diff --git a/.env.example b/.env.example
index 56bbd20..f679020 100644
--- a/.env.example
+++ b/.env.example
@@ -115,6 +115,11 @@ FAL_EDIT_TIER=balanced
 # next try, so those are retried in place (3 attempts total). `false` reverts
 # to fail-fast. A deterministic decliner costs up to 3 attempts before erroring.
 # RETRY_NO_MEDIA=true
+# Per-attempt wall-clock deadline on every fal call. Live-caught: one edit
+# call hung 12+ minutes while the SSE heartbeat kept the dead generation
+# open forever (endless shimmer, no banner). A timeout is NOT retried —
+# it fails fast to the friendly "took too long — hit retry" frame.
+# FAL_CALL_TIMEOUT_S=240
 # Errors reaching the browser are mapped to short human messages (the raw fal
 # body echoes the WHOLE prompt back — it must not render in the UI). Full
 # exception stays in logs + the additive `detail` field. `false` = raw again.
diff --git a/apps/modal-backend/providers/image.py b/apps/modal-backend/providers/image.py
index 26012fd..4081d9a 100644
--- a/apps/modal-backend/providers/image.py
+++ b/apps/modal-backend/providers/image.py
@@ -15,6 +15,7 @@
 
 from __future__ import annotations
 
+import asyncio
 import base64
 import os
 from dataclasses import dataclass
@@ -668,6 +669,17 @@ async def _fal_subscribe(
     no_media_generated. Default False keeps non-image callers (segmenter,
     video) byte-identical.
     """
+    # Wall-clock deadline per attempt (live-caught 2026-07-20: one fal edit
+    # call hung 12+ minutes with no error — and the SSE heartbeat happily
+    # kept the dead generation open forever, endless shimmer, no banner).
+    # A timeout is NON-retryable by design: fail fast to the friendly
+    # "took too long — hit retry" frame instead of stacking 3x the deadline.
+    # 240s clears the slowest legitimate fal path (gpt-image-2 edits ran
+    # ~170s in calibration); openrouter-hosted models don't ride this seam.
+    try:
+        deadline_s = max(30.0, float(os.environ.get("FAL_CALL_TIMEOUT_S", "") or 240.0))
+    except ValueError:
+        deadline_s = 240.0
     async for attempt in AsyncRetrying(
         stop=stop_after_attempt(3),
         wait=wait_exponential(multiplier=0.5, min=0.5, max=4),
@@ -675,8 +687,11 @@ async def _fal_subscribe(
         reraise=True,
     ):
         with attempt:
-            result = await fal_client.subscribe_async(
-                model, arguments=arguments, with_logs=False
+            result = await asyncio.wait_for(
+                fal_client.subscribe_async(
+                    model, arguments=arguments, with_logs=False
+                ),
+                timeout=deadline_s,
             )
             # Two success shapes exist: `images: [...]` (nano/kontext/fill) and
             # BRIA Expand's singular `image` object — accept either, so a
diff --git a/apps/modal-backend/tests/conftest.py b/apps/modal-backend/tests/conftest.py
index 75a4462..8bda351 100644
--- a/apps/modal-backend/tests/conftest.py
+++ b/apps/modal-backend/tests/conftest.py
@@ -69,6 +69,7 @@
     # for hermeticity).
     "RETRY_NO_MEDIA",
     "FRIENDLY_ERRORS",
+    "FAL_CALL_TIMEOUT_S",
     # Candidate empty-roll retry (default ON — same reasoning).
     "PRECOMPUTE_EMPTY_RETRY",
     # Interior enters (default ON — indoor register + interior judge; scrub
diff --git a/apps/modal-backend/tests/test_image_provider.py b/apps/modal-backend/tests/test_image_provider.py
index 7f900f2..da9ab15 100644
--- a/apps/modal-backend/tests/test_image_provider.py
+++ b/apps/modal-backend/tests/test_image_provider.py
@@ -704,3 +704,57 @@ async def fake_subscribe(model: str, arguments: dict, with_logs: bool) -> dict:
     monkeypatch.setattr(image.fal_client, "subscribe_async", fake_subscribe)
     result = await image._fal_subscribe("fal-ai/sam-3", {})
     assert result == {"masks": []}
+
+
+async def test_fal_subscribe_deadline_fails_fast(monkeypatch) -> None:
+    # Live-caught 2026-07-20: a fal edit call hung 12+ min; heartbeats kept
+    # the dead stream open forever. The per-attempt deadline turns a hang
+    # into TimeoutError (→ the friendly "took too long" frame) and is NOT
+    # retried — one deadline, not three stacked.
+    import asyncio
+
+    from providers import image as image_provider
+
+    calls = {"n": 0}
+
+    async def hung_subscribe(model, arguments=None, with_logs=False):
+        calls["n"] += 1
+        await asyncio.sleep(30)
+
+    monkeypatch.setenv("FAL_CALL_TIMEOUT_S", "30")  # floor clamps to 30s
+    monkeypatch.setattr(image_provider.fal_client, "subscribe_async", hung_subscribe)
+    # Shrink the effective deadline below the floor via a direct patch of the
+    # env read is not possible (floor 30s) — instead patch wait_for's timeout
+    # source: use a tiny sleep vs a tiny deadline by patching asyncio.wait_for
+    # would test the stdlib, not us. So: patch subscribe to outlive a 30s
+    # deadline is too slow for CI — instead verify the wiring by patching
+    # asyncio.wait_for to observe the timeout value and raise immediately.
+    seen = {}
+
+    async def spy_wait_for(coro, timeout=None):
+        seen["timeout"] = timeout
+        coro.close()
+        raise TimeoutError("deadline")
+
+    monkeypatch.setattr(image_provider.asyncio, "wait_for", spy_wait_for)
+    with pytest.raises(TimeoutError):
+        await image_provider._fal_subscribe("fal-ai/x", {})
+    assert seen["timeout"] == 30.0  # env honored (with the 30s floor)
+    assert calls["n"] == 0  # wait_for raised before subscribe ran (spy short-circuit)
+
+
+async def test_fal_subscribe_deadline_not_retried(monkeypatch) -> None:
+
+    from providers import image as image_provider
+
+    attempts = {"n": 0}
+
+    async def spy_wait_for(coro, timeout=None):
+        attempts["n"] += 1
+        coro.close()
+        raise TimeoutError("deadline")
+
+    monkeypatch.setattr(image_provider.asyncio, "wait_for", spy_wait_for)
+    with pytest.raises(TimeoutError):
+        await image_provider._fal_subscribe("fal-ai/x", {})
+    assert attempts["n"] == 1  # a hang fails FAST — no 3x deadline stack