#!/usr/bin/env python3
"""把「格狀排列的半身講師動作表」(例如 4 欄 x 2 列,每格一個動作,格上有編號與文字標籤)切開、去背、對齊,
產生與 avatar-sheet.py 相同格式的輸出(pose{i}.png 與 blend_{a}_{b}_{k}.png),給 make-video2.mjs 使用。
用法:python3 -I tools/avatar-grid.py <動作表圖> <輸出資料夾> <人物高度px> <欄x列,例如 4x2> <動作順序,例如 3,5,1,6,7,4,8,2>
動作順序:輸出第 i 個動作 = 動作表上第 order[i] 格(由左到右、由上到下,從 1 算起);最後一個動作會當成開場與收尾的招呼動作。
做法:每格內以邊界填色去掉白底,只保留最大的那塊(編號圓點與文字標籤因此被丟掉);以頭頂與身體中線對齊,
動作底部(切到腰部)會補滿到畫面底邊,疊在影片右下角時看起來像從畫面底部冒出來。
"""
import sys
import numpy as np
from PIL import Image, ImageFilter
from scipy import ndimage

BF = 4
sheet, outdir, H = sys.argv[1], sys.argv[2], int(sys.argv[3])
cols, rows = (int(v) for v in sys.argv[4].split('x'))
order = [int(v) - 1 for v in sys.argv[5].split(',')]
N = len(order)
BOTTOM_PAD = 0   # 半身圖疊圖時貼齊影片底邊與右緣(make-video2 的 --avatar-grid 模式)
INSET = 8        # 格線寬度,往內縮避免切到黑色格線

im = Image.open(sheet).convert('RGB')
W, Hh = im.size
cw, ch = W // cols, Hh // rows
figs = []
for idx in order:
    r, c = divmod(idx, cols)
    cell = im.crop((c * cw + INSET, r * ch + INSET, (c + 1) * cw - INSET, (r + 1) * ch - INSET))
    a = np.asarray(cell).astype(np.int16)
    near_white = a.min(axis=2) >= 222
    lab, n = ndimage.label(near_white)
    border = set(np.unique(np.concatenate([lab[0], lab[-1], lab[:, 0], lab[:, -1]]))) - {0}
    bg = np.isin(lab, list(border))
    fg = ~bg
    flab, fn = ndimage.label(fg)
    sizes = ndimage.sum(fg, flab, range(1, fn + 1))
    keep = flab == (1 + int(np.argmax(sizes)))
    keep = ndimage.binary_fill_holes(keep)
    alpha = Image.fromarray((keep * 255).astype(np.uint8)).filter(ImageFilter.MinFilter(3)).filter(ImageFilter.GaussianBlur(1.0))
    rgba = cell.convert('RGBA'); rgba.putalpha(alpha)
    m = np.asarray(alpha) > 40
    ys, xs = np.where(m)
    top, bottom, left, right = ys.min(), ys.max(), xs.min(), xs.max()
    hf = bottom - top
    def cx(r0, r1):
        rr = m[top + int(hf * r0): top + int(hf * r1)]
        return np.where(rr)[1].mean()
    axis = (cx(0.0, 0.18) + cx(0.55, 0.75)) / 2
    figs.append(dict(img=rgba.crop((left, top, right + 1, bottom + 1)), fcx=axis - left, h=bottom - top + 1, w=right - left + 1))

# 共同縮放:以所有動作高度的中位數為準(每格拍攝比例相同,只有手勢不同),頭頂對齊
med = float(np.median([f['h'] for f in figs]))
s = H / med
scaled = []
for f in figs:
    im2 = f['img'].resize((max(1, round(f['w'] * s)), max(1, round(f['h'] * s))), Image.LANCZOS)
    scaled.append(dict(img=im2, fcx=f['fcx'] * s))
pad = 10
maxL = max(f['fcx'] for f in scaled) + pad
maxR = max(f['img'].width - f['fcx'] for f in scaled)   # 右側不留邊:半身圖的右肩常被切在格邊,疊圖時讓它貼齊影片右緣
CW = int(maxL + maxR)
CH = max(f['img'].height for f in scaled) + BOTTOM_PAD
poses = []
for i, f in enumerate(scaled):
    canvas = Image.new('RGBA', (CW, CH), (0, 0, 0, 0))
    x0 = int(round(maxL - f['fcx']))
    canvas.paste(f['img'], (x0, 0), f['img'])
    # 底部補滿:把最後一列往下複製到畫面底邊(動作切在腰部,底邊本來就是被切掉的)
    arr = np.asarray(canvas).copy()
    last = f['img'].height - 1
    if last < CH - 1:
        # 最底下幾列常是邊緣羽化(半透明、混到背景灰),改從往上 10 列的實色列開始往下複製
        src = max(0, last - 10)
        arr[src + 1:, :, :] = arr[src, :, :]
    canvas = Image.fromarray(arr, 'RGBA')
    canvas.save(f'{outdir}/pose{i}.png')
    poses.append(canvas)

def premult(img):
    a = np.asarray(img).astype(np.float32)
    a[:, :, :3] *= a[:, :, 3:4] / 255.0
    return a

pm = [premult(p) for p in poses]
for a in range(N):
    for b in range(N):
        if a == b:
            continue
        for k in range(1, BF + 1):
            t = k / (BF + 1)
            mix = pm[a] * (1 - t) + pm[b] * t
            al_ = mix[:, :, 3:4]
            out = np.zeros_like(mix)
            out[:, :, :3] = np.where(al_ > 0.5, mix[:, :, :3] * 255.0 / np.maximum(al_, 1e-3), 0)
            out[:, :, 3:4] = al_
            Image.fromarray(np.clip(out, 0, 255).astype(np.uint8), 'RGBA').save(f'{outdir}/blend_{a}_{b}_{k}.png')
print(f'{CW}x{CH} poses={N} blend={BF}')
