' if botsu is not None else ""
wd = f"width:{w}px;" if w else ""
return f'
{lab}{body}{bs}
'
def stamp(text, t, style):
return f'
{text}
'
def tag(text):
return f'
{text}
'
def paper(inner, t_in=0.0, flights=(), style=""):
"""flights: [(飛んでいく秒, 戻る秒), ...] 脱線のときに紙が画面の外へ飛ぶ"""
an = [f"paperIn .35s ease-out {t_in:.3f}s both"] if t_in is not None else []
for a, b in flights:
an.append(f"fly .5s cubic-bezier(.5,0,.9,.4) {a:.3f}s forwards")
an.append(f"back .5s cubic-bezier(.2,1,.4,1) {b:.3f}s forwards")
return f'
{inner}
'
def dassen(img, t_in, t_out, words="", word_style="top:200px", src_note=""):
"""画面いっぱいの写真・イラスト（t_in〜t_out）。words は大きな一言（ は boom で出る）"""
p = "file://" + os.path.abspath(img)
out = (f'')
if words:
spans, k = [], 0
for part in words.split("|"):
if part == " ":
spans.append(" "); continue
st = ""
if part.startswith("!"):
part, st = part[1:], "color:#ffd84a;"
spans.append(f'{part}')
k += 1
out += f'
{"".join(spans)}
'
if src_note:
out += (f'
{src_note}
')
return out
def shake_js(windows, el="F", amp=5):
"""判子などで画面を少し揺らす（windows: [(開始秒, 長さ)]）"""
w = json.dumps(windows)
return (f'')
def counter_js(el, t0, t1, v1, dec=1, suffix=""):
"""数え上げ（t0〜t1 で 0→v1）"""
return (f'')
# ---------------- 場面の時刻 ----------------
class Timeline:
def __init__(self, first=None):
self.idx = load(first)
self.total = layout(self.idx)
def start(self, i):
return self.idx[i]["start"] / FPS
def speech(self, i):
return self.idx[i]["speech"] / FPS
def end(self, i):
return (self.idx[i]["start"] + self.idx[i]["frames"]) / FPS
# ---------------- ドット絵のエイイチ ----------------
def fill_teeth(im):
"""顔の中の透明な小さい穴（歯）を白で埋める（v4 の差分は歯が透明）"""
from scipy import ndimage as nd
a = np.array(im.convert("RGBA"))
holes = a[..., 3] < 128
lab, n = nd.label(holes)
edge = set(np.unique(np.concatenate([lab[0], lab[-1], lab[:, 0], lab[:, -1]]))) | {0}
for k in range(1, n + 1):
if k in edge:
continue
m = lab == k
if m.sum() < 80:
a[m] = (250, 250, 246, 255)
return Image.fromarray(a)
class Eiichi:
POSES = {1: "1_腕組み", 2: "2_うなずく", 3: "3_指さす", 4: "4_首をかしげる", 5: "5_ひらめく",
6: "6_のけぞって驚く", 7: "7_半目でツッコむ", 8: "8_吹き出して笑う"}
def __init__(self, folder="src/ei"):
self.raw = {}
for k, n in self.POSES.items():
fs = [f for f in os.listdir(folder) if f.startswith(f"エイイチ_{n}_100px")]
im = Image.open(os.path.join(folder, sorted(fs)[-1])).convert("RGBA")
self.raw[k] = fill_teeth(im)
self._cache = {}
def sprite(self, pose, scale=4, glow=False, crop=1.0):
key = (pose, scale, glow, crop)
if key not in self._cache:
im = self.raw[pose]
im = im.resize((im.width * scale, im.height * scale), Image.NEAREST)
if crop < 1.0:
im = im.crop((0, 0, im.width, int(im.height * crop)))
if glow:
pad = 60
g = Image.new("RGBA", (im.width + 2 * pad, im.height + 2 * pad), (0, 0, 0, 0))
al = Image.new("L", g.size, 0); al.paste(im.split()[3], (pad, pad))
al = al.filter(ImageFilter.GaussianBlur(28)).point(lambda v: int(v * 0.35))
g = Image.merge("RGBA", (Image.new("L", g.size, 245), Image.new("L", g.size, 243), Image.new("L", g.size, 238), al))
g.alpha_composite(im, (pad, pad))
im = g
self._cache[key] = (np.array(im).astype(np.float32), (-60 if glow else 0))
return self._cache[key]
def blend(dst, src, x, y, alpha=1.0):
"""dst: H×W×3 uint8、src: h×w×4 float（0-255）を (x,y) に重ねる（はみ出しは切る）"""
h, w = src.shape[:2]
DH, DW = dst.shape[:2] # 縦長（ショート）でも使えるように、重ね先の大きさで切る
x0, y0, x1, y1 = max(0, x), max(0, y), min(DW, x + w), min(DH, y + h)
if x0 >= x1 or y0 >= y1:
return
s = src[y0 - y:y1 - y, x0 - x:x1 - x]
a = s[..., 3:4] / 255.0 * alpha
d = dst[y0:y1, x0:x1].astype(np.float32)
dst[y0:y1, x0:x1] = (d * (1 - a) + s[..., :3] * a + 0.5).clip(0, 255).astype(np.uint8)
def rotate_about(img, deg, pivot, pad=760):
"""img（PIL RGBA）を pivot（img 内の座標）のまわりに回す。返り値: (回した画像, img の左上からのずれ)"""
c = Image.new("RGBA", (img.width + 2 * pad, img.height + 2 * pad), (0, 0, 0, 0))
c.paste(img, (pad, pad))
c = c.rotate(-deg, resample=Image.BICUBIC, center=(pad + pivot[0], pad + pivot[1]))
bb = c.getbbox()
if not bb:
return c, (-pad, -pad)
return c.crop(bb), (bb[0] - pad, bb[1] - pad)
def ease(u):
u = min(1.0, max(0.0, u))
return u * u * (3 - 2 * u)
# ---------------- 字幕 ----------------
SUBCSS = """*{margin:0;padding:0}html,body{width:1920px;height:1080px;background:transparent}
svg{position:absolute;left:0;top:0}
.t{font-family:'Noto Sans CJK JP',sans-serif;font-size:52px;font-weight:800;fill:#fff;stroke:#0D0D0D;stroke-width:11px;stroke-linejoin:round;stroke-linecap:round;paint-order:stroke fill}"""
def fill_holes(path, ink=(13, 13, 13), maxarea=400):
from scipy import ndimage as _nd
a = np.array(Image.open(path).convert('RGBA')).astype(np.float32)
al = a[..., 3] / 255.0
lab, n = _nd.label(al < 0.98)
if not n:
return
edge = set(np.unique(np.concatenate([lab[0], lab[-1], lab[:, 0], lab[:, -1]]))) | {0}
sz = _nd.sum(np.ones_like(al), lab, range(1, n + 1))
keep = np.array([0] + [(k not in edge and sz[k - 1] <= maxarea) for k in range(1, n + 1)])
m = keep[lab].astype(bool)
if not m.any():
return
for c in range(3):
a[..., c][m] = a[..., c][m] * al[m] + ink[c] * (1 - al[m])
a[..., 3][m] = 255
Image.fromarray(a.clip(0, 255).astype(np.uint8)).save(path)
def make_subs(idx, pg):
"""render.py と同じ字幕（中央・白・黒縁・2行まで）。png/sub/{番号}_{枚目}.png"""
os.makedirs("png/sub", exist_ok=True)
bad = []
for o in idx:
if not o["text"]:
continue
for pk, ptxt in enumerate(sub_pages(o["text"])):
lines = split_lines(ptxt) if ptxt else None
if lines is None:
bad.append(f"字幕 {o['i']:03d}: 2行に割れない「{o['text']}」"); continue
out = f"png/sub/{o['i']:03d}_{pk}.png"
key = hashlib.md5(repr(lines).encode()).hexdigest()[:10]
if os.path.exists(out) and os.path.exists(out + ".key") and open(out + ".key").read() == key:
continue
svg_lines = ''.join(f'{_html.escape(ln)}' for ln in lines)
pg.set_content(f"""
""")
pg.evaluate(f"""() => {{
const ts=[...document.querySelectorAll('text.t')]; const lh=72; const n=ts.length;
let fs=52; for (const t of ts){{ const w=t.getComputedTextLength(); if (w>{SUB_MAXW}) fs=Math.min(fs, Math.floor(52*{SUB_MAXW}/w)); }}
ts.forEach((t,k)=>{{ t.style.fontSize=fs+'px'; t.setAttribute('y', 1080-62-(n-1-k)*lh); }}); }}""")
pg.screenshot(path=out, omit_background=True)
fill_holes(out)
open(out + ".key", "w").write(key)
if bad:
sys.exit("\n".join(bad))
def sub_layer(path):
im = Image.open(path).convert("RGBA")
bb = im.getbbox()
return np.array(im.crop(bb)).astype(np.float32), bb[0], bb[1]
# ---------------- 紙の映像を撮る ----------------
def block_hash(b, n):
return hashlib.md5((b.get("html", "") + repr(b.get("video")) + str(n)).encode()).hexdigest()[:10]
def render_block(b, n, out, pg=None):
"""b: dict(name, html | video=(映像, 開始秒, テロップpng)), n: コマ数 → out（mp4・crf 14）"""
tmp = out[:-4] + ".part.mp4"
if "video" in b:
src, ss, ovl = b["video"]
ins = ["-ss", f"{ss:.3f}", "-stream_loop", "-1", "-i", src]
fc = f"[0:v]scale=1920:1080:force_original_aspect_ratio=increase,crop=1920:1080,fps={FPS},setsar=1[b]"
if ovl:
ins += ["-loop", "1", "-i", ovl]
fc += ";[b][1:v]overlay[v]"
else:
fc += ";[b]null[v]"
subprocess.run(["ffmpeg", "-v", "error", "-y"] + ins + ["-filter_complex", fc, "-map", "[v]", "-frames:v", str(n),
"-c:v", "libx264", "-preset", "ultrafast", "-crf", "14", "-pix_fmt", "yuv420p", "-r", str(FPS), "-an", tmp], check=True)
os.replace(tmp, out)
return
with open("_block.html", "w", encoding="utf-8") as f:
f.write(b["html"])
pg.goto("file://" + os.path.abspath("_block.html"))
pg.evaluate("document.fonts.ready")
pg.wait_for_timeout(300)
p = subprocess.Popen(["ffmpeg", "-v", "error", "-y", "-f", "image2pipe", "-framerate", str(FPS), "-c:v", "mjpeg", "-i", "-",
"-c:v", "libx264", "-preset", "ultrafast", "-crf", "14", "-pix_fmt", "yuv420p", "-r", str(FPS), "-an", tmp],
stdin=subprocess.PIPE)
last, shots = None, 0
for f in range(n):
t = f / FPS
if last is None or pg.evaluate(f"window.busy({t})"):
pg.evaluate(f"window.setT({t})")
last = pg.screenshot(type="jpeg", quality=93); shots += 1
p.stdin.write(last)
p.stdin.close()
if p.wait():
sys.exit(f"紙の映像の書き出しに失敗: {b['name']}")
os.replace(tmp, out)
print(f" {b['name']}: {n}コマ（撮影 {shots}）", flush=True)
def still(b, t, pg, path):
with open("_block.html", "w", encoding="utf-8") as f:
f.write(b["html"])
pg.goto("file://" + os.path.abspath("_block.html"))
pg.evaluate("document.fonts.ready")
pg.wait_for_timeout(200)
pg.evaluate(f"window.setT({t})")
pg.screenshot(path=path, type="jpeg", quality=90)
# ================= 重ねる（ハコブくん・エイイチ・字幕） =================
class Cast:
"""フレームごとのハコブくんとエイイチの姿を決めて、背景に重ねる"""
def __init__(self, sc, last_i):
self.sc, self.idx = sc, sc.tl.idx[:last_i + 1]
idx = self.idx
self.total = idx[-1]["start"] + idx[-1]["frames"]
ei_lines = set(_ei_lines())
self.ei_lines = ei_lines
pcm = narration(idx)
self.pcm = pcm
# --- ハコブくん ---
self.hu = hk.Hakobu()
self.hf = hk.Hakobu(mode="full")
bb = self.hf.body.getbbox()
self.hf.w = self.hu.w
self.hf.h = round((bb[3] - bb[1]) * self.hu.w / (bb[2] - bb[0]))
self.hf._cache = {}
tline = [(o["start"], o["start"] + o["frames"], o["start"] + o["speech"], o.get("face", "base"),
bool(o["text"]) and o["i"] not in ei_lines) for o in idx]
mp = pcm.astype(np.float32) / 32768.0
speak = np.zeros(self.total, bool)
eispk = np.zeros(self.total, bool)
for o in idx:
a0, a1 = int(o["start"] / FPS * SR), int((o["start"] + o["frames"]) / FPS * SR)
if o["i"] in ei_lines:
mp[a0:a1] = 0
eispk[o["start"]:o["start"] + o["speech"]] = True
elif o["text"]:
speak[o["start"]:o["start"] + o["speech"]] = True
self.params = self.hu.animate(tline, mp)
for t0, t1, f in sc.HK_FACE: # 文の途中で替える表情
ex = hk.EXPR[f]
for i in range(max(0, int(t0 * FPS)), min(self.total, int(t1 * FPS))):
eyes, brows, mouth, ring, dy = self.params[i]
if not speak[i]:
mouth = ex["rest"]
elif mouth == "closed" and "closed" in ex:
mouth = ex["closed"]
self.params[i] = (ex["eyes"], ex["brows"], mouth, ring, dy)
# --- エイイチ: 声に合わせて 4px（ドット1つ）弾む ---
ei_pcm = pcm.astype(np.float32) / 32768.0
hop = SR // FPS
rms = np.array([np.sqrt(np.mean(ei_pcm[i * hop:(i + 1) * hop] ** 2)) if eispk[i] else 0.0 for i in range(self.total)])
ref = np.percentile(rms[rms > 0.01], 85) if (rms > 0.01).any() else 1.0
bob = np.zeros(self.total, int)
for i in range(0, self.total, 3):
v = -4 if rms[i:i + 3].mean() > 0.45 * ref else 0
bob[i:i + 3] = v
self.bob = bob
self.ei = Eiichi()
self.subs = {}
def ei_state(self, f):
t = f / FPS
pose = 1
for a, p in self.sc.EI_POSE:
if t >= a:
pose = p
place = "normal"
for a, b, k in self.sc.EI_PLACE:
if a <= t < b:
place, pa, pb = k, a, b
x, y, scale, glow, crop = 1490, 680, 4, False, 1.0
if place == "hide":
return None
if place == "big":
scale, glow, x, y = 5, True, 1920 - 60 - 500, 1080 - 500
elif place == "buried":
scale, crop = 3, 0.62
x, y = 840, 870 - int(300 * 0.62) # 紙の山の真ん中の黒い穴から顔を出す（字幕より上）
elif place == "slide":
pose = 3
go = ease((t - pa) / 0.7)
back = ease((t - (pb - 0.5)) / 0.5)
x = int(1490 + (520 - 1490) * go + (1490 - 520) * back)
x -= x % 4
y += self.bob[f] if place in ("normal", "slide") else 0
return pose, x, y, scale, glow, crop
def hk_state(self, f):
"""→ (体, dx, dy, 回す角度, 回す中心) 体は upper / full"""
t = f / FPS
for m in self.sc.HK_MOVES:
u = t - m["t0"]
if m["kind"] == "jump" and 0 <= u < m["dur"]:
y0 = self.hu.y
y_peak = H - self.hf.h - 30
if u < 0.32:
k = np.sin(u / 0.32 * np.pi / 2) * 1.12
elif u < 0.5:
k = 1.12 - 0.12 * ease((u - 0.32) / 0.18)
elif u < m["dur"] - 0.45:
k = 1.0 + 0.02 * np.sin((u - 0.5) * 6)
else:
k = 1.0 - ease((u - (m["dur"] - 0.45)) / 0.45)
ytop = y0 + (y_peak - y0) * k
rot = -7 * min(1, k) + 2.5 * np.sin(u * 5) * min(1, k)
return "full", 30 * min(1, k), ytop - self.hf.y, rot, (self.hf.w / 2, self.hf.h)
if m["kind"] == "reflex" and 0 <= u < m["dur"]:
e = ease(u / 0.1) * (1 - ease((u - 0.25) / (m["dur"] - 0.25)))
# 上半身の絵を回すと、下の切り口が斜めに見える（2026-11-02 エイイチの指摘）。全身の絵を、上半身と同じ位置に置いて回す
return "full", -24 * e, self.hu.y - self.hf.y, -16 * e, (0, self.hf.h)
if m["kind"] == "fall": # 全身を出してから、前のめりに倒れる（体が切れないように。2026-11-02 エイイチの指摘）
y0 = self.hu.y
ys = H - 14 - self.hf.h # 全身が見える位置（足もとが画面の下から14px）
piv = (self.hf.w * 0.95, self.hf.h)
t1 = m["t1"]
if 0 <= u < 0.18:
k = ease(u / 0.18)
return "full", 0, y0 + (ys - y0) * k - self.hf.y, -6 * k, piv
if 0.18 <= u < 0.5:
r = -6 + 98 * ease((u - 0.18) / 0.32)
return "full", 0, ys - self.hf.y, r, piv
if u >= 0.5 and t < t1:
r = 88 + 4 * np.exp(-(u - 0.5) * 8) * np.cos((u - 0.5) * 28)
return "full", 0, ys - self.hf.y, r, piv
if t1 <= t < t1 + 0.45:
return "full", 0, ys - self.hf.y, 88 * (1 - ease((t - t1) / 0.45)), piv
if t1 + 0.45 <= t < t1 + 0.75:
k = ease((t - t1 - 0.45) / 0.3)
return "full", 0, ys + (y0 - ys) * k - self.hf.y, 0, piv
return "upper", 0, 0, 0, None
def hk_alpha(self, f):
t = f / FPS
for a, b, al in self.sc.HK_ALPHA:
if a <= t < b:
return al
return 1.0
def sub_at(self, f):
for o in self.idx:
if o["start"] <= f < o["start"] + o["frames"]:
if not o["text"]:
return None
r = f - o["start"]
for k, a, b in o["parts"]:
if a <= r < b:
key = (o["i"], k if k is not None else 0)
if key not in self.subs:
self.subs[key] = sub_layer(f"png/sub/{key[0]:03d}_{key[1]}.png")
return self.subs[key]
return None
def draw(self, bg, f):
"""bg: H×W×3 uint8（書き換える）"""
st = self.ei_state(f)
if st:
pose, x, y, scale, glow, crop = st
spr, off = self.ei.sprite(pose, scale, glow, crop)
blend(bg, spr, x + off, y + off)
body, dx, dy, rot, pivot = self.hk_state(f)
hh = self.hu if body == "upper" else self.hf
img = hh.frame(self.params[f])
x, y = hh.x + int(round(dx)), hh.y + int(round(dy))
if rot:
img, (ox, oy) = rotate_about(img, rot, pivot)
x, y = x + ox, y + oy
blend(bg, np.asarray(img).astype(np.float32), x, y, self.hk_alpha(f))
s = self.sub_at(f)
if s is not None:
blend(bg, s[0], s[1], s[2])
return bg
def _ei_lines():
import importlib.util as ilu
sp = ilu.spec_from_file_location("_narr_here", os.path.join(os.getcwd(), "narration.py"))
nm = ilu.module_from_spec(sp); sp.loader.exec_module(nm)
return list(getattr(nm, "EI_LINES", []))
def prepare(sc, blocks, pg):
os.makedirs("png/ovl", exist_ok=True); os.makedirs("blocks", exist_ok=True)
for k, h in sc.OVERLAYS.items():
pg.set_content(h); pg.wait_for_timeout(80)
pg.screenshot(path=f"png/ovl/{k}.png", omit_background=True)
last_i = blocks[-1]["sents"][-1]
make_subs(sc.tl.idx[:last_i + 1], pg)
def block_frames(sc, b):
idx = sc.tl.idx
a = idx[b["sents"][0]]["start"]
e = idx[b["sents"][-1]]["start"] + idx[b["sents"][-1]]["frames"]
return a, e - a
def main():
args = sys.argv[1:]
cmd = args[0] if args else "sheet"
import importlib
sc = importlib.import_module("scenes_kabe")
from playwright.sync_api import sync_playwright
blocks = sc.BLOCKS
if "--first-blocks" in args:
blocks = blocks[:int(args[args.index("--first-blocks") + 1])]
names = [a for a in args[1:] if not a.startswith("--") and not a.isdigit()]
with sync_playwright() as p:
br = p.chromium.launch()
pg = br.new_page(viewport={"width": W, "height": H})
prepare(sc, blocks, pg)
if cmd == "sheet":
cast = Cast(sc, blocks[-1]["sents"][-1])
os.makedirs("png/sheet", exist_ok=True)
shots = []
for b in blocks:
if names and b["name"] not in names:
continue
a, n = block_frames(sc, b)
ts = sc.SHEET.get(b["name"], [n / FPS - 0.4])
for k, t in enumerate(ts):
f = min(a + int(t * FPS), a + n - 1)
path = f"png/sheet/{b['name']}_{k}.jpg"
if "video" in b:
src, ss, ovl = b["video"]
dur = float(subprocess.check_output(["ffprobe", "-v", "quiet", "-show_entries", "format=duration", "-of", "csv=p=0", src]))
subprocess.run(["ffmpeg", "-v", "error", "-y", "-ss", f"{(ss + t) % (dur - 0.2):.2f}", "-i", src, "-frames:v", "1",
"-vf", "scale=1920:1080:force_original_aspect_ratio=increase,crop=1920:1080", path], check=True)
im = Image.open(path).convert("RGBA")
im.alpha_composite(Image.open(ovl).convert("RGBA"))
im.convert("RGB").save(path)
else:
still(b, t, pg, path)
bg = np.array(Image.open(path).convert("RGB"))
Image.fromarray(cast.draw(bg, f)).save(path, quality=88)
shots.append((b["name"], k, f / FPS, path))
# 一覧
cols, tw = 4, 480
th = tw * 9 // 16
rows = (len(shots) + cols - 1) // cols
from PIL import ImageDraw, ImageFont
fnt = ImageFont.truetype("/usr/share/fonts/opentype/noto/NotoSansCJK-Bold.ttc", 22)
sheet = Image.new("RGB", (cols * tw, rows * (th + 34)), (240, 238, 232))
d = ImageDraw.Draw(sheet)
for j, (nm, k, t, path) in enumerate(shots):
x, y = (j % cols) * tw, (j // cols) * (th + 34)
sheet.paste(Image.open(path).resize((tw - 6, th - 4)), (x + 3, y + 32))
d.text((x + 8, y + 4), f"{nm} {int(t // 60)}:{t % 60:04.1f}", font=fnt, fill=(30, 30, 30))
sheet.save("一覧_場面.jpg", quality=85)
print("一覧_場面.jpg", len(shots))
if cmd in ("render", "build"):
for b in blocks:
if names and b["name"] not in names:
continue
a, n = block_frames(sc, b)
out = f"blocks/{b['name']}_{block_hash(b, n)}.mp4"
b["mp4"] = out
if not os.path.exists(out):
render_block(b, n, out, pg)
br.close()
if cmd != "build":
return
for b in blocks:
a, n = block_frames(sc, b)
b["mp4"] = f"blocks/{b['name']}_{block_hash(b, n)}.mp4"
last_i = blocks[-1]["sents"][-1]
cast = Cast(sc, last_i)
idx = cast.idx
write_wav("narration.wav", cast.pcm)
audio_in = "narration.wav"
if os.path.exists("sound.py"):
from audio_mix import mix
audio_in = mix(idx, cast.pcm)
total = cast.total
enc = subprocess.Popen(["ffmpeg", "-v", "error", "-y", "-f", "rawvideo", "-pix_fmt", "rgb24", "-s", f"{W}x{H}", "-r", str(FPS), "-i", "-",
"-i", audio_in, "-map", "0:v", "-map", "1:a", "-c:v", "libx264", "-preset", "medium", "-crf", "20", "-pix_fmt", "yuv420p",
"-af", "loudnorm=I=-16:TP=-1.5:LRA=11", "-c:a", "aac", "-b:a", "192k", "-ar", "48000",
"-movflags", "+faststart", "-frames:v", str(total), "out.mp4"], stdin=subprocess.PIPE)
f = 0
for b in blocks:
a, n = block_frames(sc, b)
dec = subprocess.Popen(["ffmpeg", "-v", "error", "-i", b["mp4"], "-f", "rawvideo", "-pix_fmt", "rgb24", "-"], stdout=subprocess.PIPE)
for k in range(n):
raw = dec.stdout.read(W * H * 3)
if len(raw) < W * H * 3:
sys.exit(f"紙の映像が短い: {b['name']} {k}/{n}")
bg = np.frombuffer(raw, np.uint8).reshape(H, W, 3).copy()
enc.stdin.write(cast.draw(bg, a + k).tobytes())
f += 1
if f % 900 == 0:
print(f" {f}/{total}", flush=True)
dec.stdout.close(); dec.wait()
enc.stdin.close()
if enc.wait():
sys.exit("書き出しに失敗")
# 字幕 SRT
def ts(fr):
ms = round(fr / FPS * 1000); h_, ms = divmod(ms, 3600000); m_, ms = divmod(ms, 60000); s_, ms = divmod(ms, 1000)
return f"{h_:02d}:{m_:02d}:{s_:02d},{ms:03d}"
srt, k = [], 1
for o in idx:
if not o["text"]:
continue
pages_ = sub_pages(o["text"])
for (pk, a_, b_), ptxt in zip(o["parts"], pages_):
end = min(b_, o["speech"]) if pk == len(pages_) - 1 else b_
srt.append(f"{k}\n{ts(o['start'] + a_)} --> {ts(o['start'] + end)}\n{ptxt}\n"); k += 1
open("out.srt", "w", encoding="utf-8").write("\n".join(srt))
json.dump(idx, open("timeline.json", "w"), ensure_ascii=False)
for crf in (29, 32, 35):
subprocess.run(["ffmpeg", "-v", "error", "-y", "-i", "out.mp4", "-c:v", "libx264", "-preset", "medium", "-crf", str(crf),
"-c:a", "aac", "-b:a", "112k", "-movflags", "+faststart", "out_確認用.mp4"], check=True)
mb = os.path.getsize("out_確認用.mp4") / 1e6
if mb <= 30:
break
print(f"frames {total} = {total / FPS:.1f}s 確認用 {mb:.1f}MB (crf {crf})")
if __name__ == "__main__":
main()
```
### kabe_short.py（縦長ショートの版2）
```python
# -*- coding: utf-8 -*-
"""考察回「壁打ちの型」の縦長ショート（1080×1920）・2版。
2026-11-02 エイイチの指摘で作り直した:
「おれが上半身で、ハコブくんが全身は変」→ エイイチも全身のドット絵（src/ei/エイイチ_全身_*_170px_v1.png・Gemini で写真の服装から作った）
「この始まりだとすぐ離脱しそう」「宇宙・紙の山・引退式の背景を使って、ハコブくんが大きく動く所から始める」
→ 最初の数秒は、画面いっぱいの絵の前で、大きなハコブくん（高さ900）が動く。本題に入ったら、絵を暗くして紙（考察ノート）を出し、
ハコブくんは左下の全身（高さ400）に戻る
画面の割り振りは video_rules.md 9-2: 0〜150 何も置かない／170〜440 見出し／480〜1130 紙／左下 ハコブくん・右下 エイイチ（足もと y=1640）／字幕は2人の間の上
使い方: python3 kabe_short.py [番号 ...]
"""
import os, sys, json, wave, subprocess, html as _html
import numpy as np
from PIL import Image
HERE = os.path.dirname(os.path.abspath(__file__))
sys.path.insert(0, HERE)
import kabe_kit as K
from kabe_kit import ink, land, card, counter_js, shake_js, ease, blend, rotate_about, hk, FPS, SR
sys.path.insert(1, K.BASE)
import short_kit as SK
from scene_kit import LOGO
W, H = 1080, 1920
SANS = "font-family:'Noto Sans CJK JP',sans-serif;"
FOOT = 1880 # 2人の足もと（下の方まで使ってよい・2026-11-02 エイイチ）
SUB_MID = 1150 # 字幕の真ん中
SUB_CX, SUB_MAXC = 540, 11 # 字幕は画面の真ん中（2026-11-02 エイイチの指摘で 470 から直した）・1行11字
EI_SCALE, EI_X = 4.5, None # エイイチの全身（170px の4.5倍＝高さ765・右端に寄せる）。身長はエイイチ 1700mm・ハコブくん 1150mm（2026-11-02 エイイチ）
HK_SMALL_H, HK_X = 515, 10 # ふだんのハコブくん（全身）。765 × 1150/1700
HK_BIG_H = 900 # 最初の数秒の大きなハコブくん
EI_POSES = {1: "立つ", 2: "腕組み", 3: "驚く", 4: "笑う"}
VCSS = """
html, body { width:1080px; height:1920px; }
.f { width:1080px; height:1920px; }
.vhead { position:absolute; left:60px; right:60px; top:170px; height:270px; display:flex; flex-direction:column; justify-content:center; align-items:center; text-align:center; z-index:5; }
.vhead img { height:54px; align-self:flex-start; margin-bottom:18px; filter:invert(1) drop-shadow(0 0 6px rgba(0,0,0,.8)); opacity:.9; }
.vhead .q { font-family:'Noto Sans CJK JP',sans-serif; font-size:84px; font-weight:900; line-height:1.2; color:#f5f3ee; text-shadow:0 0 18px rgba(0,0,0,.9), 0 4px 0 #141311; }
.vpaper { position:absolute; left:100px; top:480px; width:880px; height:590px; background:#f5f3ee; box-shadow:0 30px 60px rgba(0,0,0,.45); opacity:0; }
.vpaper .card { font-size:40px; }
.vpaper .card small { font-size:24px; }
.vstack { position:absolute; left:0; right:0; top:40px; bottom:40px; display:flex; flex-direction:column; align-items:center; justify-content:center; gap:22px; }
.bgimg { position:absolute; inset:0; background-repeat:no-repeat; background-color:#fff; }
@keyframes veil { to { opacity:.74; } }
@keyframes pIn { from { opacity:0; transform:translateY(60px) rotate(-1.5deg); } to { opacity:1; transform:none; } }
.vbig { position:absolute; left:0; right:0; text-align:center; font-family:'Noto Sans CJK JP',sans-serif; font-weight:900; color:#fff; line-height:1.15;
-webkit-text-stroke:5px #141311; paint-order:stroke fill; text-shadow:0 10px 0 #c4472a; z-index:6; }
.vbig span { display:inline-block; opacity:0; }
"""
def vpage(head, body):
return K.page(body + f'
'
def grp(inner, t_in=0.0, t_out=None):
an = f"show .01s linear {t_in:.3f}s forwards" + (f", hide .2s linear {t_out:.3f}s forwards" if t_out is not None else "")
return f'
{inner}
'
def line(text, t, size=40, weight=700, step=0.05, style=""):
return f'
{ink(text, t, step)}
'
def words(rows, t0, t_out, top, size=130, gap=0.32):
"""大きな一言を1行ずつ叩きつける。rows は [(文字, 色)]"""
out = []
for k, (txt, col) in enumerate(rows):
out.append(f'{txt}')
return f'
', T(4, 0.05)), tp)
+ shake_js([(T(1, 0.4), 0.25)])))
# segs: (本編の文番号, 後ろの間[, 頭を切る秒, 字幕の文])
# big: 大きなハコブくんの出る間（end の文の頭で、左下の大きさへ）。kind: rise（下から飛び出す）/ hop（跳ねて喜ぶ）
SHORTS = [
dict(no=1, name="遅っ", fn=short_uchu, lead=0.5,
segs=[(18, 0.3), (19, 0.4), (20, 0.5), (21, 3.0), (22, 0.3), (23, 0.5), (25, 1.2)],
big=dict(kind="rise", end=1),
face=[(-1, 0.0, 6.5, "surprise"), (1, 0.0, 1.6, "trouble"), (3, 0.0, 5.5, "think"), (4, 0.0, 1.0, "smile"), (5, 0.0, 1.4, "trouble")],
pose=[(-1, 0.0, 3), (1, 0.0, 2), (2, 0.5, 1), (5, 0.0, 2), (6, 0.0, 3)], ei_rise=None,
se=[(-1, 0.0, "se_whoosh", -8), (0, 0.3, "se_pop", -10), (0, 1.2, "se_pop", -10), (0, 2.1, "se_pop", -8), (5, 0.0, "se_pop", -6)]
+ [(3, 0.2 + k, "se_tick", -16) for k in range(5)]),
dict(no=2, name="捨てられない人", fn=short_kami, lead=0.4,
segs=[(59, 0.5, 1.15, "エイイチは、記事を全部、全文で残しています。番号は、424番まで来ています。……捨てられない人ですか？"), (60, 0.4), (61, 0.6, 0.0, "情報がたくさんある｜だけじゃ、言葉や動きは良くならない。"), (63, 1.2)],
big=dict(kind="rise", end=1),
face=[(-1, 0.0, 9.5, "doubt"), (1, 0.0, 4.0, "surprise"), (3, 0.0, 4.5, "smile")],
pose=[(-1, 0.0, 3), (1, 0.0, 2), (2, 0.0, 1), (3, 0.0, 4)],
ei_rise=(0, 7.4), # 「捨てられない人ですか？」で、紙の山の下からエイイチが出てくる
se=[(-1, 0.0, "se_whoosh", -8), (0, 5.0, "se_pop", -6), (0, 7.4, "se_rise", -10), (0, 7.6, "se_pop", -8)]),
dict(no=3, name="勝手に引退させないで", fn=short_intai, lead=0.6,
segs=[(13, 1.0, 1.05, "エイイチの研究室の π0.5 は……引退ですね。お疲れさまでした。"), (14, 0.4), (12, 0.35), (15, 0.5), (36, 1.2)],
big=dict(kind="hop", end=1),
face=[(-1, 0.0, 7.0, "joy"), (1, 0.0, 2.0, "surprise"), (3, 0.0, 4.4, "trouble")],
pose=[(-1, 0.0, 1), (0, 3.0, 3), (1, 0.0, 2), (2, 0.0, 1), (3, 0.0, 2), (4, 0.0, 1)], ei_rise=None,
se=[(-1, 0.0, "se_rise", -8), (-1, 0.2, "se_pop", -8), (1, 0.1, "se_pop", -6), (2, 0.0, "se_whoosh", -12)]),
]
def vpages10(text):
"""字幕の枚（1行10字・3行まで）。short_kit.wrap の折り方で"""
out = []
for part in text.split("｜"):
if not part:
continue
ls = SK.wrap(part, SUB_MAXC)
if ls is None:
out.append(None); continue
while ls:
n = len(ls)
k = 3 if n == 3 or n >= 5 else min(n, 2)
out.append("\n".join(ls[:k])); ls = ls[k:]
return out
def build(sh):
idx_main = json.load(open("tts/index.json"))
ei_main = set(K._ei_lines())
lead = round(sh.get("lead", 0) * FPS)
segs, t = [], lead
for k, sg in enumerate(sh["segs"]):
i, pause = sg[0], sg[1]
trim = sg[2] if len(sg) > 2 else 0.0
o = idx_main[i]
dur = o["dur"] - trim
frames = round((dur + pause) * FPS)
segs.append(dict(k=k, i=i, trim=trim, text=sg[3] if len(sg) > 3 else o["text"], face=o.get("face", "base"),
start=t, speech=round(dur * FPS), frames=frames, ei=i in ei_main))
t += frames
total = t
if total / FPS > 30.5:
sys.exit(f"ショート{sh['no']} が {total / FPS:.1f}秒あります（30秒以内）")
Tf = lambda k, off=0.0: (segs[k]["start"] / FPS if k >= 0 else 0.0) + off
Sf = lambda k: segs[k]["speech"] / FPS
html = sh["fn"](Tf, Sf)
folder = f"ショート/{sh['no']:02d}_{sh['name']}"
os.makedirs(folder + "/png", exist_ok=True)
# --- 声 ---
pcm = np.zeros(total * SR // FPS + SR, np.int16)
for s in segs:
with wave.open(f"tts/{s['i']:03d}.wav") as w:
a = np.frombuffer(w.readframes(w.getnframes()), dtype=np.int16)
a = a[int(s["trim"] * SR):][:s["frames"] * SR // FPS]
if s["trim"]:
a = (a * np.minimum(1, np.arange(len(a)) / (0.02 * SR))).astype(np.int16) # 切った頭を短くなじませる
s0 = s["start"] * SR // FPS
pcm[s0:s0 + len(a)] = a
pcm = pcm[:total * SR // FPS]
# --- 字幕と紙の映像 ---
from playwright.sync_api import sync_playwright
SUBCSS = (f"*{{margin:0;padding:0}}html,body{{width:{W}px;height:{H}px;background:transparent}}svg{{position:absolute;left:0;top:0}}"
f".t{{font-family:'Noto Sans CJK JP',sans-serif;font-size:{SK.SUB_FS}px;font-weight:800;fill:#fff;stroke:#0D0D0D;stroke-width:12px;stroke-linejoin:round;paint-order:stroke fill}}")
subs = []
with sync_playwright() as p:
br = p.chromium.launch()
pg = br.new_page(viewport={"width": W, "height": H})
for s in segs:
pages_ = vpages10(s["text"])
n, tot, acc = len(pages_), sum(len(x or "") for x in pages_), 0
for pk, ptxt in enumerate(pages_):
if ptxt is None:
sys.exit(f"字幕が折れない: {s['text']}")
svg = ''.join(f'{_html.escape(ln)}' for ln in ptxt.split("\n"))
pg.set_content(f'')
pg.evaluate(f"""() => {{ const ts=[...document.querySelectorAll('text.t')]; const n=ts.length; let fs={SK.SUB_FS};
for (const t of ts) {{ const w=t.getComputedTextLength(); if (w>{SK.SUB_MAXW}) fs=Math.min(fs, Math.floor({SK.SUB_FS}*{SK.SUB_MAXW}/w)); }}
ts.forEach((t,k)=>{{ t.style.fontSize=fs+'px'; t.setAttribute('y', Math.round({SUB_MID}+(k-(n-1)/2)*{SK.SUB_LH}+fs*0.36)); }}); }}""")
path = f"{folder}/png/sub_{s['k']}_{pk}.png"
pg.screenshot(path=path, omit_background=True)
K.fill_holes(path)
a = s["start"] + round(s["speech"] * acc / max(1, tot))
acc += len(ptxt)
b = s["start"] + (s["frames"] if pk == n - 1 else round(s["speech"] * acc / max(1, tot)))
subs.append((a, b, K.sub_layer(path)))
mp4 = f"{folder}/_paper.mp4"
K.render_block(dict(name=sh["name"], html=html), total, mp4, pg)
br.close()
# --- ハコブくん: 全身。最初は大きく、本題で左下の大きさへ ---
def full(hh):
o = hk.Hakobu(mode="full")
bb = o.body.getbbox()
o.h = hh; o.w = round((bb[2] - bb[0]) * hh / (bb[3] - bb[1])); o._cache = {}
return o
hs, hb = full(HK_SMALL_H), full(HK_BIG_H)
tline = [(s["start"], s["start"] + s["frames"], s["start"] + s["speech"], s["face"], not s["ei"]) for s in segs]
if lead:
tline = [(0, lead, lead, "base", False)] + tline
mp = pcm.astype(np.float32) / 32768.0
speak = np.zeros(total, bool); eispk = np.zeros(total, bool)
for s in segs:
a0, a1 = s["start"] * SR // FPS, (s["start"] + s["frames"]) * SR // FPS
if s["ei"]:
mp[a0:a1] = 0; eispk[s["start"]:s["start"] + s["speech"]] = True
else:
speak[s["start"]:s["start"] + s["speech"]] = True
params = hs.animate(tline, mp)
for k, a, d, f in sh["face"]:
ex = hk.EXPR[f]
for i in range(int(Tf(k, a) * FPS), min(total, int((Tf(k, a) + d) * FPS))):
eyes, brows, mouth, ring, dy = params[i]
if not speak[i]:
mouth = ex["rest"]
params[i] = (ex["eyes"], ex["brows"], mouth, ring, dy)
t_end_big = Tf(sh["big"]["end"])
SHR = 0.5 # 小さくなるのにかける秒
def hk_place(t):
"""→ (高さ, 足もとの中心 x, 足もと y, 角度)"""
sx, sy = HK_X + hs.w / 2, FOOT
if t >= t_end_big + SHR:
return HK_SMALL_H, sx, sy, 0.0
if sh["big"]["kind"] == "rise":
k = 1.15 * np.sin(min(1, t / 0.45) * np.pi / 2) if t < 0.45 else 1.0 + 0.15 * np.exp(-(t - 0.45) * 6) * np.cos((t - 0.45) * 14)
bx, by = 540 + 40 * np.sin(t * 1.1), 1990 + (1900 - 1990) * k + 14 * np.sin(t * 2.2)
rot = 6 * np.sin(t * 1.3)
else: # hop: 跳ねて喜ぶ
ph = (t * 1.9) % 1.0
bx, by = 380 + 30 * np.sin(t * 2.0), 1900 - 160 * 4 * ph * (1 - ph)
rot = 8 * np.sin(t * 3.8)
if t < 0.3:
by = 1920 + (by - 1920) * ease(t / 0.3)
if t < t_end_big:
return HK_BIG_H, bx, by, rot
u = ease((t - t_end_big) / SHR)
return HK_BIG_H + (HK_SMALL_H - HK_BIG_H) * u, bx + (sx - bx) * u, by + (sy - by) * u, rot * (1 - u)
# --- エイイチ（全身・2倍） ---
sprites = {}
for k_, nm in EI_POSES.items():
im = Image.open(f"src/ei/エイイチ_全身_{nm}_170px_v2.png").convert("RGBA")
im = im.resize((round(im.width * EI_SCALE), round(im.height * EI_SCALE)), Image.NEAREST)
# 暗い背景に服が沈むので、うすい光の縁を付ける（本編の暗い場面と同じ）
from PIL import ImageFilter
pad = 20
g = Image.new("L", (im.width + 2 * pad, im.height + 2 * pad), 0)
g.paste(im.split()[3], (pad, pad))
g = g.filter(ImageFilter.GaussianBlur(7)).point(lambda v: int(min(255, v * 1.6) * 0.45))
lay = Image.merge("RGBA", (Image.new("L", g.size, 235), Image.new("L", g.size, 232), Image.new("L", g.size, 225), g))
lay.alpha_composite(im, (pad, pad))
sprites[k_] = np.asarray(lay.crop((pad, 0, pad + im.width, pad + im.height))).astype(np.float32)
ei_pcm = pcm.astype(np.float32) / 32768.0
hop = SR // FPS
rms = np.array([np.sqrt(np.mean(ei_pcm[i * hop:(i + 1) * hop] ** 2)) if eispk[i] else 0.0 for i in range(total)])
ref = np.percentile(rms[rms > 0.01], 85) if (rms > 0.01).any() else 1.0
bob = np.zeros(total, int)
for i in range(0, total, 3):
bob[i:i + 3] = -EI_SCALE if rms[i:i + 3].mean() > 0.45 * ref else 0
poses = sorted((Tf(k, a), p_) for k, a, p_ in sh["pose"])
def ei_dy(t):
"""ei_rise: その時刻まで画面の下に隠れていて、下から出てくる"""
if not sh.get("ei_rise"):
return 0
t0 = Tf(*sh["ei_rise"])
if t < t0:
return 760
return int(760 * (1 - ease((t - t0) / 0.35)))
# --- 音 ---
voice = pcm.astype(np.float32) / 32768.0
ref_v = np.percentile(np.abs(voice[voice != 0]), 90) if (voice != 0).any() else 0.1
raw = subprocess.run(["ffmpeg", "-v", "error", "-i", "sound/bgm_明るいB_2.mp3", "-ac", "1", "-ar", str(SR), "-f", "s16le", "-"], capture_output=True, check=True).stdout
bgm = np.frombuffer(raw, np.int16).astype(np.float32) / 32768.0
bgm = np.tile(bgm, int(np.ceil(len(voice) / len(bgm))) + 1)[:len(voice)]
bgm *= ref_v / (np.percentile(np.abs(bgm), 90) + 1e-6) * 10 ** (-24 / 20)
fade = np.ones(len(voice)); nfo = int(1.0 * SR); fade[-nfo:] = np.linspace(1, 0, nfo)
mix = voice + bgm * fade
for k, off, name, lv in sh["se"]:
raw = subprocess.run(["ffmpeg", "-v", "error", "-i", os.path.join(K.BASE, "sound", name + ".wav"), "-ac", "1", "-ar", str(SR), "-f", "s16le", "-"], capture_output=True, check=True).stdout
se = np.frombuffer(raw, np.int16).astype(np.float32) / 32768.0
se *= ref_v / (np.abs(se).max() + 1e-6) * 10 ** (lv / 20)
s0 = int(Tf(k, off) * SR)
n = min(len(se), len(mix) - s0)
if n > 0:
mix[s0:s0 + n] += se[:n]
mix *= 0.98 / max(0.98, np.abs(mix).max())
K.write_wav(f"{folder}/_mix.wav", (mix * 32767).astype(np.int16))
# --- 重ねる ---
out = f"{folder}/ショート{sh['no']}_{sh['name']}_版2.mp4"
enc = subprocess.Popen(["ffmpeg", "-v", "error", "-y", "-f", "rawvideo", "-pix_fmt", "rgb24", "-s", f"{W}x{H}", "-r", str(FPS), "-i", "-",
"-i", f"{folder}/_mix.wav", "-map", "0:v", "-map", "1:a", "-c:v", "libx264", "-preset", "medium", "-crf", "20", "-pix_fmt", "yuv420p",
"-af", "loudnorm=I=-14:TP=-1.5:LRA=11", "-c:a", "aac", "-b:a", "160k", "-ar", "48000", "-movflags", "+faststart",
"-frames:v", str(total), out], stdin=subprocess.PIPE)
dec = subprocess.Popen(["ffmpeg", "-v", "error", "-i", mp4, "-f", "rawvideo", "-pix_fmt", "rgb24", "-"], stdout=subprocess.PIPE)
for f in range(total):
t = f / FPS
bgf = np.frombuffer(dec.stdout.read(W * H * 3), np.uint8).reshape(H, W, 3).copy()
pose = 1
for a, p_ in poses:
if t >= a:
pose = p_
spr = sprites[pose]
blend(bgf, spr, W - spr.shape[1] - 12, FOOT - spr.shape[0] + bob[f] + ei_dy(t)) # エイイチは右端に寄せる
size, cx, fy, rot = hk_place(t)
img = (hb if size > HK_SMALL_H + 1 else hs).frame(params[f])
if abs(size - img.height) > 1:
img = img.resize((round(img.width * size / img.height), round(size)), Image.LANCZOS)
x, y = int(cx - img.width / 2), int(fy - img.height)
if abs(rot) > 0.2:
img, (ox, oy) = rotate_about(img, rot, (img.width / 2, img.height))
x, y = x + ox, y + oy
blend(bgf, np.asarray(img).astype(np.float32), x, y)
for a, b, (sl, sx_, sy_) in subs:
if a <= f < b:
blend(bgf, sl, sx_, sy_)
enc.stdin.write(bgf.tobytes())
dec.stdout.close(); dec.wait()
enc.stdin.close(); enc.wait()
print(f"{out} {total / FPS:.1f}秒")
return out
if __name__ == "__main__":
want = [int(a) for a in sys.argv[1:]] or [s["no"] for s in SHORTS]
for sh in SHORTS:
if sh["no"] in want:
build(sh)
```
### sound.py（BGM と章の効果音。効果音の時刻は scenes_kabe.py の SE）
```python
# 音の置き方（audio_mix.py が読む）。効果音の時刻は scenes_kabe.py の SE に書く（場面の動きと同じ所で決めるため）
# 考察の場面は静かに、脱線の場面は効果音を大きめに（考察回の型）
from scenes_kabe import SE as _SE
BGM = [
dict(file='sound/bgm_明るいB_2.mp3', start=0, end=None, level=-26, gap=-19, fade_in=0.6, fade_out=3.0), # いつもの曲（ElevenLabs Music）を、考察回なのでいつもより3dB小さく
]
SE = list(_SE)
CHAPTER_SE = dict(file='se_chapter_marimba', files={'title': 'se_chapter_glock'}, level=-10, offset=-0.3,
scenes=['title', 'bring', 'think', 'split', 'fail_ei', 'log', 'sum', 'outro'])
```
### gemini_bg.py（Gemini の画像の API で、脱線の絵・全身のドット絵・サムネイルの背景を作る）
```python
#!/usr/bin/env python3
"""考察回の脱線の場面の背景を Gemini の画像の API で作る（2026-11-02）
使い方: python3 gemini_bg.py <プロンプト.txt> <出力の頭> [札=参照画像のパス ...]
16:9。鍵は KIOXIA/gemini_api_key.txt から読むだけ（表示しない）。もとは 4コマの道具 gemini_gen_labeled.py"""
import sys, json, base64, time, urllib.request, urllib.error, mimetypes, os, datetime
MODEL = "gemini-3.1-flash-image"
SIZE = os.environ.get("BG_SIZE", "2K")
KEY_PATHS = [os.path.expanduser("~/mnt/KIOXIA/gemini_api_key.txt"), "/Volumes/KIOXIA/gemini_api_key.txt"]
key = next(open(p).read().strip() for p in KEY_PATHS if os.path.exists(p))
prompt_path, out_head, refs = sys.argv[1], sys.argv[2], sys.argv[3:]
parts = []
for r in refs:
label, path = r.split("=", 1)
parts.append({"text": f"Reference image — {label}:"})
parts.append({"inlineData": {"mimeType": mimetypes.guess_type(path)[0] or "image/png", "data": base64.b64encode(open(path, "rb").read()).decode()}})
parts.append({"text": open(prompt_path, encoding="utf-8").read()})
body = {"contents": [{"role": "user", "parts": parts}],
"generationConfig": {"responseModalities": ["IMAGE"], "imageConfig": {"aspectRatio": os.environ.get("BG_ASPECT", "16:9"), "imageSize": SIZE}}}
req = urllib.request.Request(f"https://generativelanguage.googleapis.com/v1beta/models/{MODEL}:generateContent",
data=json.dumps(body).encode(), headers={"Content-Type": "application/json", "x-goog-api-key": key})
t = time.time()
try:
res = json.load(urllib.request.urlopen(req, timeout=300))
except urllib.error.HTTPError as e:
print("失敗", e.code, e.read().decode()[:800]); sys.exit(1)
el = time.time() - t
stamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
img = None
for c in res.get("candidates", []):
for p in c.get("content", {}).get("parts", []):
if "inlineData" in p: img = p["inlineData"]
ext = ".jpg" if img and "jpeg" in img["mimeType"] else ".png"
base = f"{out_head}_{stamp}"
if img: open(base + ext, "wb").write(base64.b64decode(img["data"]))
for c in res.get("candidates", []):
for p in c.get("content", {}).get("parts", []): p.pop("inlineData", None)
json.dump({"model": MODEL, "size": SIZE, "elapsed_s": round(el, 2), "prompt": prompt_path, "refs": refs, "response": res}, open(base + ".json", "w", encoding="utf-8"), ensure_ascii=False, indent=1)
u = res.get("usageMetadata", {})
print(os.path.basename(base + ext) if img else "画像なし", f"{el:.1f}秒", u.get("promptTokenCount"), u.get("candidatesTokenCount"), res.get("candidates", [{}])[0].get("finishReason"))
```
### 脱線の絵のプロンプト：宇宙（宇宙_プロンプト.txt）
```text
Photorealistic wide cinematic photograph, 16:9 landscape.
A colossal AI data center floating in deep space: a huge ring-shaped station packed with rows of glowing server racks, cooling fins and tiny blinking status lights, seen at a dramatic angle.
A giant blue planet fills the upper right, a dense star field and a faint purple nebula behind.
The mood is awe-inspiring and slightly absurd, like a movie poster that went too far.
Leave generous empty starry space in the center and left of the frame so that large white text can be placed over it. One continuous seamless photograph — no bands, no stripes, no split panels, no hard horizontal edges.
No text, no letters, no numbers, no logos, no people, no astronauts.
```
### 脱線の絵のプロンプト：引退式（引退式_プロンプト.txt・ハコブくんの正面と横の絵を参考に渡した）
```text
Black and white Japanese manga illustration, one wide 16:9 panel, clean ink line art with screentone, the artwork fills the whole image with no panel border and no frame line.
USE THE ATTACHED REFERENCE for the robot "Hakobu" — match its body shape, screen face, wheels, line weight and screentone exactly. Hakobu is a hospital delivery robot: a rounded box body with a screen for a face. Hakobu has NO arms and NO hands.
Scene: a retirement ceremony on a hospital break-room table.
- On the table stands a small white 6-axis desktop robot arm. It holds a bouquet of flowers in its gripper, slightly drooping, as if bowing in thanks.
- Above it hangs a paper garland and a blank ribbon banner with NO letters on it. Confetti falls.
- Hakobu stands beside the table, deeply moved: its screen face shows big tear drops streaming down and a wobbly happy smile. Small sparkle marks around it.
- Comedic, heartwarming, deliberately over-the-top, like a farewell party for a coworker.
No text, no letters, no numbers, no speech bubbles anywhere in the image. No people.
```
### 脱線の絵のプロンプト：割れたコップ（コップ_プロンプト.txt）
```text
Photorealistic high-speed photograph, 16:9 landscape.
A clear drinking glass shattering on a polished pale hospital corridor floor, water splashing up in a frozen crown shape, glass shards suspended in mid-air, every droplet sharp.
Low camera angle close to the floor, shallow depth of field, the corridor softly blurred behind.
Cool clinical lighting with one warm accent light, dramatic slow-motion feel, a little comedic in its seriousness.
Keep the upper center area relatively plain so text can be placed over it.
No people, no robots, no text, no letters, no logos.
```
### 脱線の絵のプロンプト：紙の山（紙の山_プロンプト.txt）
```text
Photorealistic wide photograph, 16:9 landscape.
A small Japanese home office completely buried under mountains of printed paper documents: towers of stacked A4 sheets reaching the ceiling, leaning piles, sheets sliding off and scattered across the floor, binders and sticky notes everywhere. A desk with a laptop and a small white robot arm is barely visible under the paper.
In the lower center, leave a clear dark gap in the paper pile, as if someone has just been buried there and could pop their head out.
Warm evening desk-lamp light mixed with cool light from a window, slightly comedic exaggeration, like a magazine photo of a hoarder of documents.
Keep the upper center area calm so text can be placed over it.
No people, no faces, no hands, no readable text on any paper, no logos.
```
### 作り手の全身のドット絵のプロンプト1：立つ（f1.txt）
```text
Pixel art, FULL BODY from head to shoes, the SAME character as the first attached reference image (the pixel-art bust): same face, same black messy hair with bangs, same thick black rectangular glasses, same warm orange skin tone with reddish-brown shading, same chunky pixel size, same limited color palette, same black outlines, no anti-aliasing. Outfit exactly as the person in the second attached photo: a loose dark charcoal gray tailored jacket, a black crew-neck T-shirt, loose wide black trousers, and light gray-white chunky sneakers with lavender-blue accents. Standing, the whole figure centered and fully visible with a small margin, feet at the bottom. Plain flat pure white background. No text, no letters, no shadow on the ground.
Pose and expression: relaxed standing pose with both hands in the trouser pockets, a calm slight smile, facing the viewer.
```
### 作り手の全身のドット絵のプロンプト2：ツッコミの手つき（f2.txt・出てきた絵は腕組み）
```text
Pixel art, FULL BODY from head to shoes, the SAME character as the first attached reference image (the pixel-art bust): same face, same black messy hair with bangs, same thick black rectangular glasses, same warm orange skin tone with reddish-brown shading, same chunky pixel size, same limited color palette, same black outlines, no anti-aliasing. Outfit exactly as the person in the second attached photo: a loose dark charcoal gray tailored jacket, a black crew-neck T-shirt, loose wide black trousers, and light gray-white chunky sneakers with lavender-blue accents. Standing, the whole figure centered and fully visible with a small margin, feet at the bottom. Plain flat pure white background. No text, no letters, no shadow on the ground.
Pose and expression: a comedic retort (tsukkomi): body turned slightly to the left, one hand thrust out to the left with the back of the hand facing out as if lightly hitting someone, eyes half-closed in an unimpressed look, a wry lopsided smile.
```
### 作り手の全身のドット絵のプロンプト3：驚く（f3.txt）
```text
Pixel art, FULL BODY from head to shoes, the SAME character as the first attached reference image (the pixel-art bust): same face, same black messy hair with bangs, same thick black rectangular glasses, same warm orange skin tone with reddish-brown shading, same chunky pixel size, same limited color palette, same black outlines, no anti-aliasing. Outfit exactly as the person in the second attached photo: a loose dark charcoal gray tailored jacket, a black crew-neck T-shirt, loose wide black trousers, and light gray-white chunky sneakers with lavender-blue accents. Standing, the whole figure centered and fully visible with a small margin, feet at the bottom. Plain flat pure white background. No text, no letters, no shadow on the ground.
Pose and expression: leaning back in big surprise, eyes wide open behind the glasses, eyebrows raised high, mouth open in an O shape, both hands raised with palms open, one foot slightly lifted.
```
### 作り手の全身のドット絵のプロンプト4：笑う（f4.txt）
```text
Pixel art, FULL BODY from head to shoes, the SAME character as the first attached reference image (the pixel-art bust): same face, same black messy hair with bangs, same thick black rectangular glasses, same warm orange skin tone with reddish-brown shading, same chunky pixel size, same limited color palette, same black outlines, no anti-aliasing. Outfit exactly as the person in the second attached photo: a loose dark charcoal gray tailored jacket, a black crew-neck T-shirt, loose wide black trousers, and light gray-white chunky sneakers with lavender-blue accents. Standing, the whole figure centered and fully visible with a small margin, feet at the bottom. Plain flat pure white background. No text, no letters, no shadow on the ground.
Pose and expression: bursting out laughing, eyes squinted shut with joy, shoulders raised, one hand covering the mouth, the other hand on the stomach, leaning forward slightly.
```
### サムネイルの背景のプロンプト：最初の案・写真のような脳（脳と腕_プロンプト.txt）
```text
A photorealistic, cinematic photograph for a YouTube thumbnail background. On the right third of the frame: a realistic anatomical model of a human brain seen from the side, with the cerebrum (large upper part) lit in warm amber light and the cerebellum (small lower back part) glowing in soft cyan light, so the two parts are clearly different. Next to it, a small white desktop robot arm reaching toward it. Dark charcoal background with soft studio lighting and slight depth of field. The left two-thirds of the image and the bottom quarter must be calm, dark, empty space (for text added later). Seamless single photograph, no bands or seams. No text, no letters, no numbers, no logos, no people, no watermark. 16:9.
```
### サムネイルの背景のプロンプト：つるっとした立体（脳_光_プロンプト.txt）
```text
A sleek, minimal 3D render for a YouTube thumbnail background, in the style of a modern tech product illustration (smooth matte plastic, not realistic, no anatomical detail, nothing fleshy or wet). On the right third: a smooth, simplified brain shape made of soft rounded matte material, split into two clearly separate pieces with a small gap between them: the large upper piece (cerebrum) in warm amber orange, the small lower back piece (cerebellum) in bright cyan, each softly glowing. Next to it, a small white robot arm in the same smooth matte style. Dark charcoal background with soft studio light. The left two-thirds and the bottom quarter must be empty, calm, dark space for text added later. No text, no letters, no numbers, no logos, no people, no watermark. 16:9.
```
### サムネイルの背景のプロンプト：平面（脳_平面_プロンプト.txt）
```text
A clean, friendly flat graphic illustration for a YouTube thumbnail background (not photorealistic, no anatomical detail, no wet or fleshy texture). On the right third: a simple, stylized brain icon seen from the side, drawn with smooth rounded shapes and only a few soft curved lines, made of two clearly separated parts: the large upper part (cerebrum) in warm amber orange, the small lower back part (cerebellum) in bright cyan. Thin soft glow around the shape. Next to it, a simple white robot arm drawn in the same flat style. Deep charcoal navy background with a subtle gradient. The left two-thirds and the bottom quarter must be empty, calm, dark space for text added later. No text, no letters, no numbers, no logos, no people, no watermark. 16:9.
```
### サムネイルの背景のプロンプト：右下に大きく（脳_光2_プロンプト.txt）
```text
A sleek, minimal 3D render for a YouTube thumbnail background, in the style of a modern tech product illustration (smooth matte plastic, not realistic, no anatomical detail, nothing fleshy or wet). Composition: the subject is LARGE and placed in the LOWER-RIGHT part of the frame, occupying the right 45% of the width and reaching down close to the bottom edge. The subject: a smooth, simplified brain shape made of soft rounded matte material, split into two clearly separate pieces with a small gap between them: the large upper piece (cerebrum) in warm amber orange, the small lower back piece (cerebellum) in bright cyan, each softly glowing. A small white robot arm stands at the far right, its gripper pointing at the cyan piece, fully inside the frame. Dark charcoal background with soft studio light. The left 55% of the image must be empty, calm, dark space for text added later. No text, no letters, no numbers, no logos, no people, no watermark. 16:9.
```
### thumb_A.py（サムネイル A 案の文言と背景）
```python
# サムネイル A 案（2026-11-02）: 背景は Gemini で作った「脳の模型とロボットアーム」（サムネ素材/脳と腕.jpg）
PHOTO = 'サムネ素材/脳と腕.jpg'
PHOTO_POS_FULL = '50% 50%'
HAKOBU = True
KICKER = '考察回'
LINES = ['ロボットの頭脳は', '脳のように', '分かれる？']
LABEL = '考えるのに18.6秒'
VARIANTS = ['photo4']
SIZE_L1, SIZE_L2, SIZE_L3 = 120, 190, 150
TXT_TOP = 60
HK_H = 560
LABEL_BOTTOM = 250
```
### thumb_B.py（サムネイル B 案の文言と背景。この案を、背景を「右下に大きく」に替えて使った）
```python
# サムネイル B 案（2026-11-02）: タイトルの後ろ半分をそのまま大きく。大脳＝暖色、小脳＝水色（背景の模型の光と同じ色）
PHOTO = 'サムネ素材/脳と腕.jpg'
PHOTO_POS_FULL = '50% 50%'
HAKOBU = True
KICKER = 'ロボットの頭脳は、脳のように分かれる？'
LINES = ['大きなAIは大脳', '専用AIは小脳', '']
LABEL = '僕はどっち？'
VARIANTS = ['photo4']
SIZE_L1, SIZE_L2, SIZE_L3 = 150, 150, 10
TXT_TOP = 60
HK_H = 560
LABEL_BOTTOM = 250
```
### 概要欄の下書き（概要欄_下書き.md）
```markdown
ロボットを動かすAIが1回考えるのに、専用のAIは0.05秒、何にでも使える大きなAIは18.6秒。エイイチが持ち込んだ話に、案内役のハコブくんが付き合う「考察回」です。答えを伝える回ではなく、考えている途中をそのまま見せます。人の脳が大脳・小脳・反射に分かれているように、ロボットのAIも分かれていくのか。速い手を育てるのは何か。
詳しい数字と考えた道すじは、サイトの記事にまとめています。
記事：ロボットのAIは脳のように分業する ― 小脳を育てるのは失敗の手直し https://pixelstomotors.com/record/robot-ai-divides-like-brain
▼目次
0:00 0.05秒と、18.6秒
{t:6} この番組について
{t:10} 持ち込み：ロボット専用の頭脳は、いらなくなる？
{t:21} やり取り1：遅い頭
{t:36} 人の脳も、分かれている
{t:49} やり取り2：失敗の手直し
{t:59} やり取り3：記憶の整理
{t:66} まとめ
{t:69} 残った問い
{t:74} エンディング
▼数字の出どころ
・0.05秒・18.6秒・11〜19秒（使い方による違い）：NVIDIA「Astranomy」（2026年10月14日）。π0.5 は 46.96ミリ秒、OpenAI の GPT-6 Astra は使い方によって 11.50秒〜18.56秒 https://research.nvidia.com/labs/srl/projects/astranomy/
・370回ほど：18.6秒 ÷ 0.05秒 で出した目安です
・海外の投資家の書き込み：Robonomics（Freda Duan）「Robotics' New Era After Astra」 https://robonomics.substack.com/p/robotics-new-era-after-astra ／ X の投稿 https://x.com/fredaduan/status/2116727410972590380
・イーロン・マスクさんの投稿（"~100% of SI will be deep space compute. Obviously." を訳しました） https://x.com/elonmusk/status/2114772074438988256
・小脳の神経細胞が、大脳の表面（大脳皮質）の4倍ほど：Suzana Herculano-Houzel の研究 https://www.frontiersin.org/journals/neuroanatomy/articles/10.3389/fnana.2010.00012/full ／ インタビュー https://intelligence.org/2014/04/22/suzana-herculano-houzel/
▼この動画について
・エイイチのセリフは、記事を書く前のやり取りで本人が話した言葉を、話し言葉に整えたものです。短い相づちやツッコミの一部は、本人が了承した案です
・「大きなAIは、賢くなるほど考える時間が長くなる方向に進んでいる」は、ハコブくん（Claude）の見立てです
・手直しの区間を分けて学ばせる方法は、まだ試していません。続きは机の上の実験で確かめます
▼この番組について
「フィジカルAI解説｜Pixels to Motors」は、フィジカルAIのニュースや実機の記録から、現場で使える形を考える番組です。
エイイチは、本業のデザイナーをしながら、ロボティクスの研究と開発をしています。その中で分かったことと、そこから考えた見立てを、案内役のハコブくんがお伝えします。
・ナレーションは AI 音声（ElevenLabs）です。話しているのは Pixels to Motors のキャラクター、ハコブくんです
・エイイチの声も AI 音声で、本人の声ではありません
・引退式・宇宙・割れたコップ・紙の山の絵と、サムネイルの背景（脳の模型とロボットアーム）は、生成AI（Gemini）で作成しています
・実機の映像はエイイチが撮ったものです
▼ロボットの開発・導入のご相談
メール：pixelstomotors@adaptdesign.jp
▼リンク
Webサイト：https://pixelstomotors.com
4コママンガ『ハコブくん』：https://pixelstomotors.com/manga-archive
GitHub：https://github.com/PixelstoMotors
X：https://x.com/adaptdesignjp
LinkedIn：https://www.linkedin.com/in/eiichi-shimazu
チャンネル登録・高評価もよろしくお願いします。コメントは、エイイチが全部読んでいます。
#フィジカルAI #ロボット #AI
```
### ショートのタイトルと説明文（ショート_タイトルと説明文.txt）
```text
ショートのタイトルと説明文（本編「ロボットの頭脳は、脳のように分かれるのか｜大きなAIは大脳、専用AIは小脳」から作った3本・版2）
関連動画には本編を設定してください。
━━━━━━━━━━━━━━━━
ショート1_遅っ（24.3秒）
タイトル：
深宇宙で計算？ その前に、考えるのに18.6秒 #Shorts
説明文：
「超知能のほぼ100%は、深宇宙で計算される」。話がでかい。地上のロボットでは、何にでも使える大きなAIが1回考えるのに18.6秒。その間に、ロボット専用のAI（π0.5）は動きを370回ほど直せます。
数字の出どころ：NVIDIA「Astranomy」（2026年10月14日）・イーロン・マスクさんの X の投稿。370回は 18.6秒 ÷ 0.05秒 の目安です。
宇宙の絵は生成AI（Gemini）で作成しました。声は AI 音声です（エイイチの声も AI 音声で、本人の声ではありません）。
本編：ロボットの頭脳は、脳のように分かれるのか
#フィジカルAI #ロボット #AI
━━━━━━━━━━━━━━━━
ショート2_捨てられない人（24.0秒）
タイトル：
記事424本を全部残す人は、捨てられない人？ #Shorts
説明文：
記事を全部、全文で残しているのは、捨てられないからではなく、全部の積み重ねが「基幹情報」だから。ただ、情報が多いだけでは言葉や動きは良くならない。脳も、その日の出来事をいったん海馬に残し、大事なものを大脳へ移していきます。
紙の山の絵は生成AI（Gemini）で作成しました。声は AI 音声です（エイイチの声も AI 音声で、本人の声ではありません）。
本編：ロボットの頭脳は、脳のように分かれるのか
#フィジカルAI #ロボット #AI
━━━━━━━━━━━━━━━━
ショート3_勝手に引退させないで（26.8秒）
タイトル：
π0.5 は引退？「勝手に引退させないで」 #Shorts
説明文：
何にでも使える大きなAIが、追加の練習なしでブロックを器に入れた。では、ロボット専用のAIは引退？ 答えは、組み合わせ。遅く考える頭と、速く直す手に分けます。
引退式の絵は生成AI（Gemini）で作成しました。声は AI 音声です（エイイチの声も AI 音声で、本人の声ではありません）。
本編：ロボットの頭脳は、脳のように分かれるのか
#フィジカルAI #ロボット #AI
```
### 書き出しの記録（kabe_kit.py build の出力・3版の本編）
```text
BGM 1本・効果音 35個 → mix.wav
89.67秒 文013 se_whoosh -10dB 「ということは、エイイチの研究室の」
94.17秒 文014 se_pop -8dB 「勝手に引退させないで。」
94.77秒 文014 se_pop -12dB 「勝手に引退させないで。」
110.25秒 文018 se_whoosh -10dB 「超知能の、ほぼ100%は、深宇宙」
118.57秒 文020 se_whoosh -14dB 「宇宙は、今日は置いておきます。」
130.53秒 文023 se_pop -8dB 「遅っ。」
145.93秒 文026 se_whoosh -10dB 「コップが手から滑り始めても、僕が」
183.03秒 文034 se_pop -12dB 「考えてみました。……半分、おっし」
210.77秒 文037 se_pop -8dB 「人も、小脳と大脳で役割が分かれて」
213.23秒 文038 se_rise -10dB 「……それです。」
229.08秒 文041 se_pop -10dB 「ちょっと実演します。」
284.20秒 文052 se_whoosh -12dB 「では僕も、小脳を鍛えるために、今」
284.55秒 文052 se_pop -6dB 「では僕も、小脳を鍛えるために、今」
286.70秒 文053 se_pop -12dB 「失敗だけ覚えちゃって、良くないね」
330.00秒 文059 se_whoosh -10dB 「記録といえば、エイイチは、記事を」
333.45秒 文060 se_whoosh -14dB 「LOGは、全部の積み重ね。基幹情」
377.05秒 文068 se_pop -14dB 「ちなみに、ボツになった3つは、全」
377.45秒 文068 se_pop -14dB 「ちなみに、ボツになった3つは、全」
377.85秒 文068 se_pop -14dB 「ちなみに、ボツになった3つは、全」
404.80秒 文072 se_ding -12dB 「今の返事、速かったね。」
122.90秒 文021 se_tick -16dB 「では、なぜ組み合わせになるのか。」
123.90秒 文021 se_tick -16dB 「では、なぜ組み合わせになるのか。」
124.90秒 文021 se_tick -16dB 「では、なぜ組み合わせになるのか。」
125.90秒 文021 se_tick -16dB 「では、なぜ組み合わせになるのか。」
126.90秒 文021 se_tick -16dB 「では、なぜ組み合わせになるのか。」
127.90秒 文021 se_tick -16dB 「では、なぜ組み合わせになるのか。」
36.80秒 文006 se_chapter_glock -10dB 「Pixels to Motors」
65.60秒 文010 se_chapter_marimba -10dB 「夜、エイイチから、画面写真が7枚」
122.30秒 文021 se_chapter_marimba -10dB 「では、なぜ組み合わせになるのか。」
131.60秒 文024 se_chapter_marimba -10dB 「大きなAIなら、ここで18.6秒」
200.07秒 文036 se_chapter_marimba -10dB 「だから、分けるしかありません。遅」
267.13秒 文050 se_chapter_marimba -10dB 「失敗を手直しするデータの積み重ね」
322.50秒 文059 se_chapter_marimba -10dB 「記録といえば、エイイチは、記事を」
365.27秒 文066 se_chapter_marimba -10dB 「ここまでを、1枚にまとめます。」
408.97秒 文074 se_chapter_marimba -10dB 「この動画が役に立ったら、高評価と」
900/12761
1800/12761
2700/12761
3600/12761
4500/12761
5400/12761
6300/12761
7200/12761
8100/12761
9000/12761
9900/12761
10800/12761
11700/12761
12600/12761
frames 12761 = 425.4s 確認用 19.3MB (crf 29)
```