diff --git a/.memory/worklog.json b/.memory/worklog.json
index 6857c51..f3c50d0 100644
--- a/.memory/worklog.json
+++ b/.memory/worklog.json
@@ -3160,6 +3160,19 @@
"type": "session-heartbeat",
"message": "Codex 会话活跃 · 最近命令:codex · 3 项未提交变更 · 最近提交:auto-save 2026-05-14 04:04 (~6)",
"files_changed": 3
+ },
+ {
+ "ts": "2026-05-14T04:10:26+08:00",
+ "type": "commit",
+ "message": "auto-save 2026-05-14 04:10 (~4)",
+ "hash": "0448d28",
+ "files_changed": 4
+ },
+ {
+ "ts": "2026-05-13T20:13:12Z",
+ "type": "session-heartbeat",
+ "message": "Claude 会话活跃 · 最近命令:claude · 3 项未提交变更 · 最近提交:auto-save 2026-05-14 04:10 (~4)",
+ "files_changed": 3
}
]
}
diff --git a/api/main.py b/api/main.py
index 1d81476..3173cfa 100644
--- a/api/main.py
+++ b/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):
diff --git a/docs/source-analysis.html b/docs/source-analysis.html
index 9e47a2e..c9e0469 100644
--- a/docs/source-analysis.html
+++ b/docs/source-analysis.html
@@ -723,7 +723,7 @@ api/main.py
| 输入 Input |
- 创建/上传任务,显示视频就绪;视频缩略图上方提供自动抽帧快捷工具条,可快速选目标、张数并多次追加;单击视频缩略图打开画布内抽帧面板。 |
+ 创建/上传任务,显示视频就绪;每个视频缩略图上方都有绑定自己的自动抽帧快捷工具条,可快速选目标、张数并多次追加;单击视频缩略图打开画布内抽帧面板。 |
不要自动一路跑到 ASR 或生图;用户需要控制解析节奏。 |
page.tsx、InputNode、VideoFramePanelNode、api/main.py |
@@ -817,6 +817,18 @@ api/main.py
变更记录
这个记录不是 git log 的替代品。它记录“产品理解发生了什么变化、影响了哪些源码、你以后描述需求时该怎么说”。后续每次改功能都要补一条。
+
+
+ 2026-05-14 · 每个输入视频缩略图绑定自己的抽帧工具条
+ Input
+ UX
+
+
+
问题:统一放在缩略图浮条上方的抽帧工具条仍然不够明确,用户无法一眼判断当前会抽哪一个视频。
+
改动:抽帧目标、张数和抽帧按钮改为渲染在每个视频缩略图正上方,并且每个 job 独立保存目标和张数设置。点击某个缩略图上方的抽帧按钮时,前端直接把该 jobId 传给 analyzeJob,同时切换 active job 并进入该 job 的进度轮询。
+
影响:web/app/page.tsx、web/components/nodes/index.tsx、docs/source-analysis.html。后续与输入视频相关的快捷操作都应优先贴近对应缩略图,不再依赖全局当前选择的心智。
+
+