diff --git a/web/enrichment.py b/web/enrichment.py
index fb9bc42..7e8b103 100644
--- a/web/enrichment.py
+++ b/web/enrichment.py
@@ -35,8 +35,10 @@ IMAGES_DIR.mkdir(parents=True, exist_ok=True)
 
 MAX_IMAGES = 12
 MAX_IMAGE_BYTES = 5_000_000
+MIN_IMAGE_BYTES = 15_000           # Below this, it's an icon / tracking pixel
+MIN_IMAGE_DIMENSION = 400          # Shortest side in pixels — filters thumbs
 IMAGE_TIMEOUT = 15
-PHASH_DUPLICATE_THRESHOLD = 5  # Hamming distance ≤ N → considered the same picture
+PHASH_DUPLICATE_THRESHOLD = 5      # Hamming distance ≤ N → same picture
 
 
 class EnrichmentError(Exception):
@@ -87,19 +89,19 @@ def _ext_for(content_type: str, url: str) -> str:
     return ext.lower() or ".jpg"
 
 
-def _phash(data: bytes):
-    """Return an imagehash.ImageHash for the bytes, or None if unsupported."""
+def _image_info(data: bytes):
+    """Return (phash, (width, height)) for the bytes, or (None, None) on failure."""
     try:
         from PIL import Image
         import imagehash
     except ImportError:
-        return None
+        return None, None
     try:
         with Image.open(io.BytesIO(data)) as img:
             img.load()
-            return imagehash.phash(img)
+            return imagehash.phash(img), img.size
     except Exception:
-        return None
+        return None, None
 
 
 def _download_images(flat_id: str, urls: list[str], referer: str) -> int:
@@ -149,12 +151,20 @@ def _download_images(flat_id: str, urls: list[str], referer: str) -> int:
 
         data_bytes = bytes(data)
 
+        # Size filter: files under ~15 KB are icons/tracking pixels.
+        if len(data_bytes) < MIN_IMAGE_BYTES:
+            continue
+
         sha = hashlib.sha256(data_bytes).hexdigest()
         if sha in seen_sha:
             continue
         seen_sha.add(sha)
 
-        ph = _phash(data_bytes)
+        ph, dims = _image_info(data_bytes)
+        # Pixel filter: anything smaller than 400 px on the short side is a
+        # thumbnail/avatar, not a real flat photo.
+        if dims is not None and min(dims) < MIN_IMAGE_DIMENSION:
+            continue
         if ph is not None:
             if any((ph - prev) <= PHASH_DUPLICATE_THRESHOLD for prev in seen_phash):
                 continue
diff --git a/web/llm.py b/web/llm.py
index 73afa88..de9299b 100644
--- a/web/llm.py
+++ b/web/llm.py
@@ -36,11 +36,18 @@ TOOL_SCHEMA = {
 
 SYSTEM_PROMPT = (
     "Du bekommst eine Liste von Bild-URLs einer Wohnungsanzeige. Wähle nur "
-    "die URLs aus, die ein Foto der Wohnung zeigen (Innenraum, Außenansicht "
-    "des Gebäudes, Grundriss). Verwerfe Logos, Icons, Banner, Ads, "
-    "Bewertungs-Sterne, Karten/Stadtpläne, Mitarbeiter-Portraits, Tracking-"
-    "Pixel. Behalte die Reihenfolge der Input-Liste bei. Antworte "
-    "ausschließlich über den Tool-Call."
+    "die URLs aus, die ein Foto der Wohnung zeigen (Innenraum, Außenansicht, "
+    "Grundriss). Verwerfe Logos, Icons, Banner, Ads, Bewertungs-Sterne, "
+    "Karten/Stadtpläne, Mitarbeiter-Portraits, Tracking-Pixel.\n\n"
+    "WICHTIG: Viele Wohnungsbaugesellschaften liefern ihre Fotos über ein "
+    "CDN mit kryptischen Dateinamen aus (z.B. "
+    "fileadmin/_processed_/2/3/xcsm_abc123def.webp.pagespeed.ic.xyz.webp). "
+    "Solche URLs sind IMMER zu behalten, wenn sie vom selben Host wie die "
+    "Anzeige kommen und in /fileadmin/_processed_/ oder /assets/media/ "
+    "oder ähnlichen Media-Pfaden liegen. Lehne nichts nur wegen eines "
+    "unleserlichen Dateinamens ab.\n\n"
+    "Behalte die Reihenfolge der Input-Liste bei. Antworte ausschließlich "
+    "über den Tool-Call."
 )
 
 
@@ -92,5 +99,13 @@ def select_flat_image_urls(candidates: list[str], page_url: str,
             # Constrain to the original candidate set so the model can't
             # invent URLs (it sometimes lightly rewrites them otherwise).
             allowed = set(candidates)
-            return [u for u in urls if u in allowed]
+            kept = [u for u in urls if u in allowed]
+            if not kept and candidates:
+                # The model rejected everything — usually because the URLs
+                # are opaque CDN hashes it couldn't identify positively.
+                # Trust the Playwright pre-filter instead of returning [].
+                logger.warning("image-select returned 0 of %d — falling back to unfiltered list",
+                               len(candidates))
+                return candidates
+            return kept
     return candidates