diff --git a/app/api/scenes.py b/app/api/scenes.py
index 3f0813a..55b556c 100644
--- a/app/api/scenes.py
+++ b/app/api/scenes.py
@@ -1,7 +1,9 @@
 """GET /scenes, lista i szczegóły scen z bazy kanonicznej."""
 from __future__ import annotations
 
+import collections
 import logging
+import math
 import re
 import hashlib
 import uuid
@@ -55,6 +57,47 @@ router = APIRouter(prefix="/scenes", tags=["scenes"], dependencies=[Depends(requ
 #: Ile wynikow filtra bierze udzial w losowaniu (patrz sort=random z filtrem).
 _RANDOM_POOL = 5000
 
+#: Glowny feed "newest ingest": najwiekszy udzial jednego zrodla na stronie i ile pierwszych
+#: stron tak ukladamy (dalej zwykla kolejnosc po created_at). Patrz _diversify_pages.
+_FEED_MAX_SHARE = 1 / 3
+_FEED_DIVERSE_PAGES = 8
+
+
+def _diversify_pages(
+    items: list[tuple[uuid.UUID, str]], page_size: int, max_share: float
+) -> list[list[uuid.UUID]]:
+    """Rozklada liste (id, zrodlo), juz w kolejnosci czasu, na strony tak, by jedno zrodlo
+    zajmowalo najwyzej `max_share` strony, dopoki sa sceny z innych zrodel. Strona bierze
+    najpierw najnowsze sceny w ramach limitu na zrodlo; gdy innych zrodel brakuje, dopelnia
+    sie najstarszymi z nadmiaru, wiec zadna strona nie zostaje dziurawa i nic nie ginie.
+    W obrebie strony i w obrebie zrodla kolejnosc czasu zostaje.
+
+    Po co: eporner dokladal ~2 tys. anonimowych uploadow dziennie przy kilkuset nowych
+    scenach z calej reszty i po odcieciu archiwum (backfill) zajmowal 121 ze 150 pozycji
+    pierwszych trzech stron (2026-09-19).
+    """
+    cap = max(1, math.ceil(page_size * max_share))
+    remaining = list(enumerate(items))
+    pages: list[list[uuid.UUID]] = []
+    while remaining:
+        taken: list[tuple[int, tuple[uuid.UUID, str]]] = []
+        overflow: list[tuple[int, tuple[uuid.UUID, str]]] = []
+        counts: collections.Counter = collections.Counter()
+        for pos, item in remaining:
+            if len(taken) < page_size and counts[item[1]] < cap:
+                taken.append((pos, item))
+                counts[item[1]] += 1
+            else:
+                overflow.append((pos, item))
+        while len(taken) < page_size and overflow:
+            taken.append(overflow.pop(0))
+        taken.sort(key=lambda t: t[0])
+        pages.append([item[0] for _, item in taken])
+        used = {pos for pos, _ in taken}
+        remaining = [(pos, item) for pos, item in remaining if pos not in used]
+    return pages
+
+
 _VALID_SORTS = {
     "created_at",
     "release_date",
@@ -599,6 +642,21 @@ def list_scenes(
         # stubem. Filtruje anonymous tube-only sceny z newUrl/categories ingestu.
         base = base.where(stub_exclusion_clause())
 
+    # Glowny feed "newest ingest" bez zawezen: bez scen z dociagania archiwum (backfill).
+    # deep-crawl tubepornclassic dokladal 7,9 tys. scen z lat 2016-2018 dziennie, a xhamster,
+    # xnxx, xvideos i sxyprn kolejne 3,3 tys., wszystkie z dzisiejszym created_at, wiec
+    # zajmowaly caly szczyt listy (zgloszenie 2026-09-19: "glownie theclassicporn i eporner").
+    # Z filtrem (szukanie, tag, obsada, studio, strona) i w sekcjach JAV/Anime zostaja, bo
+    # tam user chce wszystkiego z danego zakresu, a nie przegladu nowosci.
+    _plain_feed = (
+        sort == "created_at"
+        and not (q or tag_slug_list or perf_id_strings or studio_slug_list or origin)
+        and vertical is None
+        and not jav
+    )
+    if _plain_feed:
+        base = base.where(Scene.backfill.is_(False))
+
     _is_pure_default = (
         not include_stubs and not q and not studio_slug_list and not tag_slug_list
         and not perf_id_strings and origin is None and has_playback is None
@@ -759,17 +817,40 @@ def list_scenes(
     # psuje early-termination i przy filtrach EXISTS planer robi gather-all+sort (sekundy)
     # zamiast limit-aware index-walk po `ix_scenes_created_at_desc`. page/per_page to
     # walidowane inty (Query ge=1, le=200), więc literal_column jest bezpieczne.
-    rows = (
-        studio_rows
-        if studio_rows is not None
-        else session.execute(
-            ordered.offset(literal_column(str(_off))).limit(literal_column(str(_want)))
+    if _plain_feed and page <= _FEED_DIVERSE_PAGES:
+        # Kandydatki (id + zrodlo) w kolejnosci czasu, ulozone przez _diversify_pages. Pula
+        # ma STALA wielkosc (2x tyle, ile stron ukladamy), bo inaczej kazda strona liczylaby
+        # uklad z innej puli i sceny przeskakiwalyby miedzy stronami. Pelne wiersze ladujemy
+        # tylko dla zwracanej strony.
+        cand_n = _FEED_DIVERSE_PAGES * per_page * 2
+        origin_col = (
+            select(func.min(PlaybackSource.origin))
+            .where(PlaybackSource.scene_id == Scene.id, PlaybackSource.dead_at.is_(None))
+            .correlate(Scene)
+            .scalar_subquery()
         )
-        .scalars()
-        .all()
-    )
-    has_more = len(rows) > per_page
-    rows = rows[:per_page]
+        cand = session.execute(
+            ordered.with_only_columns(Scene.id, origin_col).limit(literal_column(str(cand_n)))
+        ).all()
+        feed_pages = _diversify_pages([(r[0], r[1] or "") for r in cand], per_page, _FEED_MAX_SHARE)
+        page_ids = feed_pages[page - 1] if page <= len(feed_pages) else []
+        by_id = {
+            s.id: s for s in session.execute(select(Scene).where(Scene.id.in_(page_ids))).scalars()
+        } if page_ids else {}
+        rows = [by_id[i] for i in page_ids if i in by_id]
+        has_more = page < len(feed_pages) or len(cand) >= cand_n
+    else:
+        rows = (
+            studio_rows
+            if studio_rows is not None
+            else session.execute(
+                ordered.offset(literal_column(str(_off))).limit(literal_column(str(_want)))
+            )
+            .scalars()
+            .all()
+        )
+        has_more = len(rows) > per_page
+        rows = rows[:per_page]
 
     # Filtrowane listy: total = dolna granica z dotychczas-widzianych wierszy, a
     # total_capped=has_more daje UI "N+" (jest kolejna strona). Bez osobnego count query.
diff --git a/tests/test_feed_diversify.py b/tests/test_feed_diversify.py
new file mode 100644
index 0000000..3fd988b
--- /dev/null
+++ b/tests/test_feed_diversify.py
@@ -0,0 +1,51 @@
+import uuid
+
+from app.api.scenes import _diversify_pages
+
+
+def _ids(n, origin):
+    return [(uuid.uuid4(), origin) for _ in range(n)]
+
+
+def _mixed(flood_share=0.7, n=300, others=("b", "c", "d")):
+    """Jak feed 2026-09-19: jedno zrodlo zalewa, reszta kapie, wszystko w kolejnosci czasu."""
+    items, k = [], 0
+    for i in range(n):
+        if (i % 10) < flood_share * 10:
+            items.append((uuid.uuid4(), "flood"))
+        else:
+            items.append((uuid.uuid4(), others[k % len(others)]))
+            k += 1
+    return items
+
+
+def test_flood_is_capped_while_other_sources_exist():
+    items = _mixed()
+    pages = _diversify_pages(items, 10, 1 / 3)
+    origin = dict(items)
+    # 90 scen z innych zrodel wystarcza na 15 stron po 6; tyle stron trzyma limit
+    for pg in pages[:15]:
+        assert len(pg) == 10
+        assert sum(origin[i] == "flood" for i in pg) <= 4
+
+
+def test_nothing_is_lost_and_pages_are_full():
+    items = _mixed()
+    pages = _diversify_pages(items, 10, 1 / 3)
+    assert sorted(i for pg in pages for i in pg) == sorted(i for i, _ in items)
+    assert all(len(pg) == 10 for pg in pages[:-1])
+
+
+def test_two_sources_still_fill_pages():
+    items = _ids(40, "a") + _ids(40, "b")
+    pages = _diversify_pages(items, 10, 1 / 3)
+    assert all(len(pg) == 10 for pg in pages)
+
+
+def test_order_within_source_is_kept():
+    items = _mixed()
+    flat = [i for pg in _diversify_pages(items, 10, 1 / 3) for i in pg]
+    for src in ("flood", "b", "c", "d"):
+        mine = [i for i, o in items if o == src]
+        mine_set = set(mine)
+        assert [i for i in flat if i in mine_set] == mine