feat: VRSub 单体应用（WOV 单机版）初始提交

为视频生成 VR 双眼字幕的单体实现：FastAPI 后端、调度器与全部节点（提音/转写/翻译/ASS/抽帧/OCR/LLM 过滤）在单进程内运行。 - 节点协议（wov_sdk 数据模型）与分布式版保持一致，预留回退桥梁 - 工作流即数据：DAG 存于 workflows/*.json，模型/链路改动只改数据 - 调度器：拓扑顺序执行、断点续跑（产物重建）、任务暂停/继续 - 抽帧按帧间隔（select 按帧号精确取帧），VLM OCR 与 LLM 过滤使用自适应线程池弹性并发，并打印数据处理速度进度日志 - 100% 行覆盖率（pytest --cov-fail-under=100）
2026-08-16 23:58:25 +08:00
commit 4746e0363f
75 changed files with 10969 additions and 0 deletions
@@ -0,0 +1,178 @@
+"""视频抽帧节点。
+
+按**帧间隔**从视频抽取帧：解析视频帧率后，帧间隔 step = round(间隔秒 × fps)，
+用 ffmpeg 的 select 过滤器按帧号（n mod step == 0）精确取帧——每帧都是真实
+视频帧，帧时间 = 帧号 / fps，避免按时间 seek（-ss 秒）造成的取整漂移。
+再按 crop 相对比例 [x,y,w,h]（0~1）裁切出字幕区域，供 vlm-ocr 等下游节点
+使用。产物为 frames.json 清单：[{"time": 帧时间秒, "image_uri": 帧图片路径}, ...]。
+"""
+
+from __future__ import annotations
+
+import json
+import re
+import subprocess
+from pathlib import Path
+
+from nodes.ffmpeg import _ffmpeg_bin
+from wov_app.logging import get_logger
+from wov_sdk.models import InvokeRequest, InvokeResponse
+
+logger = get_logger("frame-extract")
+
+# 默认裁切：画面底部 18% 区域（常见字幕位置）。
+DEFAULT_CROP = [0.0, 0.82, 1.0, 0.18]
+
+
+def _parse_crop(raw) -> list[float] | None:
+    """解析并校验 crop 相对比例 [x,y,w,h]（0~1 且不越出画面）。"""
+    try:
+        crop = [float(value) for value in raw]
+    except (TypeError, ValueError):
+        return None
+    if len(crop) != 4:
+        return None
+    x, y, w, h = crop
+    if not (0 <= x <= 1 and 0 <= y <= 1 and 0 <= w <= 1 and 0 <= h <= 1):
+        return None
+    if x + w > 1.001 or y + h > 1.001:
+        return None
+    return crop
+
+
+def _video_size(video: Path, ffmpeg_bin: str) -> tuple[int, int] | None:
+    """从 ffmpeg -i 输出解析视频分辨率，避免依赖 ffprobe。"""
+    result = subprocess.run(
+        [ffmpeg_bin, "-i", str(video)],
+        capture_output=True,
+        text=True,
+    )
+    for line in (result.stderr or "").splitlines():
+        if "Video:" not in line:
+            continue
+        match = re.search(r"(\d{2,5})x(\d{2,5})", line)
+        if match:
+            return int(match.group(1)), int(match.group(2))
+    return None
+
+
+def _video_duration(video: Path, ffmpeg_bin: str) -> float | None:
+    """从 ffmpeg -i 输出解析总时长（秒）。"""
+    result = subprocess.run(
+        [ffmpeg_bin, "-i", str(video)],
+        capture_output=True,
+        text=True,
+    )
+    match = re.search(r"Duration:\s*(\d+):(\d+):(\d+(?:\.\d+)?)", result.stderr or "")
+    if match:
+        hours = int(match.group(1))
+        minutes = int(match.group(2))
+        seconds = float(match.group(3))
+        return hours * 3600 + minutes * 60 + seconds
+    return None
+
+
+def _video_fps(video: Path, ffmpeg_bin: str) -> float | None:
+    """从 ffmpeg -i 输出解析帧率，支持小数（29.97 fps）与有理数（30000/1001 fps）。"""
+    result = subprocess.run(
+        [ffmpeg_bin, "-i", str(video)],
+        capture_output=True,
+        text=True,
+    )
+    for line in (result.stderr or "").splitlines():
+        if "Video:" not in line:
+            continue
+        match = re.search(r"(\d+)/(\d+)\s*fps", line)
+        if match:
+            return int(match.group(1)) / int(match.group(2))
+        match = re.search(r"(\d+(?:\.\d+)?)\s*fps", line)
+        if match:
+            return float(match.group(1))
+    return None
+
+
+def _frame_step(fps: float, interval: float) -> int:
+    """帧间隔换算：每 step 帧取一帧（step=round(间隔秒×fps)，至少为 1）。"""
+    return max(1, int(round(interval * fps)))
+
+
+def invoke(request: InvokeRequest) -> InvokeResponse:
+    """按帧间隔抽取并裁切视频帧，输出 frames.json 清单。"""
+    video_uri = request.inputs.get("video_uri")
+    if not video_uri:
+        return InvokeResponse(status="failed", error="video_uri is required")
+    video = Path(video_uri)
+    if not video.is_file():
+        return InvokeResponse(status="failed", error="video file not found")
+
+    interval = float(request.params.get("interval_seconds", 0.5))
+    crop = _parse_crop(request.params.get("crop", DEFAULT_CROP))
+    if crop is None:
+        return InvokeResponse(status="failed", error="invalid crop")
+    if interval <= 0:
+        return InvokeResponse(status="failed", error="invalid interval_seconds")
+
+    ffmpeg_bin = _ffmpeg_bin()
+    size = _video_size(video, ffmpeg_bin)
+    if size is None:
+        return InvokeResponse(status="failed", error="cannot read video size")
+    duration = _video_duration(video, ffmpeg_bin)
+    if duration is None or duration <= 0:
+        return InvokeResponse(status="failed", error="cannot read video duration")
+    fps = _video_fps(video, ffmpeg_bin)
+    if fps is None or fps <= 0:
+        return InvokeResponse(status="failed", error="cannot read video fps")
+
+    # 把"每多少秒一帧"换算为"每多少帧取一帧"，按帧号取帧是帧精确的。
+    step = _frame_step(fps, interval)
+    width, height = size
+    x_px = int(round(crop[0] * width))
+    y_px = int(round(crop[1] * height))
+    w_px = max(1, int(round(crop[2] * width)))
+    h_px = max(1, int(round(crop[3] * height)))
+
+    output_dir = Path(request.output_dir)
+    frames_dir = output_dir / "frames"
+    frames_dir.mkdir(parents=True, exist_ok=True)
+    # 单次解码全片：select 按帧号（n mod step == 0）精确取帧，随后对选中帧
+    # 裁切字幕区域并压缩到 720p 内（仅缩小）——过大的输入图会触发 glm-ocr
+    # 的重复循环 bug（M-RoPE delta），源头规避。
+    result = subprocess.run(
+        [
+            ffmpeg_bin,
+            "-y",
+            "-hide_banner",
+            "-loglevel",
+            "error",
+            "-i",
+            str(video),
+            "-vf",
+            (
+                f"select='not(mod(n\\,{step}))',"
+                f"crop={w_px}:{h_px}:{x_px}:{y_px},"
+                "scale=1280:720:force_original_aspect_ratio=decrease:force_divisible_by=2"
+            ),
+            # 只写出被选中的帧，避免 CFR 补帧产生重复文件。
+            "-vsync",
+            "vfr",
+            str(frames_dir / "frame_%04d.png"),
+        ],
+        capture_output=True,
+        text=True,
+    )
+    if result.returncode != 0:
+        return InvokeResponse(status="failed", error=result.stderr[-500:] or "ffmpeg failed")
+
+    # 第 k 个输出文件对应原始帧号 k×step，时间 = 帧号 / fps（帧精确，无累计偏差）。
+    files = sorted(frames_dir.glob("frame_*.png"))
+    manifest = [
+        {"time": round((index * step) / fps, 3), "image_uri": str(path)}
+        for index, path in enumerate(files)
+    ]
+    manifest_path = output_dir / "frames.json"
+    manifest_path.write_text(json.dumps(manifest, ensure_ascii=False), encoding="utf-8")
+    logger.info("抽帧完成: %d 帧, 帧间隔 %d, 裁切 %dx%d+%d+%d", len(files), step, w_px, h_px, x_px, y_px)
+    return InvokeResponse(
+        status="completed",
+        outputs={"frames_manifest": str(manifest_path), "frame_count": len(files)},
+    )