HK-hub/AgentSkills / bangumi-frames
bangumi-frames
VerifiedExtract and organize frames from a Bilibili video (bangumi episode, UP upload, or a local file) into scenery shots and per-character image groups, using anime-specific person detection + CCIP character-identity embeddings. Two modes — cluster everyone, or pull out one (or several) named characters via reference folders. Use when the user wants to collect, extract, or organize anime frames/screenshots by character or by scenery from a Bilibili video. Read-only download for personal viewing/analysis; uploads nothing.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 7.1 KB |
| references/modes.md | markdown | 4.5 KB |
| references/pipeline.md | markdown | 6.8 KB |
| scripts/bangumi_frames.py | script | 36.3 KB |
| scripts/remove_overlay.py | script | 5.3 KB |
scripts/remove_overlay.py
View on GitHub#!/usr/bin/env python3
"""Remove bilibili hard subtitles (bottom) and the corner watermark from anime
frames, then inpaint the holes with LaMa. Standalone test tool.
Subtitles are located with RapidOCR (robust to low-contrast white-on-light text
that pixel heuristics miss) and restricted to a bottom band so dialogue painted
into the scene is left alone. The semi-transparent bilibili logo is invisible to
OCR, so its fixed top-right corner box is masked directly. Holes are filled by
LaMa, which reconstructs structure (clothing lines, colour edges) instead of the
smeared band cv2.inpaint leaves behind. Frames with nothing to remove skip the
model entirely.
python3 remove_overlay.py IMG [IMG ...] --out-dir /tmp/clean [--debug]
python3 remove_overlay.py FRAMES_DIR --out-dir DIR # whole folder
LaMa runs on Apple MPS / CUDA when available (~2s/frame), else CPU (~50s).
"""
import argparse
import os
from pathlib import Path
# some LaMa ops have no MPS kernel; let them fall back to CPU silently
os.environ.setdefault("PYTORCH_ENABLE_MPS_FALLBACK", "1")
import warnings
warnings.filterwarnings("ignore", message="An output with one or more elements")
import cv2
import numpy as np
_OCR = None
_LAMA = None
def _ocr():
global _OCR
if _OCR is None:
from rapidocr_onnxruntime import RapidOCR
_OCR = RapidOCR()
return _OCR
def _lama():
global _LAMA
if _LAMA is None:
import torch
from simple_lama_inpainting import SimpleLama
dev = ("mps" if torch.backends.mps.is_available()
else "cuda" if torch.cuda.is_available() else "cpu")
_LAMA = SimpleLama(device=torch.device(dev))
print(f"[lama] device={dev}")
return _LAMA
def _poly_bounds(box):
xs = [p[0] for p in box]
ys = [p[1] for p in box]
return min(xs), min(ys), max(xs), max(ys)
def build_mask(im, sub_band, min_conf, pad, wm_h, wm_w, no_watermark):
h, w = im.shape[:2]
s = max(h, w) / 2000.0
pad = max(3, int(pad * s))
mask = np.zeros((h, w), np.uint8)
# --- subtitles via OCR, kept only inside the bottom band ---
y_min = int(h * (1 - sub_band))
res, _ = _ocr()(im)
boxes = []
for box, _txt, conf in (res or []):
if conf < min_conf:
continue
x0, y0, x1, y1 = _poly_bounds(box)
if (y0 + y1) / 2 < y_min: # not in the subtitle band
continue
if (x1 - x0) > 0.95 * w: # full-width bar = not a subtitle
continue
boxes.append((x0, y0, x1, y1))
cv2.rectangle(mask, (int(x0) - pad, int(y0) - pad),
(int(x1) + pad, int(y1) + pad), 255, -1)
# --- watermark: fixed top-right corner box (OCR can't see it) ---
if not no_watermark:
ch, cw = int(h * wm_h), int(w * wm_w)
mask[:ch, w - cw:] = 255
return mask, boxes
def inpaint(im, mask):
from PIL import Image
rgb = Image.fromarray(cv2.cvtColor(im, cv2.COLOR_BGR2RGB))
out = _lama()(rgb, Image.fromarray(mask))
out = cv2.cvtColor(np.array(out), cv2.COLOR_RGB2BGR)
# LaMa pads to a multiple of 8, so the result can be a few px larger
return out[:im.shape[0], :im.shape[1]]
def clean(path, out_dir, debug=False, **kw):
im = cv2.imread(str(path))
if im is None:
print(f" skip (unreadable): {path}")
return
mask, boxes = build_mask(im, **kw)
if mask.any():
res = inpaint(im, mask)
else:
res = im # nothing to remove
out = out_dir / path.name
cv2.imwrite(str(out), res)
print(f" {path.name}: {len(boxes)} subtitle box(es), "
f"masked {int(mask.sum() / 255)} px -> {out}")
if debug:
dbg = im.copy()
dbg[mask > 0] = (0, 0, 255)
cv2.imwrite(str(out_dir / f"dbg_{path.name}"), dbg)
def main():
ap = argparse.ArgumentParser()
ap.add_argument("images", nargs="+", type=Path,
help="image files, or a directory of *.jpg")
ap.add_argument("--out-dir", type=Path, default=Path("/tmp/overlay_clean"))
ap.add_argument("--sub-band", type=float, default=0.22,
help="bottom fraction where subtitles are accepted")
ap.add_argument("--min-conf", type=float, default=0.6,
help="min OCR confidence for a subtitle box")
ap.add_argument("--pad", type=int, default=8,
help="px padding around each OCR box (scaled to image)")
ap.add_argument("--wm-h", type=float, default=0.14,
help="top fraction of the watermark corner box")
ap.add_argument("--wm-w", type=float, default=0.18,
help="right fraction of the watermark corner box")
ap.add_argument("--no-watermark", action="store_true",
help="skip the corner box (subtitles only)")
ap.add_argument("--debug", action="store_true",
help="also write dbg_* overlays of the mask")
args = ap.parse_args()
args.out_dir.mkdir(parents=True, exist_ok=True)
paths = []
for p in args.images:
paths.extend(sorted(p.glob("*.jpg")) if p.is_dir() else [p])
kw = dict(sub_band=args.sub_band, min_conf=args.min_conf, pad=args.pad,
wm_h=args.wm_h, wm_w=args.wm_w, no_watermark=args.no_watermark)
for p in paths:
clean(p, args.out_dir, debug=args.debug, **kw)
if __name__ == "__main__":
main()
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack." | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |