From 1eacee8a22b10751616a033139855cedc3c9269c Mon Sep 17 00:00:00 2001
From: johnnyzhang-eng <johnnyzhang-eng@users.noreply.github.com>
Date: Tue, 18 Aug 2026 16:33:55 +0800
Subject: [PATCH 1/2] =?UTF-8?q?feat(postprocess):=20=E6=8C=89=E8=84=9A?=
 =?UTF-8?q?=E7=BA=BF=E4=B8=8E=E6=9C=AC=E4=BD=93=E4=B8=AD=E5=BF=83=E6=8C=91?=
 =?UTF-8?q?=E5=87=BA=E7=AB=99=E4=BD=8D=E5=81=8F=E7=A6=BB=E5=85=A8=E5=BA=8F?=
 =?UTF-8?q?=E5=88=97=E7=9A=84=E5=B8=A7?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Refs #395
---
 .../src/windup_ai_engine/postprocess/pack.py  | 103 +++++++++++++++++-
 backend/tests/test_pack_align.py              |  91 ++++++++++++++++
 2 files changed, 193 insertions(+), 1 deletion(-)

diff --git a/backend/packages/ai_engine/src/windup_ai_engine/postprocess/pack.py b/backend/packages/ai_engine/src/windup_ai_engine/postprocess/pack.py
index 2fc0321a..1818758c 100644
--- a/backend/packages/ai_engine/src/windup_ai_engine/postprocess/pack.py
+++ b/backend/packages/ai_engine/src/windup_ai_engine/postprocess/pack.py
@@ -13,7 +13,8 @@
 
 _logger = logging.getLogger(__name__)
 
-__all__ = ["CELL", "CORE_THICKNESS", "FILL_H", "FILL_W", "FOOT_LINE",
+__all__ = ["CELL", "CORE_THICKNESS", "DRIFT_CX_TOL", "DRIFT_FOOT_TOL",
+           "FILL_H", "FILL_W", "FOOT_LINE", "drifted_frames",
            "align_bottom_center", "core_span", "sprite_sheet", "save_gif"]
 
 # 交付画布的几何 —— 提成模块常量而不是只当默认参数,是因为**入口预检要按同一套几何
@@ -284,6 +285,106 @@ def align_bottom_center(
     return out
 
 
+# 逐帧锚点相对中位数的容差(比例,相对画布高/宽)。超出即认为该帧生成得偏。
+# 用比例而不是像素:交付画布尺寸随项目变,像素阈值在 256 与 512 上不是同一个严格度。
+#
+# 脚底那档按归档里 58 段真实序列扫出来:0.023 时报 3 段,其中两段目视干净 —— 同批
+# 3D 渲出的走路序列自然步态起伏最大到 1.5%,而那两段擦边在 2.5% 与 2.8%。0.028 到
+# 0.045 是同一个结论(只剩一段真离群),取这段平台的中点而不是它的下沿。
+DRIFT_FOOT_TOL = 0.035   # 脚底线
+DRIFT_CX_TOL = 0.047     # 本体横向中心
+
+
+def drifted_frames(
+    frames: list[Image.Image],
+    foot_tol: float = DRIFT_FOOT_TOL,
+    cx_tol: float = DRIFT_CX_TOL,
+) -> tuple[int, ...]:
+    """挑出脚底或横向中心明显偏离全序列中位数的帧下标。
+
+    **量的是对齐之前的原始帧**:``align_bottom_center`` 会把每帧摆正,对齐后再量恒为 0,
+    那时偏差已经被搬运掉、但生成本身歪没歪的信息也一起没了。这里要的正是后者 ——
+    一帧的角色站位与其余帧差得远,通常是那一帧生成坏了(姿态崩、主体缺失、多出人物)。
+
+    判据用**中位数**不用均值:坏帧本身会把均值拖过去,于是所有帧看起来都"没偏多少"。
+
+    横向中心取本体跨度的中心而不是包围盒中心 —— 见 :func:`_core_columns`。
+    """
+    import numpy as np
+
+    feet: list[tuple[int, float]] = []
+    cxs: list[tuple[int, float]] = []
+    for i, f in enumerate(frames):
+        a = np.asarray(f)[:, :, 3]
+        ys, xs = np.where(a > 128)
+        if not len(ys):
+            continue
+        feet.append((i, float(ys.max())))
+        core = _core_columns(a > 128)
+        if core is None:
+            continue
+        cxs.append((i, float(sum(core) / 2)))
+    if len(feet) < 3:
+        return ()                        # 观测太少,中位数不成立
+
+    h = frames[0].size[1]
+    w = frames[0].size[0]
+    bad = _outliers(feet, foot_tol * h)
+    bad |= _outliers(cxs, cx_tol * w)
+    return tuple(sorted(bad))
+
+
+def _core_columns(mask) -> tuple[int, int] | None:
+    """本体占据的首尾列。判据与 :func:`core_span` 的列方向一致(以最厚的列为基准)。
+
+    不用整体包围盒:延展物的幅度随动作变,而它对中心的拉扯不是角色站位变了。实测一段
+    斧战士序列,斧头从举过头顶甩到水平前伸,包围盒中心一帧跳 38px(画布宽 256),
+    本体中心几乎不动 —— 按包围盒判,20 帧里 11 帧被误报成站位漂移。
+    """
+    import numpy as np
+
+    cols = mask.sum(axis=0)
+    if not cols.any():
+        return None
+    keep = np.flatnonzero(cols >= float(cols.max()) * CORE_THICKNESS)
+    if not len(keep):
+        return None
+    return int(keep.min()), int(keep.max())
+
+
+def _outliers(obs: list[tuple[int, float]], tol: float) -> set[int]:
+    """按中位数判离群;整段有净位移时先把这条直线除掉。
+
+    两种误报要同时躲开,而它们要求相反的处理:
+
+    - **净位移**(角色一路向右走):不除趋势的话后半段整片被判离群。实测一段斧战士序列
+      连续 7 帧横偏稳定在 30px 上下、脚偏全 0 —— 那是走位。
+    - **周期摆动**(原地走路、四足步态):除趋势反而制造离群。实测一段狼序列脚底是
+      1342/1370/1352/1377 循环三遍、净位移为 0,而逐帧差分的中位数给出 25px/帧 的
+      假斜率,残差被推到 ±120px,整段 12 帧报出 8 帧。
+
+    判据因此看**首尾净变化**而不是逐帧差分:周期序列首尾回到同一点,净变化接近 0,
+    不触发除趋势;真的一路位移才触发。
+    """
+    import numpy as np
+
+    if len(obs) < 4:
+        med = float(np.median([v for _, v in obs])) if obs else 0.0
+        return {i for i, v in obs if abs(v - med) > tol}
+
+    x = np.array([i for i, _ in obs], dtype=float)
+    y = np.array([v for _, v in obs], dtype=float)
+    # 首尾各取三帧的中位数比,单帧首尾正好是坏帧时会把整条判据带偏。
+    head = float(np.median(y[:3]))
+    tail = float(np.median(y[-3:]))
+    # 无条件除趋势,不设"位移够大才除"的门槛:那个门槛会造一个悬崖 —— 实测一段打捞员
+    # 走路净位移 24.0px、容差 24.1px,差 0.1 就不除趋势,于是首两帧被误报。周期序列首尾
+    # 回到同一点、净位移≈0,减掉一条≈0 的直线是空操作,所以无条件除对它们无害。
+    y = y - np.linspace(0.0, tail - head, len(y))
+    med = float(np.median(y))
+    return {int(i) for i, v in zip(x, y) if abs(v - med) > tol}
+
+
 def sprite_sheet(frames: list[Image.Image], bg=(0, 0, 0, 0)) -> Image.Image:
     """横向拼接为 sprite sheet。"""
     if not frames:
diff --git a/backend/tests/test_pack_align.py b/backend/tests/test_pack_align.py
index 27bd460c..cc27c13d 100644
--- a/backend/tests/test_pack_align.py
+++ b/backend/tests/test_pack_align.py
@@ -429,3 +429,94 @@ def test_missing_widths_falls_back_to_old_behaviour():
     h = [60 + 40 * i / (n - 1) for i in range(n)]
     comp, _ = scale_drift(h)                          # 不传 widths
     assert any(c != 1.0 for c in comp)
+# ── 坏帧标记:站位明显偏离全序列的帧要被挑出来 ────────────────────────────────
+
+
+def _figure(cell=256, cx=128, foot=200, w=40, h=90):
+    """在指定位置画一个不透明矩形当角色。"""
+    im = Image.new("RGBA", (cell, cell), (0, 0, 0, 0))
+    px = im.load()
+    for x in range(max(0, cx - w // 2), min(cell, cx + w // 2)):
+        for y in range(max(0, foot - h), min(cell, foot)):
+            px[x, y] = (200, 180, 160, 255)
+    return im
+
+
+def test_clean_sequence_flags_nothing():
+    from windup_ai_engine.postprocess.pack import drifted_frames
+
+    frames = [_figure(foot=200 + (i % 2), cx=128) for i in range(12)]
+    assert drifted_frames(frames) == ()
+
+
+def test_foot_outlier_is_flagged():
+    """某一帧脚底明显高于其余 —— 角色悬空,通常是那帧姿态崩了。"""
+    from windup_ai_engine.postprocess.pack import drifted_frames
+
+    frames = [_figure(foot=200) for _ in range(12)]
+    frames[5] = _figure(foot=160)
+    assert 5 in drifted_frames(frames)
+
+
+def test_horizontal_outlier_is_flagged():
+    from windup_ai_engine.postprocess.pack import drifted_frames
+
+    frames = [_figure(cx=128) for _ in range(12)]
+    frames[3] = _figure(cx=180)
+    assert 3 in drifted_frames(frames)
+
+
+def test_median_not_mean_so_one_bad_frame_cannot_hide():
+    """判据用中位数:坏帧会把均值拖过去,于是所有帧都显得"没偏多少"。"""
+    from windup_ai_engine.postprocess.pack import drifted_frames
+
+    frames = [_figure(foot=200) for _ in range(12)]
+    frames[0] = _figure(foot=60)               # 极端离群
+    assert drifted_frames(frames) == (0,), "只有那一帧该被标记"
+
+
+def test_too_few_observations_flags_nothing():
+    """观测不足时中位数不成立,不硬判。"""
+    from windup_ai_engine.postprocess.pack import drifted_frames
+
+    assert drifted_frames([_figure(), _figure(foot=100)]) == ()
+
+
+def test_steady_walk_translation_is_not_flagged():
+    """角色连续横移不是坏帧:坏帧孤立地偏离邻居,位移是连续同向的。"""
+    from windup_ai_engine.postprocess.pack import drifted_frames
+
+    frames = [_figure(cx=100 + i * 4, foot=200) for i in range(16)]
+    assert drifted_frames(frames) == ()
+
+
+def test_outlier_on_top_of_translation_still_flagged():
+    """位移之上叠一帧真离群,仍要报出来。"""
+    from windup_ai_engine.postprocess.pack import drifted_frames
+
+    frames = [_figure(cx=100 + i * 4, foot=200) for i in range(16)]
+    frames[7] = _figure(cx=100 + 7 * 4, foot=140)
+    assert 7 in drifted_frames(frames)
+
+
+def _with_extension(cell=256, cx=128, foot=200, reach=90):
+    """本体不动,只从右手伸出一条细长延展物(斧柄 / 剑 / 披风)。"""
+    im = _figure(cell=cell, cx=cx, foot=foot)
+    px = im.load()
+    for x in range(cx, min(cell, cx + reach)):
+        for y in range(foot - 60, foot - 52):
+            px[x, y] = (120, 120, 120, 255)
+    return im
+
+
+def test_extension_swing_is_not_mistaken_for_a_shifted_pose():
+    """延展物甩出去不是站位变了。
+
+    横向中心按整体包围盒算时,一柄伸出的斧子能把中心带走一大截:实测一段斧战士序列
+    包围盒中心一帧跳 38px(画布宽 256),20 帧里 11 帧被判成漂移,而本体中心几乎不动。
+    """
+    from windup_ai_engine.postprocess.pack import drifted_frames
+
+    frames = [_figure(cx=128, foot=200) for _ in range(6)]
+    frames += [_with_extension(cx=128, foot=200) for _ in range(6)]
+    assert drifted_frames(frames) == ()

From c4d1e47a6775236f89e4518bd056cf68b8b6a0d0 Mon Sep 17 00:00:00 2001
From: johnnyzhang-eng <johnnyzhang-eng@users.noreply.github.com>
Date: Tue, 18 Aug 2026 16:57:53 +0800
Subject: [PATCH 2/2] =?UTF-8?q?fix(postprocess):=20=E6=96=9C=E7=8E=87?=
 =?UTF-8?q?=E6=8C=89=E4=B8=AD=E4=BD=8D=E6=95=B0=E5=B8=A7=E4=BD=8D=E7=AE=97?=
 =?UTF-8?q?=EF=BC=8C=E6=97=A0=E4=B8=BB=E4=BD=93=E5=B8=A7=E5=B9=B6=E5=85=A5?=
 =?UTF-8?q?=E7=BB=93=E6=9E=9C?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Refs #395
---
 .../src/windup_ai_engine/postprocess/pack.py  | 17 +++++++--
 backend/tests/test_pack_align.py              | 35 +++++++++++++++++++
 2 files changed, 49 insertions(+), 3 deletions(-)

diff --git a/backend/packages/ai_engine/src/windup_ai_engine/postprocess/pack.py b/backend/packages/ai_engine/src/windup_ai_engine/postprocess/pack.py
index 1818758c..e0cce3ae 100644
--- a/backend/packages/ai_engine/src/windup_ai_engine/postprocess/pack.py
+++ b/backend/packages/ai_engine/src/windup_ai_engine/postprocess/pack.py
@@ -314,10 +314,15 @@ def drifted_frames(
 
     feet: list[tuple[int, float]] = []
     cxs: list[tuple[int, float]] = []
+    empty: set[int] = set()
     for i, f in enumerate(frames):
         a = np.asarray(f)[:, :, 3]
         ys, xs = np.where(a > 128)
         if not len(ys):
+            # 整帧没有主体就是本函数要找的那种坏帧(抠图抠穿、生成漏了角色),不是
+            # "没测到"。它不进中位数、直接进结果,也不受下面那道观测数下限的约束 ——
+            # 一帧全透明这件事本身不需要参照就能判。
+            empty.add(i)
             continue
         feet.append((i, float(ys.max())))
         core = _core_columns(a > 128)
@@ -325,13 +330,13 @@ def drifted_frames(
             continue
         cxs.append((i, float(sum(core) / 2)))
     if len(feet) < 3:
-        return ()                        # 观测太少,中位数不成立
+        return tuple(sorted(empty))      # 观测太少,中位数不成立
 
     h = frames[0].size[1]
     w = frames[0].size[0]
     bad = _outliers(feet, foot_tol * h)
     bad |= _outliers(cxs, cx_tol * w)
-    return tuple(sorted(bad))
+    return tuple(sorted(bad | empty))
 
 
 def _core_columns(mask) -> tuple[int, int] | None:
@@ -377,10 +382,16 @@ def _outliers(obs: list[tuple[int, float]], tol: float) -> set[int]:
     # 首尾各取三帧的中位数比,单帧首尾正好是坏帧时会把整条判据带偏。
     head = float(np.median(y[:3]))
     tail = float(np.median(y[-3:]))
+    # 斜率的分母取这两个中位数**各自对应的帧位**,不取序列总长:三帧窗口的中位数落在窗口
+    # 中间那一帧上,拿总长当分母会把斜率算小。实测 16 帧每帧横移 13px 的线性序列,按总长
+    # 算出 11.27/帧,两端各剩 13px 残差、双双越过 12.03 的容差,匀速位移被误报成坏帧。
+    head_i = float(np.median(x[:3]))
+    tail_i = float(np.median(x[-3:]))
     # 无条件除趋势,不设"位移够大才除"的门槛:那个门槛会造一个悬崖 —— 实测一段打捞员
     # 走路净位移 24.0px、容差 24.1px,差 0.1 就不除趋势,于是首两帧被误报。周期序列首尾
     # 回到同一点、净位移≈0,减掉一条≈0 的直线是空操作,所以无条件除对它们无害。
-    y = y - np.linspace(0.0, tail - head, len(y))
+    if tail_i > head_i:
+        y = y - (tail - head) / (tail_i - head_i) * x
     med = float(np.median(y))
     return {int(i) for i, v in zip(x, y) if abs(v - med) > tol}
 
diff --git a/backend/tests/test_pack_align.py b/backend/tests/test_pack_align.py
index cc27c13d..4f98cb6a 100644
--- a/backend/tests/test_pack_align.py
+++ b/backend/tests/test_pack_align.py
@@ -520,3 +520,38 @@ def test_extension_swing_is_not_mistaken_for_a_shifted_pose():
     frames = [_figure(cx=128, foot=200) for _ in range(6)]
     frames += [_with_extension(cx=128, foot=200) for _ in range(6)]
     assert drifted_frames(frames) == ()
+
+
+def test_linear_translation_at_the_tolerance_edge_is_not_flagged():
+    """斜率的分母取中位数各自的帧位,不取序列总长。
+
+    256px 画布、16 帧、每帧横移 13px:按总长算斜率会得到 11.27/帧,两端各剩 13px 残差、
+    双双越过 12.03 的容差,于是匀速位移被误报成首末两帧坏。
+    """
+    from windup_ai_engine.postprocess.pack import drifted_frames
+
+    frames = [_figure(cx=30 + i * 13, foot=200) for i in range(16)]
+    assert drifted_frames(frames) == ()
+
+
+def test_a_frame_with_no_subject_at_all_is_flagged():
+    """全透明帧是抠图抠穿或生成漏了角色,正是本函数要找的那种坏帧。
+
+    它不进中位数(没有可量的锚点),所以必须单独并进结果 —— 静默跳过的话,一组正常帧里
+    插一张透明 PNG 会返回空元组,而 ``_lastmile`` 只拒空字节、合法的全透明 PNG 照样通过。
+    """
+    from windup_ai_engine.postprocess.pack import drifted_frames
+    from PIL import Image
+
+    frames = [_figure(foot=200) for _ in range(12)]
+    frames[4] = Image.new("RGBA", (256, 256), (0, 0, 0, 0))
+    assert 4 in drifted_frames(frames)
+
+
+def test_no_subject_is_reported_even_when_observations_are_too_few():
+    """观测不足时中位数不成立,但"这帧没有主体"不需要参照就能判。"""
+    from windup_ai_engine.postprocess.pack import drifted_frames
+    from PIL import Image
+
+    frames = [_figure(), Image.new("RGBA", (256, 256), (0, 0, 0, 0))]
+    assert drifted_frames(frames) == (1,)