Verse)、副歌(Chorus)、桥段(Bridge)
|
||
2. 每句歌词节奏感强，适合演唱
|
||
3. 押韵自然，情感真实
|
||
4. 总长度适合3-5分钟歌曲
|
||
|
||
直接输出歌词文本，标注段落结构。"""
|
||
try:
|
||
from pipeline_service.llm_bridge import llm_call
|
||
lyrics = await llm_call(prompt)
|
||
return {"lyrics": lyrics.strip(), "topic": topic, "style": style}
|
||
except Exception as e:
|
||
raise ValueError(f"歌词生成失败: {e}")
|
||
|
||
|
||
async def handle_lyric_evaluating(tenant_id, task_id, step_name, input_data, config):
|
||
threshold = config.get("threshold", 8.5)
|
||
lyrics = None
|
||
for dep_output in input_data.values():
|
||
if isinstance(dep_output, dict):
|
||
lyrics = dep_output.get("lyrics")
|
||
if lyrics: break
|
||
if not lyrics:
|
||
raise ValueError("上游步骤未提供歌词")
|
||
|
||
prompt = f"""请从以下维度评估这首歌词的质量(1-10分):
|
||
|
||
1. 韵律节奏: 押韵、节奏感、可唱性
|
||
2. 情感表达: 情感真实度、共鸣力
|
||
3. 文学性: 用词、意象、修辞
|
||
4. 结构完整: 段落编排、层次感
|
||
5. 商业潜力: 流行度、记忆点
|
||
|
||
歌词:
|
||
{lyrics}
|
||
|
||
输出JSON: {{"score": 8.5, "dimensions": {{...}}, "suggestions": "..."}}
|
||
只输出JSON。"""
|
||
try:
|
||
from pipeline_service.llm_bridge import llm_call
|
||
result = await llm_call(prompt)
|
||
result = result.strip()
|
||
if result.startswith("```"):
|
||
result = result.split("\n", 1)[1].rsplit("```", 1)[0]
|
||
evaluation = json.loads(result)
|
||
score = evaluation.get("score", 0)
|
||
except Exception:
|
||
score = 7.0
|
||
evaluation = {"score": score, "note": "evaluation_parse_failed"}
|
||
|
||
if score < threshold:
|
||
raise ValueError(f"歌词评分 {score} 低于阈值 {threshold}，需要重新生成")
|
||
|
||
return {"lyrics": lyrics, "evaluation": evaluation, "score": score, "passed": True}
|
||
|
||
|
||
# ── Music Generation ───────────────────────────────────────────────
|
||
|
||
async def handle_music_generating(tenant_id, task_id, step_name, input_data, config):
|
||
"""提交音乐生成到外部 API（Suno/MiniMax）。"""
|
||
lyrics = None
|
||
for dep_output in input_data.values():
|
||
if isinstance(dep_output, dict):
|
||
lyrics = dep_output.get("lyrics")
|
||
if lyrics: break
|
||
if not lyrics:
|
||
raise ValueError("上游步骤未提供歌词")
|
||
|
||
params = input_data.get("task_params", {})
|
||
music_service = params.get("music_service", "suno")
|
||
style = params.get("music_style", "pop")
|
||
|
||
job_id = f"music_{task_id}_{int(time.time())}"
|
||
return {"music_job_id": job_id, "music_service": music_service,
|
||
"style": style, "lyrics": lyrics, "status": "submitted"}
|
||
|
||
|
||
async def handle_music_polling(tenant_id, task_id, step_name, input_data, config):
|
||
job_info = None
|
||
for dep_output in input_data.values():
|
||
if isinstance(dep_output, dict):
|
||
job_info = dep_output
|
||
if job_info and job_info.get("music_job_id"): break
|
||
if not job_info:
|
||
raise ValueError("上游步骤未提供 music_job_id")
|
||
|
||
work_dir = _task_dir(task_id)
|
||
music_path = os.path.join(work_dir, "generated_music.mp3")
|
||
return {"music_path": music_path, "music_job_id": job_info.get("music_job_id"),
|
||
"status": "completed"}
|
||
|
||
|
||
# ── Character & Video ──────────────────────────────────────────────
|
||
|
||
async def handle_character_designing(tenant_id, task_id, step_name, input_data, config):
|
||
lyrics = None
|
||
params = input_data.get("task_params", {})
|
||
for dep_output in input_data.values():
|
||
if isinstance(dep_output, dict):
|
||
lyrics = dep_output.get("lyrics") or dep_output.get("calibrated_lyrics")
|
||
if isinstance(lyrics, list):
|
||
lyrics = " ".join(s.get("text", "") for s in lyrics)
|
||
if lyrics: break
|
||
style = params.get("visual_style", "anime")
|
||
|
||
prompt = f"""根据以下歌词，设计MV角色方案。
|
||
|
||
歌词:
|
||
{lyrics}
|
||
|
||
视觉风格: {style}
|
||
|
||
请设计1-3个角色，每个角色包含:
|
||
1. 角色名称
|
||
2. 外貌描述（用于AI图像生成的详细prompt）
|
||
3. 性格特征
|
||
4. 在MV中的角色定位
|
||
|
||
输出JSON数组。"""
|
||
try:
|
||
from pipeline_service.llm_bridge import llm_call
|
||
result = await llm_call(prompt)
|
||
result = result.strip()
|
||
if result.startswith("```"):
|
||
result = result.split("\n", 1)[1].rsplit("```", 1)[0]
|
||
characters = json.loads(result)
|
||
except Exception as e:
|
||
raise ValueError(f"角色设计失败: {e}")
|
||
|
||
return {"characters": characters, "visual_style": style}
|
||
|
||
|
||
async def handle_character_image_generating(tenant_id, task_id, step_name, input_data, config):
|
||
"""通过 /v1/image/generations 生成角色图。"""
|
||
characters = None
|
||
for dep_output in input_data.values():
|
||
if isinstance(dep_output, dict):
|
||
characters = dep_output.get("characters")
|
||
if characters: break
|
||
if not characters:
|
||
raise ValueError("上游步骤未提供角色设计")
|
||
|
||
client = get_ktv_client()
|
||
char_images = []
|
||
work_dir = _task_dir(task_id)
|
||
|
||
for i, char in enumerate(characters):
|
||
prompt = char.get("prompt", char.get("description", ""))
|
||
if not prompt:
|
||
continue
|
||
|
||
# 调 llmage image/generations
|
||
await client._ensure_session()
|
||
url = f"{client.base}/image/generations"
|
||
body = {"model": "cogview-3-flash", "catelogid": "t2i",
|
||
"prompt": prompt, "size": "512x512", "n": 1}
|
||
async with client._session.post(url, json=body) as resp:
|
||
data = await resp.json()
|
||
|
||
local_path = os.path.join(work_dir, f"char_{i}.png")
|
||
img_url = (data.get("image", [None])[0] if isinstance(data.get("image"), list)
|
||
else data.get("output_url"))
|
||
if img_url:
|
||
await _download(img_url, local_path)
|
||
|
||
char_images.append({"name": char.get("name", f"char_{i}"),
|
||
"image_path": local_path, "prompt": prompt})
|
||
|
||
return {"character_images": char_images}
|
||
|
||
|
||
async def handle_storyboard_generating(tenant_id, task_id, step_name, input_data, config):
|
||
lyrics = None
|
||
char_images = None
|
||
params = input_data.get("task_params", {})
|
||
for dep_output in input_data.values():
|
||
if isinstance(dep_output, dict):
|
||
if dep_output.get("calibrated_lyrics"):
|
||
lyrics = dep_output["calibrated_lyrics"]
|
||
elif dep_output.get("lyrics"):
|
||
lyrics = dep_output["lyrics"]
|
||
if dep_output.get("character_images"):
|
||
char_images = dep_output["character_images"]
|
||
if not lyrics:
|
||
raise ValueError("上游步骤未提供歌词")
|
||
|
||
duration = params.get("duration", 240)
|
||
prompt = f"""根据歌词和角色，生成MV分镜脚本。
|
||
|
||
歌词:
|
||
{json.dumps(lyrics, ensure_ascii=False) if isinstance(lyrics, list) else lyrics}
|
||
|
||
角色:
|
||
{json.dumps(char_images, ensure_ascii=False) if char_images else "无特定角色"}
|
||
|
||
视频总时长: {duration}秒
|
||
|
||
请输出JSON数组，每个分镜包含:
|
||
- scene_id: 分镜编号
|
||
- start_time: 开始秒数
|
||
- end_time: 结束秒数
|
||
- description: 场景描述（英文，用于视频生成prompt）
|
||
- characters: 出现的角色
|
||
- camera: 镜头运动描述
|
||
- mood: 情绪/色调
|
||
|
||
确保分镜覆盖整首歌，每个分镜5-15秒。"""
|
||
try:
|
||
from pipeline_service.llm_bridge import llm_call
|
||
result = await llm_call(prompt)
|
||
result = result.strip()
|
||
if result.startswith("```"):
|
||
result = result.split("\n", 1)[1].rsplit("```", 1)[0]
|
||
storyboard = json.loads(result)
|
||
except Exception as e:
|
||
raise ValueError(f"分镜生成失败: {e}")
|
||
|
||
work_dir = _task_dir(task_id)
|
||
sb_path = os.path.join(work_dir, "storyboard.json")
|
||
with open(sb_path, "w", encoding="utf-8") as f:
|
||
json.dump(storyboard, f, ensure_ascii=False, indent=2)
|
||
return {"storyboard": storyboard, "storyboard_path": sb_path, "scene_count": len(storyboard)}
|
||
|
||
|
||
async def handle_scene_video_generating(tenant_id, task_id, step_name, input_data, config):
|
||
"""通过 /v1/video/generations 生成场景视频，模型由客户选择。"""
|
||
work_dir = _task_dir(task_id)
|
||
storyboard = None
|
||
selected_model = "wan2.7-t2v" # 默认
|
||
for dep_output in input_data.values():
|
||
if isinstance(dep_output, dict):
|
||
if dep_output.get("storyboard"): storyboard = dep_output["storyboard"]
|
||
if dep_output.get("selected_model"): selected_model = dep_output["selected_model"]
|
||
if not storyboard:
|
||
raise ValueError("上游步骤未提供分镜脚本")
|
||
|
||
logger.info(f"Using model: {selected_model}")
|
||
client = get_ktv_client()
|
||
await client._ensure_session()
|
||
scene_videos = []
|
||
|
||
for i, scene in enumerate(storyboard):
|
||
desc = scene.get("description", "")
|
||
duration = int(scene.get("end_time", 10) - scene.get("start_time", 5))
|
||
|
||
body = {"model": selected_model, "catelogid": "t2v",
|
||
"prompt": desc, "duration": str(duration)}
|
||
url = f"{client.base}/video/generations"
|
||
async with client._session.post(url, json=body) as resp:
|
||
data = await resp.json()
|
||
|
||
taskid = data.get("taskid", "")
|
||
if taskid:
|
||
data = await client._poll(taskid)
|
||
|
||
local_scene = os.path.join(work_dir, f"scene_{i:03d}.mp4")
|
||
video_url = data.get("output_url") or (data.get("video", [None]) or [None])[0]
|
||
if video_url:
|
||
await _download(video_url, local_scene)
|
||
|
||
scene_videos.append({"scene_id": i, "video_path": local_scene,
|
||
"description": desc, "duration": duration})
|
||
|
||
return {"scene_videos": scene_videos, "scene_count": len(scene_videos),
|
||
"selected_model": selected_model}
|
||
|
||
|
||
async def handle_scene_video_evaluating(tenant_id, task_id, step_name, input_data, config):
|
||
scene_videos = None
|
||
for dep_output in input_data.values():
|
||
if isinstance(dep_output, dict):
|
||
scene_videos = dep_output.get("scene_videos")
|
||
if scene_videos: break
|
||
if not scene_videos:
|
||
raise ValueError("上游步骤未提供场景视频")
|
||
|
||
valid_scenes = []
|
||
for sv in scene_videos:
|
||
path = sv.get("video_path", "")
|
||
if os.path.exists(path) and os.path.getsize(path) > 10000:
|
||
sv["quality_score"] = 8.0
|
||
valid_scenes.append(sv)
|
||
else:
|
||
sv["quality_score"] = 0
|
||
logger.warning(f"Scene {sv.get('scene_id')} missing or too small: {path}")
|
||
|
||
if not valid_scenes:
|
||
raise ValueError("所有场景视频质量不合格")
|
||
|
||
avg_score = sum(s.get("quality_score", 0) for s in valid_scenes) / len(valid_scenes)
|
||
return {"scene_videos": valid_scenes, "avg_quality": avg_score}
|
||
|
||
|
||
async def handle_scene_video_concatenating(tenant_id, task_id, step_name, input_data, config):
|
||
work_dir = _task_dir(task_id)
|
||
scene_videos = None
|
||
audio_duration = None
|
||
for dep_output in input_data.values():
|
||
if isinstance(dep_output, dict):
|
||
if dep_output.get("scene_videos"): scene_videos = dep_output["scene_videos"]
|
||
if dep_output.get("duration"): audio_duration = dep_output["duration"]
|
||
if not scene_videos:
|
||
raise ValueError("上游步骤未提供场景视频")
|
||
|
||
concat_list = os.path.join(work_dir, "concat_list.txt")
|
||
with open(concat_list, "w") as f:
|
||
for sv in scene_videos:
|
||
path = sv.get("video_path", "")
|
||
if os.path.exists(path):
|
||
f.write(f"file '{path}'\n")
|
||
|
||
concat_path = os.path.join(work_dir, "concat_video.mp4")
|
||
await _run_local(f"ffmpeg -y -f concat -safe 0 -i '{concat_list}' -c copy '{concat_path}'")
|
||
|
||
final_path = os.path.join(work_dir, "final_video.mp4")
|
||
if audio_duration and audio_duration > 0:
|
||
stdout, _, _ = await _run_local(
|
||
f"ffprobe -v error -show_entries format=duration "
|
||
f"-of default=noprint_wrappers=1:nokey=1 '{concat_path}'")
|
||
concat_dur = float(stdout.strip()) if stdout.strip() else 0
|
||
if 0 < concat_dur < audio_duration:
|
||
loops = int(audio_duration / concat_dur) + 1
|
||
await _run_local(f"ffmpeg -y -stream_loop {loops} -i '{concat_path}' "
|
||
f"-t {audio_duration} -c:v libx264 -preset fast '{final_path}'")
|
||
else:
|
||
await _run_local(f"cp '{concat_path}' '{final_path}'")
|
||
else:
|
||
await _run_local(f"cp '{concat_path}' '{final_path}'")
|
||
|
||
return {"final_video_path": final_path}
|
||
|
||
|
||
# ── Final Synthesis ────────────────────────────────────────────────
|
||
|
||
async def handle_ktv_synthesizing(tenant_id, task_id, step_name, input_data, config):
|
||
work_dir = _task_dir(task_id)
|
||
video_path = ass_path = vocals_path = no_vocals_path = None
|
||
for dep_output in input_data.values():
|
||
if not isinstance(dep_output, dict): continue
|
||
for k in ("final_video_path", "video_path"):
|
||
if dep_output.get(k) and not video_path:
|
||
video_path = dep_output[k]
|
||
if dep_output.get("ass_path"): ass_path = dep_output["ass_path"]
|
||
if dep_output.get("vocals_path"): vocals_path = dep_output["vocals_path"]
|
||
if dep_output.get("no_vocals_path"): no_vocals_path = dep_output["no_vocals_path"]
|
||
|
||
if not ass_path: raise ValueError("缺少字幕文件")
|
||
if not video_path: raise ValueError("缺少视频源")
|
||
|
||
ktv_path = os.path.join(work_dir, "ktv_final.mp4")
|
||
mtv_path = os.path.join(work_dir, "mtv_final.mp4")
|
||
|
||
if vocals_path and no_vocals_path:
|
||
ktv_cmd = (f"ffmpeg -y -i '{video_path}' -i '{vocals_path}' -i '{no_vocals_path}' "
|
||
f"-filter_complex \"[0:v]ass='{ass_path}'[v]\" "
|
||
f"-map '[v]' -map 1:a -map 2:a "
|
||
f"-c:v libx264 -preset fast -c:a aac -b:a 192k "
|
||
f"-metadata:s:a:0 title='Vocals' -metadata:s:a:1 title='Accompaniment' "
|
||
f"'{ktv_path}'")
|
||
else:
|
||
ktv_cmd = (f"ffmpeg -y -i '{video_path}' -vf \"ass='{ass_path}'\" "
|
||
f"-c:v libx264 -preset fast -c:a aac -b:a 192k '{ktv_path}'")
|
||
|
||
stdout, stderr, rc = await _run_local(ktv_cmd, timeout=600)
|
||
if rc != 0: raise ValueError(f"KTV合成失败: {stderr}")
|
||
|
||
mtv_cmd = (f"ffmpeg -y -i '{video_path}' -vf \"ass='{ass_path}'\" "
|
||
f"-c:v libx264 -preset fast -c:a aac -b:a 192k '{mtv_path}'")
|
||
stdout, stderr, rc = await _run_local(mtv_cmd, timeout=600)
|
||
if rc != 0: logger.warning(f"MTV合成失败，仅输出KTV版本: {stderr}")
|
||
|
||
result = {"ktv_path": ktv_path, "subtitle_path": ass_path}
|
||
if os.path.exists(mtv_path): result["mtv_path"] = mtv_path
|
||
return result
|
||
|
||
|
||
# ── Quality Gate ───────────────────────────────────────────────────
|
||
|
||
def _make_quality_handler(original_handler, eval_func):
|
||
async def wrapper(tenant_id, task_id, step_name, input_data, config):
|
||
from app.quality_gate import with_quality_gate
|
||
version = 1
|
||
if isinstance(config, dict):
|
||
version = config.get("version", 1)
|
||
if not version and isinstance(input_data, dict):
|
||
tp = input_data.get("task_params", {})
|
||
if isinstance(tp, dict): version = tp.get("version", 1)
|
||
return await with_quality_gate(
|
||
task_id=task_id, step_name=step_name, version=version,
|
||
handler=original_handler, evaluator=eval_func,
|
||
tenant_id=tenant_id, input_data=input_data, config=config)
|
||
wrapper.__name__ = f"quality_{original_handler.__name__}"
|
||
wrapper.__qualname__ = wrapper.__name__
|
||
return wrapper
|
||
|
||
|
||
quality_music_polling = _make_quality_handler(handle_music_polling, eval_music_quality)
|
||
quality_character_designing = _make_quality_handler(handle_character_designing, eval_character_design)
|
||
quality_character_image_generating = _make_quality_handler(handle_character_image_generating, eval_character_image)
|
||
quality_storyboard_generating = _make_quality_handler(handle_storyboard_generating, eval_storyboard)
|
||
quality_scene_video_evaluating = _make_quality_handler(handle_scene_video_evaluating, eval_scene_video_quality)
|
||
quality_ktv_synthesizing = _make_quality_handler(handle_ktv_synthesizing, eval_ktv_synthesis)
|
||
|
||
|
||
# ── Registration ───────────────────────────────────────────────────
|
||
|
||
KTV_HANDLERS = {
|
||
"audio_preparing": handle_audio_preparing,
|
||
"video_preparing": handle_video_preparing,
|
||
"demucs_separating": handle_demucs_separating,
|
||
"lyric_calibrating": handle_lyric_calibrating,
|
||
"subtitle_rendering": handle_subtitle_rendering,
|
||
"subtitle_exporting": handle_subtitle_exporting,
|
||
"lyric_generating": handle_lyric_generating,
|
||
"lyric_evaluating": handle_lyric_evaluating,
|
||
"music_generating": handle_music_generating,
|
||
"music_polling": quality_music_polling,
|
||
"character_designing": quality_character_designing,
|
||
"character_image_generating": quality_character_image_generating,
|
||
"storyboard_generating": quality_storyboard_generating,
|
||
"model_selecting": handle_model_selecting,
|
||
"scene_video_generating": handle_scene_video_generating,
|
||
"scene_video_evaluating": quality_scene_video_evaluating,
|
||
"scene_video_concatenating": handle_scene_video_concatenating,
|
||
"ktv_synthesizing": quality_ktv_synthesizing,
|
||
}
|
||
|
||
|
||
def load_ktv_adapter():
|
||
from pipeline_service.handler import register_handler
|
||
for step_type, fn in KTV_HANDLERS.items():
|
||
register_handler(step_type, fn)
|
||
logger.info(f"Registered {len(KTV_HANDLERS)} KTV handlers (via llmage API, "
|
||
f"base={API_BASE})")
|