diff --git a/.dedupe/index.sqlite b/.dedupe/index.sqlite
new file mode 100644
index 0000000..d2fa618
Binary files /dev/null and b/.dedupe/index.sqlite differ
diff --git a/.dedupe/report.csv b/.dedupe/report.csv
new file mode 100644
index 0000000..8056de2
--- /dev/null
+++ b/.dedupe/report.csv
@@ -0,0 +1 @@
+group_id,canonical_path,dup_path,dup_size_bytes
diff --git a/tools/image_deduper.py b/tools/image_deduper.py
index 8bbec23..715e4e0 100644
--- a/tools/image_deduper.py
+++ b/tools/image_deduper.py
@@ -2,39 +2,40 @@
 """
 image_deduper.py — Finde und bereinige Bild-Dubletten sicher & reversibel.
 
-Funktionen:
-- Scan: Verzeichnisse rekursiv scannen, sha256 + pHash berechnen
-- Report: CSV + menschenlesbare Zusammenfassung mit Gruppen
-- Apply: Duplikate auf kanonische Datei umbiegen (Hardlink/Löschen)
-- Optional: DB-Referenzen aktualisieren (SQLite/SQLModel kompatibel)
+Neu:
+- collect: Bild-URLs aus JSON/HTML/Markdown/Text sammeln und lokal in .media_cache/ speichern
+- scan/report/apply: wie gehabt (Hardlink/Delete), jetzt standardmäßig mit Cache nutzbar
 
-Nutzung (Beispiele):
-  # 1) Nur scannen + reporten (keine Änderungen):
-  python tools/image_deduper.py scan --roots media,assets/images --out-dir .dedupe --phash
+Workflows:
+  # A) Nur remote-Referenzen vorhanden → erst sammeln, dann deduplizieren
+  python tools/image_deduper.py collect --sources processed_articles.json,content/ --cache .media_cache
+  python tools/image_deduper.py scan --roots .media_cache --out-dir .dedupe --phash
+  python tools/image_deduper.py apply --report .dedupe/report.csv --mode hardlink --dry-run
+  python tools/image_deduper.py apply --report .dedupe/report.csv --mode hardlink
 
-  # 2) Report anzeigen:
-  python tools/image_deduper.py report --index .dedupe/index.sqlite --csv
+  # B) Lokale Bild-Ordner vorhanden
+  python tools/image_deduper.py scan --roots ./media,./static/images --out-dir .dedupe --phash
 
-  # 3) Anwenden (Hardlinks setzen, Dry-Run):
-  python tools/image_deduper.py apply --index .dedupe/index.sqlite --mode hardlink --dry-run
-
-  # 4) Anwenden (wirklich ändern):
-  python tools/image_deduper.py apply --index .dedupe/index.sqlite --mode hardlink
-
-  # 5) Referenzen in DB aktualisieren (optional):
-  python tools/image_deduper.py apply --index .dedupe/index.sqlite --update-db sqlite:///./rssbot.db --dry-run
+Hinweise:
+- "collect" speichert Bilder content-addressed (<sha256>.<ext>) und lädt identische Dateien nicht doppelt.
+- Für HTML/Markdown werden Bild-URLs geparst; für JSON werden alle Stringwerte mit Bild-URL-Muster extrahiert.
 """
 
 import argparse
 import csv
 import hashlib
+import json
+import mimetypes
 import os
+import re
 import sqlite3
 import sys
 from dataclasses import dataclass
 from pathlib import Path
-from typing import Iterable, List, Optional, Tuple
+from typing import Iterable, List, Optional, Tuple, Set
+from urllib.parse import urlparse
 
+# Optionale Parser/Helpers
 try:
     from PIL import Image
 except ImportError:
@@ -45,12 +46,39 @@ try:
 except ImportError:
     imagehash = None
 
+try:
+    import requests
+except ImportError:
+    requests = None
+
+try:
+    from bs4 import BeautifulSoup
+except ImportError:
+    BeautifulSoup = None
+
+try:
+    from tqdm import tqdm
+except ImportError:
+    tqdm = None
+
 
 IMAGE_EXTS = {".jpg", ".jpeg", ".png", ".webp", ".gif"}
 DEFAULT_INDEX = ".dedupe/index.sqlite"
 DEFAULT_REPORT = ".dedupe/report.csv"
+DEFAULT_CACHE = ".media_cache"
+
+URL_RE = re.compile(
+    r"""https?://[^\s"'<>]+?\.(?:jpg|jpeg|png|webp|gif)(?:\?[^\s"'<>]*)?""",
+    re.IGNORECASE,
+)
+
+def human_mb(nbytes: int) -> str:
+    return f"{nbytes/1_000_000:.2f} MB"
 
 
+# ---------------------------
+# Hashing & pHash
+# ---------------------------
 def sha256_file(path: Path, bufsize: int = 1024 * 1024) -> str:
     h = hashlib.sha256()
     with path.open("rb") as f:
@@ -74,6 +102,9 @@ def calc_phash(path: Path) -> Optional[str]:
         return None
 
 
+# ---------------------------
+# Index (SQLite)
+# ---------------------------
 def ensure_dir(p: Path):
     p.mkdir(parents=True, exist_ok=True)
 
@@ -104,6 +135,9 @@ def is_image(path: Path) -> bool:
 
 def walk_images(roots: List[Path]) -> Iterable[Path]:
     for root in roots:
+        if not root.exists():
+            # Skip silently to be more forgiving; user might pass multiple roots
+            continue
         for p in root.rglob("*"):
             if p.is_file() and is_image(p):
                 yield p
@@ -129,12 +163,9 @@ def upsert_file(db_path: Path, path: Path, sha256: str, phash: Optional[str]):
 
 
 def group_by_sha256(db_path: Path) -> List[List[Tuple[int, str, int]]]:
-    """Return list of groups: [(id, path, size), ...] where sha256 identical and len(group) > 1."""
     conn = sqlite3.connect(str(db_path))
     cur = conn.cursor()
-    cur.execute("""
-        SELECT sha256 FROM files GROUP BY sha256 HAVING COUNT(*) > 1;
-    """)
+    cur.execute("SELECT sha256 FROM files GROUP BY sha256 HAVING COUNT(*) > 1;")
     hashes = [r[0] for r in cur.fetchall()]
     groups = []
     for h in hashes:
@@ -157,8 +188,7 @@ def write_csv_report(db_path: Path, csv_path: Path) -> Tuple[int, int]:
         for g in groups:
             if not g:
                 continue
-            # Kanon: größte Datei (oder erste)
-            canonical = max(g, key=lambda x: x[2])
+            canonical = max(g, key=lambda x: x[2])  # größte Datei als Kanon
             for rid, path, size in g:
                 if path == canonical[1]:
                     continue
@@ -169,20 +199,22 @@ def write_csv_report(db_path: Path, csv_path: Path) -> Tuple[int, int]:
     return total_dups, total_savings
 
 
+# ---------------------------
+# Apply (Hardlink/Delete)
+# ---------------------------
 def apply_hardlink(canonical: Path, dup: Path, dry_run: bool) -> None:
-    # Ersetzt dup durch Hardlink auf canonical (gleiche Partition nötig)
     if dry_run:
         return
     tmp = dup.with_suffix(dup.suffix + ".dedupe.tmp")
-    dup.unlink()                # entferne dup
-    os.link(canonical, tmp)     # hardlink temp
-    tmp.replace(dup)            # atomarer move
+    dup.unlink(missing_ok=False)
+    os.link(canonical, tmp)  # gleicher FS notwendig
+    tmp.replace(dup)
 
 
 def apply_delete(dup: Path, dry_run: bool) -> None:
     if dry_run:
         return
-    dup.unlink()
+    dup.unlink(missing_ok=False)
 
 
 @dataclass
@@ -215,89 +247,42 @@ def apply_changes(csv_report: Path, mode: str, dry_run: bool) -> ApplyStats:
     return stats
 
 
-def parse_roots(roots_arg: str) -> List[Path]:
-    parts = [Path(p.strip()) for p in roots_arg.split(",") if p.strip()]
-    for p in parts:
-        if not p.exists():
-            raise FileNotFoundError(f"Root not found: {p}")
-    return parts
+# ---------------------------
+# Collect (URLs -> Cache)
+# ---------------------------
+def is_image_url(url: str) -> bool:
+    if URL_RE.search(url):
+        return True
+    # Fallback: Extension anhand des Pfads
+    path = urlparse(url).path
+    ext = Path(path).suffix.lower()
+    return ext in IMAGE_EXTS
 
 
-def cmd_scan(args):
-    out_dir = Path(args.out_dir)
-    index = Path(args.index or DEFAULT_INDEX)
-    ensure_dir(out_dir)
-    ensure_dir(index.parent)
-    init_index(index)
-    roots = parse_roots(args.roots)
-
-    count = 0
-    for path in walk_images(roots):
-        try:
-            h = sha256_file(path)
-            ph = calc_phash(path) if args.phash else None
-            upsert_file(index, path, h, ph)
-            count += 1
-            if count % 500 == 0:
-                print(f"... indexed {count} files")
-        except Exception as e:
-            print(f"[WARN] {path}: {e}", file=sys.stderr)
-
-    dups, savings = write_csv_report(index, Path(args.report or DEFAULT_REPORT))
-    print(f"Indexed {count} images. Found duplicate files: {dups}, potential savings: {savings/1_000_000:.2f} MB")
-    print(f"Index: {index}")
-    print(f"Report: {args.report or DEFAULT_REPORT}")
+def extract_urls_from_text(text: str) -> Set[str]:
+    urls = set(URL_RE.findall(text))
+    return urls
 
 
-def cmd_report(args):
-    index = Path(args.index or DEFAULT_INDEX)
-    csv_path = Path(args.report or DEFAULT_REPORT)
-    dups, savings = write_csv_report(index, csv_path)
-    print(f"Duplicates: {dups}, potential savings: {savings/1_000_000:.2f} MB")
-    if args.csv:
-        print(f"CSV written: {csv_path}")
+def extract_urls_from_html(text: str) -> Set[str]:
+    urls = set()
+    if BeautifulSoup is None:
+        return extract_urls_from_text(text)
+    try:
+        soup = BeautifulSoup(text, "html.parser")
+        for tag in soup.find_all(["img", "source"]):
+            src = tag.get("src") or tag.get("data-src")
+            if src and is_image_url(src):
+                urls.add(src)
+        # Fallback auf nackte URLs
+        urls |= extract_urls_from_text(text)
+    except Exception:
+        urls |= extract_urls_from_text(text)
+    return urls
 
 
-def cmd_apply(args):
-    csv_report = Path(args.report or DEFAULT_REPORT)
-    if not csv_report.exists():
-        raise FileNotFoundError(f"Report not found: {csv_report}")
-    stats = apply_changes(csv_report, args.mode, args.dry_run)
-    print(f"Processed: {stats.processed}, Errors: {stats.errors}, Saved: {stats.saved_bytes/1_000_000:.2f} MB (mode={args.mode}, dry_run={args.dry_run})")
-    if args.update_db:
-        # Platzhalter: hier könntest du eure DB-Referenzen aktualisieren (falls Bilder-Paths in DB gespeichert sind).
-        # Beispiel: SQLModel mit Tabelle ImageMeta(content_hash UNIQUE, local_path) → auf kanonischen Pfad umbiegen.
-        print(f"[INFO] DB update requested for: {args.update_db} (implementierung projektspezifisch)")
-
-
-def main():
-    ap = argparse.ArgumentParser(description="Bild-Deduplizierung (scan/report/apply)")
-    sub = ap.add_subparsers(dest="cmd", required=True)
-
-    sc = sub.add_parser("scan", help="Verzeichnisse scannen und Index/Report erstellen")
-    sc.add_argument("--roots", required=True, help="Kommagetrennte Wurzelpfade, z.B. 'media,assets/images'")
-    sc.add_argument("--out-dir", default=".dedupe", help="Ausgabeverzeichnis für Index/Reports")
-    sc.add_argument("--index", help="Pfad zur SQLite-Indexdatei (default .dedupe/index.sqlite)")
-    sc.add_argument("--report", help="Pfad zum CSV-Report (default .dedupe/report.csv)")
-    sc.add_argument("--phash", action="store_true", help="Perzeptuellen Hash berechnen (für zukünftige Near-Dups)")
-    sc.set_defaults(func=cmd_scan)
-
-    rp = sub.add_parser("report", help="Report neu generieren/anzeigen")
-    rp.add_argument("--index", help="Pfad zur SQLite-Indexdatei")
-    rp.add_argument("--report", help="Pfad zum CSV-Report")
-    rp.add_argument("--csv", action="store_true", help="CSV-Pfad ausgeben")
-    rp.set_defaults(func=cmd_report)
-
-    aply = sub.add_parser("apply", help="Änderungen anwenden (Hardlink/Delete)")
-    aply.add_argument("--report", help="Pfad zum CSV-Report")
-    aply.add_argument("--mode", choices=["hardlink", "delete"], default="hardlink", help="Strategie für Duplikate")
-    aply.add_argument("--dry-run", action="store_true", help="Nur anzeigen, nichts ändern")
-    aply.add_argument("--update-db", help="Optional: DB-URL für Referenz-Updates (projektspezifisch)")
-    aply.set_defaults(func=cmd_apply)
-
-    args = ap.parse_args()
-    args.func(args)
-
-
-if __name__ == "__main__":
-    main()
+def extract_urls_from_json(obj) -> Set[str]:
+    urls = set()
+    if isinstance(obj, dict):
+        for v in obj.values():
+            urls |= extract