auto-save 2026-05-14 04:15 (~4)

api/main.py

@@ -90,6 +90,7 @@ JobStatus = Literal[
 KEYFRAME_COUNT = int(os.getenv("KEYFRAME_COUNT", "5"))
 FrameExtractTarget = Literal["balanced", "subject", "transition", "expression", "motion"]
 FrameExtractMode = Literal["replace", "append"]
+FrameExtractQuality = Literal["fast", "accurate", "ultra"]
 FRAME_TARGET_LABELS: dict[FrameExtractTarget, str] = {
     "balanced": "综合关键帧",
     "subject": "清晰主体",
@@ -97,6 +98,11 @@ FRAME_TARGET_LABELS: dict[FrameExtractTarget, str] = {
     "expression": "表情瞬间",
     "motion": "动作峰值",
 }
+FRAME_QUALITY_LABELS: dict[FrameExtractQuality, str] = {
+    "fast": "快速",
+    "accurate": "精细",
+    "ultra": "极准",
+}
 
 
 class GeneratedImage(BaseModel):
@@ -399,20 +405,23 @@ def _sharpness_from_gray(g: np.ndarray) -> float:
     return float(lap.var())
 
 
-def _frame_metrics(img_path: Path, idx: int, timestamp: float) -> dict | None:
+def _frame_metrics(img_path: Path, idx: int, timestamp: float, metric_width: int = 160) -> dict | None:
     """低清候选帧的本地评分特征。只用于排序，最终仍从原视频抽原尺寸帧。"""
     try:
         with Image.open(img_path) as raw:
             img = raw.convert("RGB")
             h = imagehash.phash(img)
-            small = img.resize((160, 90))
+            src_w, src_h = img.size
+            metric_height = max(1, round(metric_width * src_h / max(src_w, 1)))
+            small = img.resize((metric_width, metric_height))
     except Exception:
         return None
 
     arr = np.asarray(small, dtype=np.float32)
     # Rec. 601 luma，保留 0-255 范围，便于和清晰度 / 对比度阈值一起看。
     gray = (0.299 * arr[:, :, 0] + 0.587 * arr[:, :, 1] + 0.114 * arr[:, :, 2]).astype(np.float32)
-    center = gray[22:68, 40:120]
+    gh, gw = gray.shape
+    center = gray[gh // 4:max(gh // 4 + 1, gh * 3 // 4), gw // 4:max(gw // 4 + 1, gw * 3 // 4)]
     rg = arr[:, :, 0] - arr[:, :, 1]
     yb = 0.5 * (arr[:, :, 0] + arr[:, :, 1]) - arr[:, :, 2]
     colorfulness = float(np.sqrt(rg.var() + yb.var()) + 0.3 * np.sqrt(rg.mean() ** 2 + yb.mean() ** 2))
@@ -432,6 +441,20 @@ def _frame_metrics(img_path: Path, idx: int, timestamp: float) -> dict | None:
     }
 
 
+def _scan_profile(duration: float, quality: FrameExtractQuality) -> tuple[float, int, int, int]:
+    """返回 scan_fps / scan_width / metric_width / estimated_count。"""
+    if quality == "ultra":
+        base_fps, scan_width, cap, metric_width = 12.0, 960, 1800, 320
+    elif quality == "accurate":
+        base_fps, scan_width, cap, metric_width = 8.0, 720, 900, 240
+    else:
+        base_fps, scan_width, cap, metric_width = 2.0, 360, 240, 160
+
+    estimated = max(1, min(int(duration * base_fps), cap))
+    scan_fps = max(0.02, min(base_fps, estimated / max(duration, 0.1)))
+    return scan_fps, scan_width, metric_width, estimated
+
+
 def _attach_temporal_metrics(items: list[dict]) -> None:
     """相邻低清帧差异：转场 / 动作目标依赖它，不需要逐帧高分辨率扫描。"""
     for i, it in enumerate(items):