ImageStudio / watermark.py
ImageStudio Maintainer
feat: watermark assets + skip R2 upload for web-UI generations
74e9ec6
Raw
History Blame Contribute Delete
10.7 kB
"""Brand watermark overlay for generated images and videos.
A watermark is requested by passing a JSON spec (a string or an already-parsed
dict) to a generation endpoint via the ``watermark`` parameter. When the spec is
valid, a small badge — the brand logo with the site domain rendered beneath it —
is composited onto the generated image (or every frame of the generated video,
via a single ffmpeg overlay pass). An empty or invalid spec is a no-op, so any
caller that omits the parameter is completely unaffected.
Spec fields (JSON object; at least one of ``logo`` / ``domain`` must be present
for the spec to count as "valid"):
logo str logo image as an http(s) URL, a ``data:`` URI, or raw
base64. Optional — a domain-only watermark is allowed.
domain str text drawn under the logo, e.g. "nsfwailab.com". Optional.
position str bottom-right (default), bottom-left, top-right, top-left,
bottom-center or top-center.
scale float badge width as a fraction of the media width (default 0.18,
clamped to 0.05..0.6).
opacity float overall badge opacity 0..1 (default 0.9).
margin float badge margin from the edge as a fraction of media width
(default 0.03).
Every public entry point is wrapped so a watermarking failure degrades to
"asset returned un-watermarked" rather than breaking generation.
"""
from __future__ import annotations
import base64
import json
import os
import subprocess
import tempfile
import urllib.request
_VIDEO_REF_W = 1024 # fallback media width when a video's size can't be probed.
# Cache decoded logos so a per-frame / per-call render doesn't refetch or
# re-decode the same image. Keyed on the raw logo string.
_logo_cache: dict[str, object] = {}
_POSITIONS = {
"bottom-right", "bottom-left", "top-right", "top-left",
"bottom-center", "top-center",
}
def parse_spec(spec) -> dict | None:
"""Validate/normalise a watermark spec; return a clean dict or None.
Accepts a JSON string or a dict. Returns None (a no-op signal) when the spec
is empty, unparseable, or carries neither a logo nor a domain.
"""
if not spec:
return None
data = spec
if isinstance(spec, str):
s = spec.strip()
if not s:
return None
try:
data = json.loads(s)
except (ValueError, TypeError):
return None
if not isinstance(data, dict):
return None
logo = data.get("logo") or data.get("logo_url") or ""
domain = (data.get("domain") or data.get("text") or "").strip()
if not (logo or domain):
return None
def _f(key, default, lo, hi):
try:
return max(lo, min(hi, float(data.get(key, default))))
except (TypeError, ValueError):
return default
position = str(data.get("position", "bottom-right")).strip().lower()
if position not in _POSITIONS:
position = "bottom-right"
return {
"logo": logo if isinstance(logo, str) else "",
"domain": domain,
"position": position,
"scale": _f("scale", 0.18, 0.05, 0.6),
"opacity": _f("opacity", 0.9, 0.1, 1.0),
"margin": _f("margin", 0.03, 0.0, 0.2),
}
def is_valid(spec) -> bool:
"""True when ``spec`` would produce a watermark."""
return parse_spec(spec) is not None
def _load_logo(logo: str):
"""Decode a logo string (URL / data-URI / base64) into an RGBA PIL image."""
if not logo:
return None
if logo in _logo_cache:
return _logo_cache[logo]
from PIL import Image
import io
img = None
try:
raw = None
if logo.startswith("data:"):
_, _, b64 = logo.partition(",")
raw = base64.b64decode(b64)
elif logo.startswith("http://") or logo.startswith("https://"):
req = urllib.request.Request(logo, headers={"User-Agent": "watermark/1.0"})
with urllib.request.urlopen(req, timeout=10) as resp:
raw = resp.read()
else:
raw = base64.b64decode(logo)
if raw:
img = Image.open(io.BytesIO(raw)).convert("RGBA")
except Exception as exc: # noqa: BLE001 - logo is best-effort
print(f"[watermark] logo decode failed: {exc}")
img = None
_logo_cache[logo] = img
return img
_FONT_CANDIDATES = [
"/usr/share/fonts/truetype/dejavu/DejaVuSans-Bold.ttf",
"/usr/share/fonts/truetype/dejavu/DejaVuSans.ttf",
"/usr/share/fonts/truetype/liberation/LiberationSans-Bold.ttf",
"DejaVuSans-Bold.ttf",
]
def _font(size: int):
from PIL import ImageFont
for path in _FONT_CANDIDATES:
try:
return ImageFont.truetype(path, size)
except Exception: # noqa: BLE001
continue
return ImageFont.load_default()
def _render_badge(cfg: dict, media_w: int):
"""Build the watermark badge (logo + domain on a translucent plate)."""
from PIL import Image, ImageDraw
badge_w = max(72, int(media_w * cfg["scale"]))
pad = max(6, badge_w // 16)
logo = _load_logo(cfg["logo"])
logo_img = None
logo_h = 0
if logo is not None and logo.width > 0:
lw = max(1, badge_w - 2 * pad)
lh = max(1, int(logo.height * lw / logo.width))
logo_img = logo.resize((lw, lh), Image.LANCZOS)
logo_h = lh
domain = cfg["domain"]
font = None
text_w = text_h = 0
bbox = (0, 0, 0, 0)
if domain:
measure = ImageDraw.Draw(Image.new("RGBA", (1, 1)))
fs = max(11, int(badge_w * 0.14))
font = _font(fs)
bbox = measure.textbbox((0, 0), domain, font=font)
text_w, text_h = bbox[2] - bbox[0], bbox[3] - bbox[1]
max_tw = badge_w - 2 * pad
if text_w > max_tw and text_w > 0:
fs = max(8, int(fs * max_tw / text_w))
font = _font(fs)
bbox = measure.textbbox((0, 0), domain, font=font)
text_w, text_h = bbox[2] - bbox[0], bbox[3] - bbox[1]
gap = pad if (logo_img is not None and domain) else 0
badge_h = logo_h + gap + text_h + 2 * pad
badge = Image.new("RGBA", (badge_w, badge_h), (0, 0, 0, 0))
draw = ImageDraw.Draw(badge)
radius = max(8, badge_w // 12)
draw.rounded_rectangle([0, 0, badge_w - 1, badge_h - 1], radius=radius,
fill=(0, 0, 0, 110))
y = pad
if logo_img is not None:
badge.alpha_composite(logo_img, ((badge_w - logo_img.width) // 2, y))
y += logo_h + gap
if domain:
tx = (badge_w - text_w) // 2 - bbox[0]
draw.text((tx, y - bbox[1]), domain, font=font, fill=(255, 255, 255, 235))
opacity = cfg["opacity"]
if opacity < 1.0:
alpha = badge.split()[3].point(lambda v: int(v * opacity))
badge.putalpha(alpha)
return badge
def _offset(position: str, media_w: int, media_h: int, bw: int, bh: int, margin: int):
"""Top-left pixel offset for the badge given a named position."""
if "right" in position:
x = media_w - bw - margin
elif "left" in position:
x = margin
else: # center
x = (media_w - bw) // 2
y = margin if position.startswith("top") else media_h - bh - margin
return max(0, x), max(0, y)
def apply_to_image(img, spec):
"""Return ``img`` with the watermark composited on, or unchanged on no-op."""
cfg = parse_spec(spec)
if cfg is None:
return img
try:
base = img.convert("RGBA")
badge = _render_badge(cfg, base.width)
margin = int(cfg["margin"] * base.width)
x, y = _offset(cfg["position"], base.width, base.height,
badge.width, badge.height, margin)
base.alpha_composite(badge, (x, y))
return base.convert("RGB")
except Exception as exc: # noqa: BLE001 - never break generation
print(f"[watermark] image overlay failed: {exc}")
return img
def _video_dims(path: str):
try:
import cv2
cap = cv2.VideoCapture(path)
w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
cap.release()
if w > 0 and h > 0:
return w, h
except Exception: # noqa: BLE001
pass
# Fallback when OpenCV is unavailable: parse the resolution out of ffmpeg's
# probe output so the badge is still sized/placed against the real frame.
try:
import re
import imageio_ffmpeg
ff = imageio_ffmpeg.get_ffmpeg_exe()
err = subprocess.run([ff, "-i", path], capture_output=True, text=True).stderr
m = re.search(r"Video:.*?(\d{2,5})x(\d{2,5})", err)
if m:
return int(m.group(1)), int(m.group(2))
except Exception: # noqa: BLE001
pass
return None
def apply_to_video(in_path: str, spec, out_path: str | None = None) -> str:
"""Overlay the watermark on a video via one ffmpeg pass.
Returns the path to the watermarked file, or the original ``in_path`` when
the spec is a no-op or anything fails (so upload still proceeds).
"""
cfg = parse_spec(spec)
if cfg is None:
return in_path
badge_png = None
try:
import imageio_ffmpeg
ffmpeg = imageio_ffmpeg.get_ffmpeg_exe()
dims = _video_dims(in_path)
media_w, media_h = dims if dims else (_VIDEO_REF_W, _VIDEO_REF_W)
badge = _render_badge(cfg, media_w)
margin = int(cfg["margin"] * media_w)
x, y = _offset(cfg["position"], media_w, media_h,
badge.width, badge.height, margin)
fd, badge_png = tempfile.mkstemp(suffix=".png")
os.close(fd)
badge.save(badge_png)
if out_path is None:
fd, out_path = tempfile.mkstemp(suffix=".mp4")
os.close(fd)
cmd = [
ffmpeg, "-y", "-i", in_path, "-i", badge_png,
"-filter_complex", f"overlay={x}:{y}:format=auto",
"-c:a", "copy", "-c:v", "libx264", "-pix_fmt", "yuv420p",
"-crf", "18", "-preset", "veryfast", "-movflags", "+faststart",
out_path,
]
subprocess.run(cmd, check=True, capture_output=True)
return out_path
except Exception as exc: # noqa: BLE001 - never break generation
detail = exc.stderr.decode("utf-8", "ignore")[-400:] if isinstance(exc, subprocess.CalledProcessError) else exc
print(f"[watermark] video overlay failed: {detail}")
return in_path
finally:
if badge_png and os.path.exists(badge_png):
try:
os.unlink(badge_png)
except OSError:
pass