V6：从生图到切片，展开整个过程
目标是把 100 帧、27 fps 的角色视频重建成可编辑二维动画。本页展开 V6 的素材、拆分、补片、网格和动作编排。坐姿结构更接近原片，但完整部件被可见面切片替代，拆分质量仍有问题；这里直接展示产生碎片的步骤和实际中间画面。
4新生图：整人参考＋3 张部件图
43最终图层，含遮挡补片和背景
9LLM 观察后编排的次级动作组
0397 个采样中的网格翻折
整个 pipeline：素材从哪里来，又被怎样处理
这版有两个不同的拆分分支。先尝试独立部件 atlas（一张图中排列多个部件），随后因装配问题改为从生成整人图切可见面。下面逐步展开输入、输出、归属规则和实际使用情况。
1 · 原片观察LLM 判断坐姿、服装与动作
2 · 生图整人参考 ＋ 3 张 atlas
3A · 独立部件提取16 个部件 → 初装后多数弃用
3B · 整人图切片多边形、颜色、归属 → 18 片可见面
4 · 遮挡补片14 个颜色延展 ＋ 1 个生成躯干
5 · 各部件网格与绑定Alpha、Canny、约束三角化、权重
6 · 光流 → LLM 动作观测证据 → 9 组目标 → 24 个笔刷
7 · 装配与逐版检查9 份中间渲染 → 当前 V6
当前拆分的问题：透明背景只给出了外轮廓。整人图里的袖子、衣领和躯干仍互相遮挡；切出的像素只包含可见面。torso_core 又接收了其他多边形未覆盖的残余像素，出现碎片。首帧能拼回、三角形不翻折，都不能证明这些部件适合独立动画。
图示依据当前保存的代码和产物。颜色候选、闭合和 Canny 是本次按代码重建的诊断视图；最终贴图、网格、权重及历史渲染均来自已保存结果。完整流程数据与来源哈希
1–2 · 观察原帧，保存提示词和全部生图
4 次生图均保留了原生 Alpha。atlas 生图同时参考原帧和生成整人图；这些图不是都进入了最终模型。逐张查看确切提示词 · 生图来源记录
3A · atlas 提取：看原图中的位置，以及最后有没有用
代码按 Alpha > 24 找连通区域，保留面积 ≥ 500 像素且中心落在指定格子的区域；用 5×5 核膨胀选区，调整 Alpha 上限，再紧裁边界。这一步只是在新生成图上提取部件。
全部 16 个独立生成部件（点击检查）
3B · 当前切片：多边形如何把整人图切成部件
身体按多边形先后顺序分配像素，后写的部件覆盖前写的部件。头发还加入橙色条件、3×3 闭合、填洞和小块归属调整。这里可以逐步查看条件，而不只看最终裁图。
连通区域统计：Alpha > 16，8 邻域，只计面积 ≥ 8 像素的区域。碎片数量用于定位问题，不能自动判断部件质量。透明区域显示棋盘底色；点击下面的缩略图可检查每一片。
4 · 隐藏补片：哪些内容是生成的，哪些只是颜色延展
14 个身体切片各有一块内部补片，取最近的已生成像素向切口外延展。普通部件半径 10 px，躯干和两条腿 38 px；范围仍受原生成身体的 Alpha 限制。另有 1 块独立生图的躯干补片。颜色延展没有生成被遮挡处的完整结构。
查看独立生成的躯干补片，以及被弃用的头发补片
5 · 从 Alpha 和线条到网格、蒙皮与笔刷权重
网格由透明度边界和 Canny 线条约束三角化得到。身体切片共用同一位置的蒙皮权重以减少裂缝；右侧宽后发的笔刷权重沿长度分布。网格忠实保留了输入里的孔和碎片，它不会自动把这些切片补成完整部件。
装配坐标、骨骼输入和 43 层绘制顺序
整人图先等比缩至高 720 px。切片的贴图坐标变为画布坐标时，采用 x′ = 0.936x + 76.25、y′ = y − 4 + max(0, y − 620) × 0.16，对应当时选定的颈部与膝部对齐。早期 atlas 的放置则是各部件单独的矩形。
主动作事件与骨盆、胸口、头部轨迹沿用 V5；手臂重新标定绑定位置并计算 IK。身体表面共用蒙皮权重，远侧手臂当前没有独立运动的有效蒙皮影响。下列顺序从后向前绘制。
6 · 原片光流 → LLM 判断 → 控制点 → 代码执行
DIS 光流及往返检查提供运动证据。LLM 查看原片、轨迹和遮挡后，编排 9 组控制目标；代码把目标编译为 24 个笔刷并执行变形。光流没有直接写入网格。查看测量箭头、可信性标记和观测轨迹
点击时间行可同步跳到该帧。
保存的原始观察说明
7 · 中间装配与修改过程：逐版、逐帧对照
下面使用实际保存的中间渲染。部分早期场景 JSON 已被后续修改覆盖，因此这里只展示当时留下的截图，并标明对应哈希；不会把当前代码重跑的画面冒充历史结果。
全部 9 个阶段概览
原片 / V5 / V6 同步播放
改进集中在首帧结构和可见的衣发摆动。转头仍主要由平面旋转和表情切换表达，不能等同于原片的立体转向。几何通过不代表动作已经完全相似。
首帧与确切生图提示词
V3 的后续生图提示词要求长裙完全覆盖腿部，改变了原片服装。V6 提示词明确保留原片的坐姿透视、服装开口和可见腿部，禁止添加长裙、长袜或站姿轮廓。后发要求清晰的大束结构，避免额外细卷发。
独立 atlas 的第一次装配仍出现重复衣领和袖口错位。因此最终可见面从本轮新生成的整人参考图按 LLM 划定的部件边界拆出，并保留独立网格；独立生图补躯干遮挡区域，少量生成像素延展补内部接缝。没有分割或使用原视频像素作为动画纹理。
查看本轮全部原始生图与确切提示词
canonical 提示词
Redraw the same character and exact first-frame pose shown in the attached animation frame as a clean reusable painted animation reference, on a true transparent background. Preserve the actual costume, seated three-quarter perspective, proportions, overlaps and asymmetry. This is a faithful production redraw, not a costume redesign. Keep the warm restrained anime linework and shading.
She sits on an implied rock, facing slightly toward image-right, eyes closed, viewer-left gloved hand resting at the upper chest. Her other arm lowers beside the torso and is mostly behind the nearer bent thigh. The thighs extend to image-right and the nearer knee bends downward. Keep the visible skin of the seated legs exactly as in the reference and the original openings of the green inner garment. Do not invent a long green skirt, stockings, extra dress panels, or a standing silhouette. The white coat spreads over the seat toward image-left and falls around the seated legs, with a narrow striped green front panel hanging between them. Keep the original coat length and its asymmetric gold-edged openings. Keep the upper chest, thin dark straps, gray collar, badge, cuffs, watch and waist ornaments as shown, without replacing the clothing design.
Hair must have the same few clearly shaped copper-orange locks, their length and swept direction from the frame, not a voluminous mass of extra thin curly strands. Preserve the hood and ears. Match shoulder-to-waist distance, elbow bend, wrist position, hip-to-knee direction and the head-to-body size. Draw the complete visible character on a canvas with a small transparent margin. Exclude scenery, seat, ground shadow and tail (the tail is already a separate approved asset). No text, grid, labels or checkerboard. Actual alpha transparency.
costume_parts 提示词
Create SIX separate production animation clothing layers from the attached approved character redraw (image 1) and original pose (image 2). Preserve this exact seated pose, costume openings, perspective, fabric stripes, gold edging and restrained painted style. Do not design a different costume. The character's seated legs remain visible in the final assembly; do not lengthen a skirt or close the coat opening.
LANDSCAPE 3-column by 2-row atlas. Each cell contains exactly one independent garment layer, fully isolated by generous transparent gutters. Parts may have different sizes; retain their visible shapes from the reference. No grid, words, numbers, cast shadow or background. True alpha.
TOP LEFT: complete central TORSO CORE, from the short gray neck collar to the hip. Include the collar, the small upper chest area and thin dark horizontal straps exactly as pictured, and the narrow vertically striped green bodice with the gold upper trim. Preserve the long slender seated torso shape and its slight perspective. No white coat lapels, shoulders, sleeves, badge, arms, hands, head or skirt flap. Extend under lapels and beneath waist by a small painted overlap.
TOP CENTER: viewer-left WHITE UPPER COAT / LAPEL layer, from shoulder and neck edge down to waist, including its original long folded white lapel and triangular pale gold inset. No sleeve, arm, green central bodice, badge or lower coat tail. Draw the hidden underlap at the waist without an outline.
TOP RIGHT: viewer-right WHITE UPPER COAT / LAPEL layer, from neck/shoulder down to waist, narrower in seated perspective, its lapel turned outward near the chest. No arm, sleeve, green bodice or lower hem. Hidden continuation at waist, no black cut seam.
BOTTOM LEFT: viewer-left LOWER WHITE COAT FRONT PANEL. Starts at waist, sweeps diagonally down-left as it drapes over the seat, then continues off the bottom. Preserve the original stepped gold inner edge and green trim. Broad leftward trailing fabric, slender waist root. It must not become a standing flared skirt or a rolled cylindrical hem. No torso, legs, green central hanging panel or other coat pieces. Paint hidden overlap above the waist.
BOTTOM CENTER: viewer-right LOWER WHITE COAT PANEL, draped over and behind the raised seated thigh toward image-right. Long tapered diagonal strip from hip at upper-left to lower-right, white cloth with gold/green edge. Preserve the seated diagonal direction, no vertical hanging bell shape. No legs or waist bow.
BOTTOM RIGHT: BACK WHITE COAT TAIL, the small spread of white fabric on the seat extending left behind the hips, with green lining visible at the edge. This is a low shallow sideways drape, not a full circular skirt. No duplicate front coat panel, torso or legs.
Reconstruct hidden attachment overlaps in the same material. Every piece isolated and complete within its cell, no pieces touching. All visible artwork is newly drawn, not original pixel cutouts.
seated_parts 提示词
Create SIX precisely matching separated animation assets from the attached character redraw (image 1) and original animation pose (image 2). Keep the actual seated posture, exact garment design, warm anime shading, and ordinary visible legs. These are production layers for the same character, not new clothing designs. No stockings, long skirt covering the knees, extra dress layers or invented glove on the lowered hand.
LANDSCAPE atlas of 3 columns and 2 rows. One isolated asset in each cell, wide transparent gutters. Every shape must be complete with a small hidden overlap at attachment. Retain the pose directions from the images. Actual transparent alpha background, no scenery, grid, labels, text or shadows.
TOP LEFT: the narrow GREEN STRIPED FRONT CLOTH PANEL hanging from the lower torso between the seated legs. Its root is upper-left and it curves slightly to lower-right toward the bottom. Preserve its taper and long vertical stripes, gold edge. Green fabric only, no torso, white coat, hips or legs. Absolutely not a wide skirt over both knees.
TOP CENTER: the complete visible nearer BENT LEG as one continuous layer, from its upper thigh under the coat to the bent knee at viewer-right and the upper part of the shin descending downward. Match the exact L-shaped sitting silhouette, soft knee shading and skin tone in the reference. No stockings, shoes, skirt, underwear or pelvis. Enough overlap at the upper-thigh attachment; it will be behind the coat. Preserve knee bend rather than a straight tube.
TOP RIGHT: the FAR THIGH visible at the lower center of the reference, diagonal from upper-left toward lower-right, behind the hanging green cloth. Match its skin tone, restrained shading and seated perspective. No pelvis, clothing, shin or other body parts. Complete hidden attachment end.
BOTTOM LEFT: viewer-left ivory UPPER SLEEVE with its softly rolled cuff, from shoulder at upper-right down toward elbow at lower-left. Same compact seated upper arm shape and fabric folds as reference. Cloth only, no skin, forearm, badges or torso. One sleeve opening at the elbow, no duplicate cuffs. Paint a broad hidden shoulder overlap.
BOTTOM CENTER: viewer-right ivory UPPER SLEEVE and cuff, narrower, from shoulder upper-left to elbow lower-right, same rolled fabric and green/tan diamond cuff band as reference. No forearm, bow, hand or torso. One cuff only; complete hidden shoulder overlap.
BOTTOM RIGHT: viewer-right lowered FOREARM with relaxed hand, mostly straight and directed down-right from elbow. Match the reference skin tone. The hand is ungloved and relaxed. This layer will be mostly hidden behind the nearer seated thigh. No sleeve, watch or coat; complete hidden elbow attachment.
hair_parts 提示词
Produce FOUR separate copper-orange hair animation layers that precisely reproduce the large clear locks in the attached approved redraw (image 1) and original video frame (image 2). Match their shapes, length, swept direction and restrained cel-painted shading. Do not add extra thin curly strands or a voluminous generic mane. These layers supplement the existing face, hood, crown and front bangs; do not redraw those elements.
Square 2-column by 2-row atlas. Exactly one hair layer per cell, generous fully transparent margins, no touching. Roots need generous hidden painted overlap, ends naturally tapered and complete. No labels, grid, symbols, face, scalp skin, accessories, character or background. Actual alpha transparency.
TOP LEFT: BACK LEFT LOCK underneath the hood on viewer-left, long narrow gently curved orange hair extending downward behind the left shoulder, with two clear pointed tips that curve slightly left. Root at upper-right, tapered tips lower-left. Preserve the long slim silhouette from the reference.
TOP RIGHT: BACK RIGHT LOWER LOCK: the broad copper hair blade behind the right shoulder, root at upper-left, bends and fans toward lower-right, ending in two broad sharply tapered tips. Two broad flowing shapes, not a spray of many fine strands. Match the original frame's large triangular swept lower-right hair silhouette.
BOTTOM LEFT: BACK RIGHT UPPER LOCK above the broad lower lock: root at upper-left, sweeps right with a gentle upward flick at the tip. One main long curved lock and one small secondary tip. Thin brown linework; no face or earrings.
BOTTOM RIGHT: RIGHT TEMPLE / CHEEK SIDE LOCK, long narrow orange strand descending alongside the right cheek and neck, rooted at top, gently bending to the right toward its pointed lower end. Preserve the original hair length and slim shape. No face, skin, collar or earring.
光流负责测量，动作由观察后的控制点决定
代码计算 99 对相邻帧的稠密光流，并做前后往返检查。箭头显示图像局部如何移动；它不能自动说明哪一束头发挡住了哪一片衣服。LLM 同时查看原片、箭头和轨迹，选择事件、判断遮挡，再写控制点位移。
例如右后发在 14 帧和 74–83 帧向内、向下收拢，39–52 帧向外展开；绿色前襟的中段先转向，末端随后跟随。左后衣摆有轨迹跟到尾巴上的情况，已剔除。脸侧发束与宽后发拆开控制，宽发束沿长度分散弯曲。
查看 LLM 编排的关键姿势及其观察依据
{
"author": "LLM visual review of original frames and flow evidence",
"times": [
0,
14,
26,
39,
52,
63,
74,
83,
88,
99
],
"controls": [
{
"id": "hair_left",
"layers": [
"hair_back_left"
],
"anchors": [
[
359,
313
],
[
330,
343
]
],
"fields": [
{
"center": [
359,
313
],
"radii": [
56,
52
],
"hardness": 0.3,
"pins": [
{
"center": [
385,
263
],
"radii": [
60,
38
],
"hardness": 0.6
}
]
},
{
"center": [
330,
343
],
"radii": [
60,
60
],
"hardness": 0.3,
"pins": [
{
"center": [
385,
263
],
"radii": [
60,
38
],
"hardness": 0.6
}
]
}
],
"offsets": [
[
[
0,
0
],
[
-14,
-4
],
[
-9,
-5
],
[
-4,
-2
],
[
0,
2
],
[
-4,
0
],
[
-12,
-3
],
[
-10,
-5
],
[
-8,
-5
],
[
-3,
-1
]
],
[
[
0,
0
],
[
-30,
-18
],
[
-25,
-20
],
[
-7,
-9
],
[
2,
2
],
[
-9,
-2
],
[
-30,
-12
],
[
-28,
-20
],
[
-22,
-18
],
[
-8,
-4
]
]
],
"reason": "Tip track retains consistent left/up sweep twice. Mid track first fails cumulative consistency at frame 38; conservatively stop using it after the preceding selected event at26. Later middle offsets manually interpolated along the strand, not copied from drift."
},
{
"id": "hair_right",
"layers": [
"hair_back_right"
],
"anchors": [
[
556,
269
],
[
607,
292
]
],
"fields": [
{
"axis": [
[
504,
221
],
[
607,
292
]
],
"profile": 1,
"center": [
556,
269
],
"radii": [
110,
95
],
"hardness": 0.3
},
{
"axis": [
[
504,
221
],
[
607,
292
]
],
"profile": 2,
"center": [
607,
292
],
"radii": [
110,
95
],
"hardness": 0.3
}
],
"offsets": [
[
[
0,
0
],
[
-12,
15
],
[
3,
12
],
[
18,
3
],
[
17,
8
],
[
8,
17
],
[
-5,
22
],
[
0,
20
],
[
10,
12
],
[
19,
8
]
],
[
[
0,
0
],
[
-28,
29
],
[
-19,
24
],
[
12,
3
],
[
15,
1
],
[
0,
14
],
[
-22,
32
],
[
-24,
34
],
[
-14,
25
],
[
6,
15
]
]
],
"reason": "Source right locks fold down/in around frames 14 and 74–83; open outward around 39–52. Middle and tip have distinct trajectories. Strand-aligned quadratic weight fields spread bending across the whole blade; radial fields previously compressed its lower edge into flips."
},
{
"id": "hair_upper",
"layers": [
"hair_sweep_right"
],
"anchors": [
[
597,
248
]
],
"fields": [
{
"center": [
597,
248
],
"radii": [
89,
61
],
"hardness": 0.3,
"pins": [
{
"center": [
510,
216
],
"radii": [
43,
35
],
"hardness": 0.6
}
]
}
],
"offsets": [
[
[
0,
0
],
[
-11,
16
],
[
-10,
14
],
[
3,
3
],
[
-1,
6
],
[
-9,
9
],
[
-10,
10
],
[
-5,
20
],
[
4,
14
],
[
9,
3
]
]
],
"reason": "Upper rear tip moves less than the lower broad lock; use its separate observed trajectory."
},
{
"id": "front_left",
"layers": [
"fringe_left"
],
"anchors": [
[
438,
281
]
],
"fields": [
{
"center": [
438,
281
],
"radii": [
83,
106
],
"hardness": 0.3,
"pins": [
{
"center": [
418,
142
],
"radii": [
84,
72
],
"hardness": 0.6
}
]
}
],
"offsets": [
[
[
0,
0
],
[
-9,
8
],
[
0,
6
],
[
9,
-1
],
[
8,
1
],
[
2,
6
],
[
-5,
10
],
[
3,
6
],
[
9,
1
],
[
11,
-3
]
]
],
"reason": "Only part of the measured fringe residual is secondary bending; reduce it because source yaw is not represented by planar head roll."
},
{
"id": "temple",
"layers": [
"hair_temple"
],
"anchors": [
[
522,
302
]
],
"fields": [
{
"center": [
522,
302
],
"radii": [
64,
94
],
"hardness": 0.3,
"pins": [
{
"center": [
513,
223
],
"radii": [
55,
41
],
"hardness": 0.6
}
]
}
],
"offsets": [
[
[
0,
0
],
[
-10,
-5
],
[
2,
2
],
[
12,
1
],
[
8,
2
],
[
-8,
0
],
[
-12,
-4
],
[
-5,
3
],
[
1,
4
],
[
3,
6
]
]
],
"reason": "The vertical cheek strand moves differently from the broad rear blade; use its own smaller trajectory and root pin."
},
{
"id": "front_right",
"layers": [
"fringe_right"
],
"anchors": [
[
590,
183
]
],
"fields": [
{
"center": [
590,
183
],
"radii": [
96,
65
],
"hardness": 0.3,
"pins": [
{
"center": [
511,
130
],
"radii": [
43,
41
],
"hardness": 0.6
}
]
}
],
"offsets": [
[
[
0,
0
],
[
-12,
4
],
[
-14,
3
],
[
-3,
-2
],
[
-1,
-1
],
[
-6,
2
],
[
-7,
3
],
[
0,
7
],
[
6,
7
],
[
11,
5
]
]
],
"reason": "Flow fails near the thin tip; these smaller keys come from source contact-sheet inspection, not the invalid cumulative track."
},
{
"id": "coat_left",
"layers": [
"coat_left",
"coat_left_underlap"
],
"anchors": [
[
345,
585
],
[
289,
681
]
],
"fields": [
{
"center": [
345,
585
],
"radii": [
119,
97
],
"hardness": 0.3,
"pins": [
{
"center": [
429,
497
],
"radii": [
67,
53
],
"hardness": 0.6
}
]
},
{
"center": [
289,
681
],
"radii": [
176,
129
],
"hardness": 0.3,
"pins": [
{
"center": [
429,
497
],
"radii": [
67,
53
],
"hardness": 0.6
}
]
}
],
"offsets": [
[
[
0,
0
],
[
-8,
-3
],
[
-11,
-4
],
[
-8,
-5
],
[
-4,
-4
],
[
-6,
-4
],
[
-10,
-5
],
[
-4,
-1
],
[
0,
-2
],
[
10,
2
]
],
[
[
0,
0
],
[
-12,
-6
],
[
-24,
-7
],
[
-21,
-7
],
[
-7,
-2
],
[
-9,
-2
],
[
-23,
-6
],
[
-18,
1
],
[
-17,
0
],
[
4,
8
]
]
],
"reason": "Seat-level hem has smaller vertical motion and later horizontal turnaround than hair; do not use the rear-hem track that drifts onto the tail."
},
{
"id": "front_cloth",
"layers": [
"front_cloth",
"front_cloth_underlap",
"torso_core",
"torso_core_underlap",
"torso_underlap"
],
"anchors": [
[
513,
631
],
[
567,
711
]
],
"fields": [
{
"center": [
513,
631
],
"radii": [
74,
93
],
"hardness": 0.3,
"pins": [
{
"center": [
458,
541
],
"radii": [
77,
56
],
"hardness": 0.6
}
]
},
{
"center": [
567,
711
],
"radii": [
95,
90
],
"hardness": 0.3,
"pins": [
{
"center": [
458,
541
],
"radii": [
77,
56
],
"hardness": 0.6
}
]
}
],
"offsets": [
[
[
0,
0
],
[
-20,
-4
],
[
-27,
-5
],
[
-9,
8
],
[
15,
8
],
[
11,
8
],
[
-15,
-3
],
[
-11,
-5
],
[
-8,
-2
],
[
11,
1
]
],
[
[
0,
0
],
[
-3,
-6
],
[
-18,
-8
],
[
-19,
-6
],
[
8,
-2
],
[
11,
-1
],
[
-10,
-6
],
[
-10,
-8
],
[
-13,
-6
],
[
7,
-1
]
]
],
"reason": "Green cloth middle reverses before the tip; preserve that lag instead of translating the whole panel."
},
{
"id": "coat_right",
"layers": [
"coat_right",
"coat_right_underlap"
],
"anchors": [
[
591,
542
]
],
"fields": [
{
"center": [
591,
542
],
"radii": [
93,
77
],
"hardness": 0.3,
"pins": [
{
"center": [
529,
482
],
"radii": [
49,
34
],
"hardness": 0.6
}
]
}
],
"offsets": [
[
[
0,
0
],
[
-5,
0
],
[
-11,
5
],
[
-11,
7
],
[
4,
0
],
[
3,
0
],
[
-4,
4
],
[
-3,
6
],
[
-2,
9
],
[
14,
7
]
]
],
"reason": "Partly occluded cloth: conservative amplitude from the visible strip; source optical flow includes leg/cloth boundary ambiguity."
}
],
"evidence": "artifacts/flow_v6/flow/tracks.json and contact0/1.jpg",
"rejected": [
"coat_back_hem after occlusion",
"hair_left_mid after frame26",
"fringe_right_tip cumulative track"
]
}
全部最终图层：处理前 / 处理后 / Alpha
包含可见部件、隐藏补片、沿用的头部素材和背景。补片的“处理前”显示其生成图来源，处理后显示实际运行的贴图。透明区域采用棋盘底色。
全部 43 层与 3 个附加表情贴图缩略图
可播放、可拆层的 HTML 模型
支持 V5 / V6 切换、逐帧播放、图层开关、单独查看部件、骨骼编辑及保存。运行时只读固定生成贴图与动画数据，无需原片或模型服务。
检查、局限与复现
| 网格翻折 | 0；0–99 帧，每 0.25 帧检查一次，共 397 次 |
|---|---|
| Python / JS 最大顶点差 | 3.07e-12 px；9 个时刻、7 种部件 |
| 尾巴 | V5 的完整尾巴图层、骨盆控制保持一致，沿用已验证的保宽截面 |
| 输入保护 | 412 个文件未改，包括 old_stuff、V4、V5 和原 skill |
| 贴图来源 | 全部来自生成素材；原片只用于观察、测量和报告对照 |
完整代码测试 48 项通过。下载源码包在临时目录中解压后，9 项 V6 测试、模型导出和第 39 帧渲染通过；无需原视频或 LLM。下载包检查记录
上一版“首帧服装结构和衣发摆动已经修复”的表述过早。坐姿外观有所恢复，但躯干残留域、被遮挡的部件缺失和颜色延展补片仍限制了拆分质量。远侧手臂目前沿用躯干运动，转头仍是平面近似。零翻折只说明采样中的三角形方向，不能证明部件完整或动画相似。
cd model
python -m http.server 8765 --bind 127.0.0.1
# 浏览器打开 http://127.0.0.1:8765/
源码目录：experiments/flow_v6/。保存的控制点由本次观察编排，重放无需再次调用生图。日期：2026-09-28。