Pascal Scholz
0b326e6e9c
Szenen aus der Figuren-Bibliothek: freigestellte Standbilder mit Posenwechsel, freigestellte Clips (WebM mit Alpha in /shared), Orte als vorberechnete Aquarell-SVGs, Untertitel statt Stimmen. Zeitplan aus den Sprechzeilen, damit ElevenLabs-Aufnahmen später nur die Dauer setzen. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
47 lines
2.6 KiB
Python
47 lines
2.6 KiB
Python
# Clips freistellen: python3 werkzeuge/clip_freistellen.py <arbeitsordner> <zielordner> clip1.mp4 ...
|
|
# Clip freistellen: Frames -> Maske -> gemeinsamer Zuschnitt -> WebM (VP9 mit Alpha)
|
|
import sys, os, subprocess, json, glob
|
|
import numpy as np
|
|
from PIL import Image, ImageFilter
|
|
from scipy import ndimage
|
|
from rembg import remove, new_session
|
|
s = new_session("isnet-general-use")
|
|
work, ziel = sys.argv[1], sys.argv[2]
|
|
os.makedirs(ziel, exist_ok=True)
|
|
for src in sys.argv[3:]:
|
|
name = os.path.basename(src)[:-4]
|
|
if os.path.exists(f"{ziel}/{name}.webm"): continue
|
|
d = f"{work}/{name}"; os.makedirs(d + "/roh", exist_ok=True); os.makedirs(d + "/frei", exist_ok=True)
|
|
subprocess.run(["ffmpeg", "-v", "error", "-y", "-i", src, f"{d}/roh/%04d.png"], check=True)
|
|
roh = sorted(glob.glob(f"{d}/roh/*.png")); boxen = []
|
|
for f in roh:
|
|
dst = f"{d}/frei/" + os.path.basename(f)
|
|
im = Image.open(f).convert("RGB")
|
|
if not os.path.exists(dst):
|
|
m = np.asarray(remove(im, session=s, only_mask=True)).astype(float)
|
|
hard = m > 110
|
|
lab, n = ndimage.label(hard)
|
|
if n:
|
|
sizes = ndimage.sum(hard, lab, range(1, n + 1)); keep = np.zeros_like(hard)
|
|
for i, sz in enumerate(sizes):
|
|
if sz > 0.02 * sizes.max(): keep |= lab == i + 1
|
|
hard = keep
|
|
holes = ndimage.binary_fill_holes(hard) & ~hard
|
|
hl, hn = ndimage.label(holes)
|
|
for i in range(1, hn + 1):
|
|
if (hl == i).sum() < 400: hard |= hl == i
|
|
hard = ndimage.binary_erosion(hard, iterations=2)
|
|
a = Image.fromarray((hard * 255).astype("uint8")).filter(ImageFilter.GaussianBlur(0.8))
|
|
o = im.convert("RGBA"); o.putalpha(a); o.save(dst)
|
|
boxen.append(Image.open(dst).getchannel("A").getbbox())
|
|
bx = [b for b in boxen if b]
|
|
box = [min(b[0] for b in bx) - 4, min(b[1] for b in bx) - 4, max(b[2] for b in bx) + 4, max(b[3] for b in bx) + 4]
|
|
box = [max(0, box[0]), max(0, box[1]), min(1284, box[2]), min(716, box[3])]
|
|
if (box[2] - box[0]) % 2: box[2] -= 1
|
|
if (box[3] - box[1]) % 2: box[3] -= 1
|
|
w, h = box[2] - box[0], box[3] - box[1]
|
|
subprocess.run(["ffmpeg", "-v", "error", "-y", "-framerate", "24", "-i", f"{d}/frei/%04d.png",
|
|
"-vf", f"crop={w}:{h}:{box[0]}:{box[1]}", "-c:v", "libvpx-vp9", "-pix_fmt", "yuva420p",
|
|
"-b:v", "0", "-crf", "28", "-auto-alt-ref", "0", f"{ziel}/{name}.webm"], check=True)
|
|
json.dump({"box": box, "boxen": boxen, "frames": len(roh)}, open(f"{ziel}/{name}.json", "w"))
|
|
print(name, box, len(roh), flush=True)
|