#!/usr/bin/env python3 """YouTube thumbnail maker: 1280x720 JPGs (under 2 MB) from a video frame or an image, with huge outlined text, an optional cut-out subject (remove_background PNG) with a glow, a face / subject zoom, and an A/B set of layouts. Examples python thumb.py --source https://.../video.mp4 --pick-best --text "I TRIED IT FOR 30 DAYS" --variants 4 python thumb.py --source frame.jpg --subject cutout.png --text "DON'T BUY THIS" --text-b "WORTH IT?" --accent "#FFD60A" python thumb.py --source shot.jpg --focus 0.7,0.35 --zoom 1.6 --text "WOW" # punch in on a face at 70% / 35% """ from __future__ import annotations import argparse import sys from pathlib import Path sys.path.insert(0, str(Path(__file__).resolve().parent)) import kmedia as K # noqa: E402 W, H = 1280, 720 LAYOUTS = ["text-left", "text-right", "text-top", "text-bottom"] SCHEMES = [((255, 255, 255), (0, 0, 0)), ((255, 214, 10), (0, 0, 0)), ((255, 255, 255), (220, 30, 60)), ((0, 0, 0), (255, 255, 255))] def sharpness(path: Path) -> float: from PIL import ImageFilter, ImageStat im = K.open_image(path).convert("L") im.thumbnail((480, 480)) return ImageStat.Stat(im.filter(ImageFilter.FIND_EDGES)).var[0] def background(src, focus, zoom: float): from PIL import ImageEnhance img = K.open_image(src).convert("RGB") if zoom > 1.0: cw, ch = img.width / zoom, img.height / zoom x0 = min(max(img.width * focus[0] - cw / 2, 0), img.width - cw) y0 = min(max(img.height * focus[1] - ch / 2, 0), img.height - ch) img = img.crop((int(x0), int(y0), int(x0 + cw), int(y0 + ch))) img = K.cover(img, W, H, focus) img = ImageEnhance.Contrast(img).enhance(1.12) img = ImageEnhance.Color(img).enhance(1.25) # thumbnails read better a bit punchier return img def compose(bg, subject, text: str, layout: str, scheme, accent, font, arrow: bool): from PIL import Image, ImageDraw, ImageFilter img = bg.convert("RGBA").copy() # darken behind the text side shade = Image.new("L", (W, H), 0) d = ImageDraw.Draw(shade) if layout == "text-left": for x in range(W // 2 + 120): d.line((x, 0, x, H), fill=int(170 * max(0, 1 - x / (W / 2 + 120)))) elif layout == "text-right": for x in range(W // 2 - 120, W): d.line((x, 0, x, H), fill=int(170 * max(0, (x - (W / 2 - 120)) / (W / 2 + 120)))) elif layout == "text-top": for y in range(H // 2): d.line((0, y, W, y), fill=int(160 * (1 - y / (H / 2)))) else: for y in range(H // 2, H): d.line((0, y, W, y), fill=int(160 * ((y - H / 2) / (H / 2)))) dark = Image.new("RGBA", (W, H), (0, 0, 0, 255)) dark.putalpha(shade) img.alpha_composite(dark) if subject is not None: s = subject.copy() s.thumbnail((int(W * 0.55), int(H * 1.0))) x = {"text-left": W - s.width - 10, "text-right": 10}.get(layout, (W - s.width) // 2) y = H - s.height halo = Image.new("RGBA", s.size, tuple(accent) + (0,)) halo.putalpha(s.getchannel("A").filter(ImageFilter.MaxFilter(9)).filter(ImageFilter.GaussianBlur(14))) img.alpha_composite(halo, (x, y)) img.alpha_composite(halo, (x, y)) img.alpha_composite(s, (x, y)) fill, stroke = scheme box = {"text-left": (40, 40, int(W * 0.56), H - 40), "text-right": (int(W * 0.44), 40, W - 40, H - 40), "text-top": (60, 24, W - 60, int(H * 0.45)), "text-bottom": (60, int(H * 0.55), W - 60, H - 24)}[layout] r = K.draw_block(img, text, box, fill=fill, max_size=200, min_size=40, kind="bold", font_override=font, stroke_ratio=0.085, stroke_fill=stroke, upper=True, shadow=True, align="left" if layout == "text-left" else ("right" if layout == "text-right" else "center"), max_lines=3) if arrow and layout in ("text-left", "text-right"): d = ImageDraw.Draw(img) bx = r["bbox"] y = min(H - 70, bx[3] + 40) if layout == "text-left": x0, x1 = bx[0] + 20, min(bx[2] + 120, int(W * 0.66)) d.line((x0, y, x1 - 30, y), fill=tuple(accent) + (255,), width=18) d.polygon([(x1, y), (x1 - 50, y - 36), (x1 - 50, y + 36)], fill=tuple(accent) + (255,)) else: x0, x1 = bx[2] - 20, max(bx[0] - 120, int(W * 0.34)) d.line((x1 + 30, y, x0, y), fill=tuple(accent) + (255,), width=18) d.polygon([(x1, y), (x1 + 50, y - 36), (x1 + 50, y + 36)], fill=tuple(accent) + (255,)) return img.convert("RGB"), r def save_under(img, dst: Path, limit=2 * 1024 * 1024) -> int: for q in (92, 88, 84, 80, 75, 70, 62): img.save(dst, quality=q, optimize=True, progressive=True) if dst.stat().st_size <= limit: return q return q def main() -> None: ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter) ap.add_argument("--source", required=True, help="video or image (URL or file)") ap.add_argument("--time", type=float, default=None, help="video: frame time in seconds") ap.add_argument("--pick-best", action="store_true", help="video: sample 12 frames and use the sharpest") ap.add_argument("--text", required=True, help="3-5 words read best") ap.add_argument("--text-b", default="", help="alternate text for the B variants") ap.add_argument("--subject", default="", help="transparent PNG cut-out (Kenerate remove_background result)") ap.add_argument("--focus", default="0.5,0.4", help="x,y 0..1 of the face / subject to keep in frame") ap.add_argument("--zoom", type=float, default=1.0, help="punch-in on --focus (1.0 = none, 1.5 = 150%%)") ap.add_argument("--accent", default="#FFD60A") ap.add_argument("--variants", type=int, default=4, help="how many layouts / colour schemes (1-8)") ap.add_argument("--no-arrow", action="store_true") K.add_font_args(ap) ap.add_argument("--name", default="thumb") ap.add_argument("--out-dir", default="kenerate-out/thumbnail-maker") ap.add_argument("--keep-temp", action="store_true") a = K.parse_args(ap) st = K.Status(4, "thumbnail-maker") K.require(status=st) tmp = K.workdir(a.keep_temp) out = Path(a.out_dir) out.mkdir(parents=True, exist_ok=True) focus = tuple(float(v) for v in a.focus.split(",")) st.step("Fetching inputs") src = K.fetch(a.source, tmp, "source", st) subject = K.open_image(K.fetch(a.subject, tmp, "subject", st)).convert("RGBA") if a.subject else None if subject is not None and subject.getchannel("A").getextrema()[0] == 255: st.warn("--subject has no transparency; run Kenerate remove_background on it first for a clean cut-out") st.step("Choosing the frame") frame = src if K.kind_of(str(src)) != "image": info = K.probe(src) D = info["duration"] if a.pick_best and D > 0: cands = [] for i in range(12): t = D * (0.05 + 0.9 * i / 11) p = K.extract_frame(src, t, tmp / f"cand_{i:02d}.jpg") cands.append((sharpness(p), t, p)) st.progress((i + 1) * 100 / 12, f"frame {i + 1}/12") score, t, frame = max(cands) st.info(f"sharpest frame at {t:.2f} s") else: t = a.time if a.time is not None else D * 0.3 frame = K.extract_frame(src, t, tmp / "frame.jpg") st.info(f"frame at {t:.2f} s") bg = background(frame, focus, a.zoom) n = max(1, min(8, a.variants)) st.step(f"Rendering {n} variant(s)") accent = K.hex_rgb(a.accent) files, rows = [], [] for i in range(n): layout = LAYOUTS[i % len(LAYOUTS)] if subject is not None and layout in ("text-top", "text-bottom") and i < 2: layout = "text-left" scheme = SCHEMES[(i // 2) % len(SCHEMES)] if i else SCHEMES[0] if i == 1: scheme = ((accent[0], accent[1], accent[2]), (0, 0, 0)) text = a.text_b if (a.text_b and i % 2 == 1) else a.text img, r = compose(bg, subject, text, layout, scheme, accent, a.font, not a.no_arrow) ab = "A" if (not a.text_b or i % 2 == 0) else "B" dst = out / f"{K.safe_name(a.name)}_{ab}{i + 1}_{layout}.jpg" q = save_under(img, dst) files.append(dst) rows.append({"file": str(dst.resolve()), "text": text, "layout": layout, "font_size": r["font_size"], "jpeg_quality": q, "bytes": dst.stat().st_size}) st.progress((i + 1) * 100 / n, dst.name) st.step("Small-size preview (how it looks in the feed)") from PIL import Image tiles = [] for p in files: im = Image.open(p).convert("RGB") tiles.append(im.resize((336, 189), Image.LANCZOS)) # 2x the 168x94 sidebar size sheet = Image.new("RGB", (len(tiles) * 352 + 16, 189 + 32), (15, 15, 15)) for i, tl in enumerate(tiles): sheet.paste(tl, (16 + i * 352, 16)) sp = out / f"{K.safe_name(a.name)}_feed_preview.jpg" sheet.save(sp, quality=90) K.finish({"thumbnails": rows, "feed_preview": str(sp.resolve()), "size": f"{W}x{H}", "elapsed_s": st.elapsed()}) if __name__ == "__main__": K.main_guard(main)