rui40000andClaude Opus 4.8
9c674bac42
上次加的 expand_beyond_cell 有个致命条件写错：
use_expand = expand_beyond_cell and mode != "none" and _HAS_SCIPY
选「不处理（输出不透明）」时整个修复被跳过，退回严格格线切分，
输出尺寸恒为格子大小（834/3 x 1112/3 = 278x371），角色照样被切。
当初这么写的理由是「不透明模式没有 alpha 可做连通分析」——
但定位和输出是两件事，内部照样可以算一份白底检测只用于圈定范围。
改为：locate_alpha() 专用于定位（三种 bg_mode 都算），
frame_alpha() 才是真正写进输出的。修复后三种模式尺寸完全一致。
连带解决「不处理」模式的一个矛盾：裁剪框为了不切断角色必然框进
邻居的像素，而该模式不做透明处理，邻居就会直接显示出来。
现在把非本角色区域按软 alpha 合成到白底 —— 保住完整角色，
也不会混进旁边那位。
crop_padding 默认 8 → 16：给描边/发光/阴影等后续特效留余量，
也免得内容贴着边看起来像被切。
验证：三种 bg_mode 全部「贴边被切的方向 = 无」。
判据按模式区分——透明模式看输出 alpha 的四边，不处理模式看非白像素的四边。
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
435 lines
20 KiB
Python
435 lines
20 KiB
Python
# -*- coding: utf-8 -*-
|
||
"""
|
||
八方向雪碧图序列拆分节点（Ruinode）
|
||
===================================
|
||
用于「8 方向行走动画」制作管线：把每帧排布着 8 个朝向的雪碧图序列，
|
||
拆成 8 条各自独立、可直接成片的动画序列。
|
||
|
||
典型管线：
|
||
角色图 → (GPTimage2) 八方向静态图 → (Seedance 首尾帧) 循环行走视频
|
||
→ VHS「Load Video」转序列帧（可选帧率）
|
||
→ 【本节点】拆分 + 方向编号 + 分组
|
||
→ 8×SaveImage 出序列帧，8×VHS「Video Combine」出透明 webm
|
||
|
||
为什么用固定网格而不是连通区域拆分（本仓库的 RuiSpriteSplitterRGBA）：
|
||
- 连通区域按包围盒排序，角色走动时位置会浮动，一旦跨过排序的行界，
|
||
方向对应关系就错乱 —— 几十上百帧里错一帧，整条动画就废了；
|
||
- 连通区域按各自 bbox 裁剪，每个 sprite 尺寸不同，无法直接合成视频。
|
||
固定网格没有这两个问题：格子位置恒定，方向对应天然稳定，尺寸也一致。
|
||
（连通区域拆分依然更适合单张静态合图，两者各有用武之地。）
|
||
|
||
白底转透明的关键点：角色身上常有白色衣物，按亮度阈值一刀切会把白衬衫
|
||
一起掏空。这里改为**从画面边缘做连通性判断**：只有与边缘相连的白色才算背景，
|
||
被角色包围的白色（衣服、高光）一律保留。
|
||
"""
|
||
import numpy as np
|
||
import torch
|
||
|
||
try:
|
||
import scipy.ndimage as ndi
|
||
_HAS_SCIPY = True
|
||
except Exception: # 理论上 ComfyUI 环境都有
|
||
_HAS_SCIPY = False
|
||
|
||
MAX_DIRS = 8
|
||
|
||
_BG_MODES = {
|
||
"白底转透明（推荐）": "white",
|
||
"已带透明通道": "keep",
|
||
"不处理（输出不透明）": "none",
|
||
}
|
||
|
||
# 与 3×3 中间留空的常见排布对应：行 1 面向观众、行 3 背对观众
|
||
_DEFAULT_NAMES = "SW,S,SE,W,E,NW,N,NE"
|
||
|
||
|
||
def _white_to_alpha(rgb, threshold, softness):
|
||
"""
|
||
白底 → alpha。rgb: (H,W,3) float[0,1]，返回 (H,W) float[0,1]。
|
||
|
||
先按「离白色多远」算出软 alpha 保住边缘抗锯齿，再用连通性把
|
||
与画面边缘相连的白色判为背景 —— 只有这一部分才真正抹成全透明。
|
||
这样角色内部的白衬衫、白高光不会被误伤。
|
||
"""
|
||
dist = 1.0 - rgb.min(axis=2) # 纯白=0，越大越不白
|
||
cut = max(1e-4, 1.0 - float(threshold))
|
||
soft = np.clip(dist / (cut * max(1e-3, softness)), 0.0, 1.0)
|
||
|
||
near_white = dist <= cut
|
||
if _HAS_SCIPY and near_white.any():
|
||
lab, n = ndi.label(near_white)
|
||
if n > 0:
|
||
border = np.concatenate([lab[0, :], lab[-1, :], lab[:, 0], lab[:, -1]])
|
||
ids = np.unique(border)
|
||
ids = ids[ids != 0]
|
||
if ids.size:
|
||
bg = np.isin(lab, ids)
|
||
soft = np.where(bg, 0.0, soft)
|
||
else:
|
||
soft = np.where(near_white, 0.0, soft)
|
||
return soft.astype(np.float32)
|
||
|
||
|
||
def _drop_fragments(alpha, ratio):
|
||
"""
|
||
清掉远小于主体的连通碎片。
|
||
|
||
网格切分难免会把相邻格子探过来的部件（手杖尖、飘起的衣角）切进本格，
|
||
这些碎片不仅难看，还会把 auto_crop 的包围盒撑大。
|
||
按「面积不足主体 ratio 倍」判定为碎片，这样与身体相连的道具不会被误删。
|
||
"""
|
||
if ratio <= 0 or not _HAS_SCIPY:
|
||
return alpha
|
||
solid = alpha > 0.1
|
||
if not solid.any():
|
||
return alpha
|
||
lab, n = ndi.label(solid)
|
||
if n <= 1:
|
||
return alpha
|
||
areas = np.bincount(lab.ravel())
|
||
areas[0] = 0
|
||
keep = areas >= areas.max() * float(ratio)
|
||
keep[0] = False
|
||
return np.where(keep[lab], alpha, 0.0).astype(np.float32)
|
||
|
||
|
||
def _assign_blocks(alpha_full, ybnd, xbnd, cols, frag_ratio):
|
||
"""
|
||
全图连通标记 + 按质心把每块归属到格子。
|
||
|
||
这样格子只负责回答「这是哪个方向」，角色的实际范围由它自己的连通块决定，
|
||
因此**角色超出格子边界也不会被切**（原本严格等分会把探出去的脚、
|
||
飘起的斗篷直接截断）。归属按质心判定，每块只属于一个格子，
|
||
相邻角色不会被重复计入。
|
||
|
||
返回 (lab, {cell_id: [块标签...]})；scipy 不可用时返回 (None, None)。
|
||
"""
|
||
if not _HAS_SCIPY:
|
||
return None, None
|
||
solid = alpha_full > 0.1
|
||
if not solid.any():
|
||
return None, {}
|
||
lab, n = ndi.label(solid)
|
||
if n == 0:
|
||
return lab, {}
|
||
areas = np.bincount(lab.ravel())
|
||
areas[0] = 0
|
||
big = areas.max()
|
||
cents = ndi.center_of_mass(solid, lab, np.arange(1, n + 1))
|
||
rows = len(ybnd) - 1
|
||
out = {}
|
||
for i in range(1, n + 1):
|
||
if frag_ratio > 0 and areas[i] < big * float(frag_ratio):
|
||
continue # 碎片，丢弃
|
||
cy, cx = cents[i - 1]
|
||
r = min(rows - 1, max(0, int(np.searchsorted(ybnd, cy, "right") - 1)))
|
||
c = min(cols - 1, max(0, int(np.searchsorted(xbnd, cx, "right") - 1)))
|
||
out.setdefault(r * cols + c, []).append(i)
|
||
return lab, out
|
||
|
||
|
||
def _bbox(alpha, thr=0.02):
|
||
"""内容包围盒 (y0,y1,x0,x1)，无内容时返回 None。"""
|
||
m = alpha > thr
|
||
if not m.any():
|
||
return None
|
||
ys = np.where(m.any(axis=1))[0]
|
||
xs = np.where(m.any(axis=0))[0]
|
||
return int(ys[0]), int(ys[-1]) + 1, int(xs[0]), int(xs[-1]) + 1
|
||
|
||
|
||
def _parse_cells(text, total):
|
||
out = set()
|
||
for tok in str(text or "").replace("，", ",").split(","):
|
||
tok = tok.strip()
|
||
if not tok:
|
||
continue
|
||
try:
|
||
v = int(tok)
|
||
except ValueError:
|
||
continue
|
||
if 0 <= v < total:
|
||
out.add(v)
|
||
return out
|
||
|
||
|
||
class RuiEightDirSplit:
|
||
"""八方向雪碧图序列 → 8 条独立动画序列。"""
|
||
|
||
@classmethod
|
||
def INPUT_TYPES(cls):
|
||
return {
|
||
"required": {
|
||
"images": ("IMAGE", {
|
||
"tooltip": "视频转出的序列帧，每帧是一张排布着多个朝向的雪碧图。"
|
||
}),
|
||
"grid_cols": ("INT", {
|
||
"default": 3, "min": 1, "max": 8, "step": 1,
|
||
"tooltip": "雪碧图的列数。"
|
||
}),
|
||
"grid_rows": ("INT", {
|
||
"default": 3, "min": 1, "max": 8, "step": 1,
|
||
"tooltip": "雪碧图的行数。"
|
||
}),
|
||
"empty_cells": ("STRING", {
|
||
"default": "4", "multiline": False,
|
||
"tooltip": "空格子的序号（行优先、从 0 开始，逗号分隔）。\n"
|
||
"3×3 布局中间留空即填 4。留空表示没有空格。"
|
||
}),
|
||
"direction_names": ("STRING", {
|
||
"default": _DEFAULT_NAMES, "multiline": False,
|
||
"tooltip": "按「跳过空格后的先后顺序」给每个方向命名，逗号分隔。\n"
|
||
"默认对应 3×3 中间留空、行 1 面向观众的排布：\n"
|
||
" SW S SE\n"
|
||
" W E\n"
|
||
" NW N NE\n"
|
||
"名字只用于 info 与你自己辨认，不影响画面内容。"
|
||
}),
|
||
"bg_mode": (list(_BG_MODES.keys()), {
|
||
"default": "白底转透明（推荐）",
|
||
"tooltip": "webm 要保留透明就必须先把白底转成 alpha。\n"
|
||
"转换只把与画面边缘相连的白色判为背景，\n"
|
||
"角色身上的白衣服、白高光会被保留。"
|
||
}),
|
||
"bg_threshold": ("FLOAT", {
|
||
"default": 0.92, "min": 0.5, "max": 1.0, "step": 0.005,
|
||
"tooltip": "白底判定阈值：像素三通道最小值高于它才算「接近白」。\n"
|
||
"背景没扣干净就调低，角色边缘被啃掉就调高。"
|
||
}),
|
||
"edge_softness": ("FLOAT", {
|
||
"default": 1.0, "min": 0.1, "max": 4.0, "step": 0.05,
|
||
"tooltip": "边缘过渡宽度。原图边缘带抗锯齿，过渡太硬会有锯齿白边；\n"
|
||
"调大更柔和，调小更锐利。"
|
||
}),
|
||
"fragment_threshold": ("FLOAT", {
|
||
"default": 0.05, "min": 0.0, "max": 0.5, "step": 0.01,
|
||
"tooltip": "清掉面积不足主体这一比例的连通碎片。\n"
|
||
"网格切分会把相邻格子探过来的部件（手杖尖、飘起的衣角）\n"
|
||
"切进本格，既难看又会撑大自动裁剪的范围。\n"
|
||
"0 = 不清理；与身体相连的道具不会被误删。"
|
||
}),
|
||
"expand_beyond_cell": ("BOOLEAN", {
|
||
"default": True,
|
||
"tooltip": "允许角色超出格子边界（强烈建议开启）。\n"
|
||
"关闭时按格子严格切分，角色只要探出格线就会被切断\n"
|
||
"——最常见的是脚、飘起的斗篷和手杖被削掉一截。\n"
|
||
"开启后格子只用来判定「这是哪个方向」，实际范围由角色\n"
|
||
"自身的连通区域决定；按质心归属，相邻角色不会被卷进来。\n"
|
||
"对三种 bg_mode 都有效：选「不处理」时也会在内部算一份\n"
|
||
"白底检测来圈定范围，输出仍保持不透明。"
|
||
}),
|
||
"auto_crop": ("BOOLEAN", {
|
||
"default": True,
|
||
"tooltip": "按内容裁掉多余空白。\n"
|
||
"裁剪框取「该方向所有帧的并集」，因此整条序列尺寸一致，\n"
|
||
"既能合成视频，角色也不会在帧间跳动。"
|
||
}),
|
||
"crop_padding": ("INT", {
|
||
"default": 16, "min": 0, "max": 200, "step": 1,
|
||
"tooltip": "裁剪时在内容外保留的边距（像素）。\n"
|
||
"给足边距不仅好看，也给后续的描边、发光、\n"
|
||
"阴影等特效留出余量，免得贴着边显得像被切了。"
|
||
}),
|
||
},
|
||
"optional": {
|
||
"masks": ("MASK", {
|
||
"tooltip": "可选。已有的透明通道（如上游抠图结果），\n"
|
||
"配合 bg_mode=已带透明通道 使用。"
|
||
}),
|
||
},
|
||
}
|
||
|
||
RETURN_TYPES = ("IMAGE",) * MAX_DIRS + ("STRING",)
|
||
RETURN_NAMES = tuple(f"dir_{i + 1}" for i in range(MAX_DIRS)) + ("info",)
|
||
FUNCTION = "split"
|
||
CATEGORY = "Rui-Node🐶/图像调节🎨"
|
||
|
||
@classmethod
|
||
def VALIDATE_INPUTS(cls, **kwargs):
|
||
return True
|
||
|
||
def split(self, images, grid_cols, grid_rows, empty_cells,
|
||
direction_names, bg_mode, bg_threshold, edge_softness,
|
||
fragment_threshold, expand_beyond_cell, auto_crop, crop_padding,
|
||
masks=None):
|
||
mode = _BG_MODES.get(bg_mode, "white")
|
||
B, H, W, C = images.shape
|
||
cols, rows = int(grid_cols), int(grid_rows)
|
||
total = cols * rows
|
||
empties = _parse_cells(empty_cells, total)
|
||
cell_ids = [i for i in range(total) if i not in empties]
|
||
n_dir = len(cell_ids)
|
||
|
||
names = [s.strip() for s in str(direction_names).replace("，", ",").split(",")
|
||
if s.strip()]
|
||
while len(names) < n_dir:
|
||
names.append(f"dir{len(names) + 1}")
|
||
|
||
arr = images.detach().cpu().float().numpy()
|
||
if C == 4:
|
||
rgb_all, a_in = arr[..., :3], arr[..., 3]
|
||
else:
|
||
rgb_all, a_in = arr[..., :3], None
|
||
if masks is not None:
|
||
m = masks.detach().cpu().float().numpy()
|
||
if m.ndim == 2:
|
||
m = m[None]
|
||
if m.shape[0] != B:
|
||
m = np.repeat(m[:1], B, axis=0)
|
||
a_in = m
|
||
|
||
# 格子边界按浮点等分再取整，避免整除不尽时累计误差（如 1112/3）
|
||
ybnd = [int(round(r * H / rows)) for r in range(rows + 1)]
|
||
xbnd = [int(round(c * W / cols)) for c in range(cols + 1)]
|
||
|
||
pad = int(crop_padding)
|
||
# 注意不要因为 bg_mode=不处理 就跳过这条路径：那样会退回严格格线切分，
|
||
# 角色照样被切。定位与输出是两件事，分开处理即可。
|
||
use_expand = bool(expand_beyond_cell) and _HAS_SCIPY
|
||
|
||
def locate_alpha(b):
|
||
"""
|
||
仅用于圈定角色范围的 alpha。
|
||
即使用户选了「不处理（输出不透明）」，这里也要照算一份 ——
|
||
否则无从判断角色到哪儿为止，只能按格线硬切。
|
||
"""
|
||
if mode == "keep" and a_in is not None:
|
||
return a_in[b]
|
||
return _white_to_alpha(rgb_all[b], bg_threshold, edge_softness)
|
||
|
||
def frame_alpha(b):
|
||
"""真正写进输出的 alpha。"""
|
||
if mode == "none":
|
||
return np.ones((H, W), dtype=np.float32)
|
||
return locate_alpha(b)
|
||
|
||
outs, notes = [], []
|
||
|
||
if use_expand:
|
||
# ===== 内容自适应：格子只定方向，范围由角色自身的连通块决定 =====
|
||
# 第一遍只求包围盒（全图坐标），不留像素，避免整段序列驻留内存
|
||
boxes = [None] * n_dir
|
||
for b in range(B):
|
||
lab, groups = _assign_blocks(locate_alpha(b), ybnd, xbnd,
|
||
cols, fragment_threshold)
|
||
if lab is None:
|
||
continue
|
||
for di, cid in enumerate(cell_ids):
|
||
blk = groups.get(cid)
|
||
if not blk:
|
||
continue
|
||
bb = _bbox(np.isin(lab, blk).astype(np.float32), 0.5)
|
||
if bb is None:
|
||
continue
|
||
boxes[di] = bb if boxes[di] is None else (
|
||
min(boxes[di][0], bb[0]), max(boxes[di][1], bb[1]),
|
||
min(boxes[di][2], bb[2]), max(boxes[di][3], bb[3]))
|
||
|
||
# 没有 auto_crop 时退回该格的格线范围，仍允许块超界的像素带出来
|
||
final = []
|
||
for di, cid in enumerate(cell_ids):
|
||
r, c = divmod(cid, cols)
|
||
if auto_crop and boxes[di] is not None:
|
||
y0, y1, x0, x1 = boxes[di]
|
||
y0 = max(0, y0 - pad); x0 = max(0, x0 - pad)
|
||
y1 = min(H, y1 + pad); x1 = min(W, x1 + pad)
|
||
elif boxes[di] is not None:
|
||
y0, y1, x0, x1 = (min(ybnd[r], boxes[di][0]),
|
||
max(ybnd[r + 1], boxes[di][1]),
|
||
min(xbnd[c], boxes[di][2]),
|
||
max(xbnd[c + 1], boxes[di][3]))
|
||
else:
|
||
y0, y1, x0, x1 = ybnd[r], ybnd[r + 1], xbnd[c], xbnd[c + 1]
|
||
final.append((y0, y1, x0, x1))
|
||
outs.append(np.zeros((B, y1 - y0, x1 - x0, 4), dtype=np.float32))
|
||
|
||
# 第二遍按并集框提取；只保留归属本格的连通块，邻居不会混进来
|
||
for b in range(B):
|
||
la = locate_alpha(b)
|
||
af = frame_alpha(b)
|
||
lab, groups = _assign_blocks(la, ybnd, xbnd, cols, fragment_threshold)
|
||
for di, cid in enumerate(cell_ids):
|
||
y0, y1, x0, x1 = final[di]
|
||
rgb = rgb_all[b, y0:y1, x0:x1]
|
||
blk = groups.get(cid) if groups else None
|
||
m = (np.isin(lab[y0:y1, x0:x1], blk) if blk
|
||
else np.zeros(rgb.shape[:2], dtype=bool))
|
||
|
||
if mode == "none":
|
||
# 不透明输出。裁剪框为了不切断角色必然会框进邻居的像素，
|
||
# 「原样保留」与「不切断」不可兼得 —— 这里把非本角色的部分
|
||
# 合成到白底，既保住完整的角色，也不会混进旁边那位。
|
||
a = (locate_alpha(b)[y0:y1, x0:x1] * m)[..., None]
|
||
outs[di][b, ..., :3] = rgb * a + (1.0 - a)
|
||
outs[di][b, ..., 3] = 1.0
|
||
continue
|
||
|
||
outs[di][b, ..., :3] = rgb
|
||
if blk:
|
||
outs[di][b, ..., 3] = af[y0:y1, x0:x1] * m
|
||
|
||
for di in range(n_dir):
|
||
y0, y1, x0, x1 = final[di]
|
||
notes.append(f"{di + 1}.{names[di]} 格{cell_ids[di]} "
|
||
f"{x1 - x0}×{y1 - y0}")
|
||
outs[di] = torch.from_numpy(outs[di])
|
||
else:
|
||
# ===== 严格按格线切分（角色探出格线会被截断）=====
|
||
per_dir = [[] for _ in range(n_dir)]
|
||
boxes = [None] * n_dir
|
||
for b in range(B):
|
||
af = frame_alpha(b)
|
||
for di, cid in enumerate(cell_ids):
|
||
r, c = divmod(cid, cols)
|
||
y0, y1, x0, x1 = ybnd[r], ybnd[r + 1], xbnd[c], xbnd[c + 1]
|
||
rgb = rgb_all[b, y0:y1, x0:x1]
|
||
alpha = af[y0:y1, x0:x1]
|
||
if mode != "none":
|
||
alpha = _drop_fragments(alpha, fragment_threshold)
|
||
per_dir[di].append((rgb, alpha))
|
||
if auto_crop:
|
||
bb = _bbox(alpha)
|
||
if bb is not None:
|
||
boxes[di] = bb if boxes[di] is None else (
|
||
min(boxes[di][0], bb[0]), max(boxes[di][1], bb[1]),
|
||
min(boxes[di][2], bb[2]), max(boxes[di][3], bb[3]))
|
||
for di in range(n_dir):
|
||
frames = per_dir[di]
|
||
ch, cw = frames[0][0].shape[:2]
|
||
if auto_crop and boxes[di] is not None:
|
||
y0, y1, x0, x1 = boxes[di]
|
||
y0 = max(0, y0 - pad); x0 = max(0, x0 - pad)
|
||
y1 = min(ch, y1 + pad); x1 = min(cw, x1 + pad)
|
||
else:
|
||
y0, y1, x0, x1 = 0, ch, 0, cw
|
||
stack = np.empty((len(frames), y1 - y0, x1 - x0, 4), dtype=np.float32)
|
||
for fi, (rgb, alpha) in enumerate(frames):
|
||
stack[fi, ..., :3] = rgb[y0:y1, x0:x1]
|
||
stack[fi, ..., 3] = alpha[y0:y1, x0:x1]
|
||
outs.append(torch.from_numpy(stack))
|
||
notes.append(f"{di + 1}.{names[di]} 格{cell_ids[di]} "
|
||
f"{x1 - x0}×{y1 - y0}")
|
||
|
||
info = (f"输入 {B} 帧 {W}×{H} → {cols}×{rows} 网格，"
|
||
f"空格 {sorted(empties) if empties else '无'}，"
|
||
f"得到 {n_dir} 个方向 × {B} 帧\n" + " | ".join(notes))
|
||
if n_dir > MAX_DIRS:
|
||
info += f"\n⚠ 方向数 {n_dir} 超过输出口数量 {MAX_DIRS}，只输出前 {MAX_DIRS} 个"
|
||
elif n_dir < MAX_DIRS:
|
||
info += (f"\n⚠ 方向数 {n_dir} 少于输出口数量 {MAX_DIRS}，"
|
||
f"dir_{n_dir + 1}~dir_{MAX_DIRS} 为占位空图，请勿使用")
|
||
print(f"[Ruinode-8Dir] {info}")
|
||
|
||
# 输出口数量固定，方向不足时补占位图，避免下游拿到 None 直接报错
|
||
blank = torch.zeros((1, 8, 8, 4), dtype=torch.float32)
|
||
result = [outs[i] if i < len(outs) else blank for i in range(MAX_DIRS)]
|
||
return tuple(result) + (info,)
|
||
|
||
|
||
NODE_CLASS_MAPPINGS = {
|
||
"RuiEightDirSplit": RuiEightDirSplit,
|
||
}
|
||
NODE_DISPLAY_NAME_MAPPINGS = {
|
||
"RuiEightDirSplit": "八方向序列拆分 / 8-Direction Sprite Split",
|
||
}
|