diff --git a/anyplotlib/plot2d/_plot2d.py b/anyplotlib/plot2d/_plot2d.py
index 32d1d9f9..43586021 100644
--- a/anyplotlib/plot2d/_plot2d.py
+++ b/anyplotlib/plot2d/_plot2d.py
@@ -45,6 +45,40 @@ def _binary_transport_active() -> bool:
     return os.environ.get("APL_BINARY_TRANSPORT") == "1"
 
 
+def _reduce_mask_any(mask, out_h: int, out_w: int):
+    """Reduce a boolean *mask* to ``(out_h, out_w)`` with a block ANY.
+
+    A block of the source is True in the output if ANY of its pixels is. For a
+    mask that is the only defensible reduction: it marks *where objects are*,
+    and an object smaller than one output block must not vanish because the
+    sampled pixel happened to miss it.
+
+    The tail rows/columns that do not divide evenly are folded into the last
+    block rather than cropped, so an object at the right or bottom edge is kept.
+    """
+    import numpy as _np
+
+    src_h, src_w = mask.shape
+    if (src_h, src_w) == (out_h, out_w):
+        return mask
+    ys = max(1, src_h // out_h)
+    xs = max(1, src_w // out_w)
+    keep_h, keep_w = (src_h // ys) * ys, (src_w // xs) * xs
+    small = mask[:keep_h, :keep_w].reshape(
+        keep_h // ys, ys, keep_w // xs, xs).any(axis=(1, 3))
+    out = _np.zeros((out_h, out_w), bool)
+    sh, sw = min(out_h, small.shape[0]), min(out_w, small.shape[1])
+    out[:sh, :sw] = small[:sh, :sw]
+    # Fold any cropped tail into the final block so edge objects survive.
+    if keep_h < src_h and sh > 0:
+        out[sh - 1, :sw] |= mask[keep_h:, :keep_w].reshape(
+            1, src_h - keep_h, keep_w // xs, xs).any(axis=(1, 3))[0, :sw]
+    if keep_w < src_w and sw > 0:
+        out[:sh, sw - 1] |= mask[:keep_h, keep_w:].reshape(
+            keep_h // ys, ys, 1, src_w - keep_w).any(axis=(1, 3))[:sh, 0]
+    return out
+
+
 class Plot2D(_BasePlot, _PanelMixin, _MarkerMixin):
     """2-D image plot panel.
 
@@ -1115,6 +1149,13 @@ def set_overlay_mask(self, mask: "np.ndarray | None",
             Boolean array aligned to the image data.  ``True`` / non-zero
             pixels are filled with *color* at transparency *alpha*.
             Pass ``None`` to clear the overlay.
+
+            In TILE mode pass the mask at the FULL image resolution: the
+            renderer composites the mask against the OVERVIEW texture, so it
+            is reduced to the base grid here.  A mask already at the base grid
+            is accepted unchanged.  (Both are accepted because neither is
+            wrong; what was wrong was accepting only the full-resolution one
+            and then shipping bytes the renderer silently discards.)
         color : str, optional
             CSS hex colour for the overlay, e.g. ``"#ff4444"``.  Default red.
             Must be in ``#RRGGBB`` format.
@@ -1137,14 +1178,32 @@ def set_overlay_mask(self, mask: "np.ndarray | None",
             self._state["overlay_mask_alpha"] = alpha
         else:
             arr = np.asarray(mask)
-            if arr.shape != (self._state["image_height"], self._state["image_width"]):
-                raise ValueError(
-                    f"mask shape {arr.shape} does not match image "
-                    f"({self._state['image_height']} x {self._state['image_width']})"
-                )
+            ih, iw = self._state["image_height"], self._state["image_width"]
+            bh = int(self._state.get("base_height") or 0)
+            bw = int(self._state.get("base_width") or 0)
+            # THE TILE-MODE SHAPE TRAP. The renderer sizes the mask against
+            # `base_width || image_width` -- the OVERVIEW grid -- but tile mode
+            # sets image_width/height to the FULL native frame. So validating
+            # only against the image shape rejected the one shape that renders
+            # and accepted the one the renderer drops on the floor
+            # (`mBytes.length===iw*ih` else `maskCache=null`, silently). Both
+            # shapes are legal here; the overview is what actually ships.
+            tiled = bw > 0 and bh > 0 and (bh, bw) != (ih, iw)
+            if arr.shape != (ih, iw) and not (tiled and arr.shape == (bh, bw)):
+                want = f"({ih} x {iw})"
+                if tiled:
+                    want += f" or the tile overview ({bh} x {bw})"
+                raise ValueError(f"mask shape {arr.shape} does not match image {want}")
             # For origin='lower' the image data was flipped; flip mask to match.
             if self._origin == "lower":
                 arr = np.flipud(arr)
+            if tiled and arr.shape == (ih, iw):
+                # Block ANY, never a subsample: a mask marks objects that are
+                # often only a few pixels across, and striding a 4096² mask
+                # down to 1024² drops three quarters of them at random. ANY
+                # keeps every object visible at the cost of fattening it, which
+                # is the right trade for an overlay you are looking at.
+                arr = _reduce_mask_any(np.asarray(arr, dtype=bool), bh, bw)
             # Convert to uint8: True/non-zero → 255, False/zero → 0
             u8 = (np.asarray(arr, dtype=bool).view(np.uint8) * 255).astype(np.uint8)
             self._state["overlay_mask_b64"]   = base64.b64encode(u8.tobytes()).decode("ascii")
diff --git a/anyplotlib/tests/test_plot2d/test_imshow.py b/anyplotlib/tests/test_plot2d/test_imshow.py
index 711a2a7f..d7b35e5b 100644
--- a/anyplotlib/tests/test_plot2d/test_imshow.py
+++ b/anyplotlib/tests/test_plot2d/test_imshow.py
@@ -575,6 +575,60 @@ def test_set_overlay_mask_shape_mismatch(self):
         with pytest.raises(ValueError, match="mask shape"):
             plot.set_overlay_mask(bad_mask)
 
+    def test_overlay_mask_in_TILE_mode_ships_the_overview_size(self):
+        """The mask must be sized for what the RENDERER checks.
+
+        The renderer sizes the mask against ``base_width || image_width`` — the
+        OVERVIEW grid — and on a mismatch sets ``maskCache=null``: no error, no
+        overlay, nothing in the log. Tile mode sets ``image_width`` to the FULL
+        native frame, so validating a mask only against the image shape both
+        rejected the one shape that renders and accepted the one the renderer
+        drops. A caller that did the reduction itself got a ValueError; one that
+        did not got 22 MB of bytes silently discarded.
+
+        Either shape in, overview-sized bytes out.
+        """
+        import base64
+
+        plot = apl.subplots(1, 1)[1].imshow(
+            np.zeros((2048, 2048), np.uint8), tile="auto")
+        st = plot._state
+        bw, bh = st.get("base_width"), st.get("base_height")
+        assert bw and bh and (bw, bh) != (st["image_width"], st["image_height"]), \
+            "this test needs a genuinely tiled plot with a smaller base grid"
+
+        for shape in [(st["image_height"], st["image_width"]), (bh, bw)]:
+            mask = np.zeros(shape, bool)
+            mask[::5, ::5] = True
+            plot.set_overlay_mask(mask)
+            sent = base64.b64decode(st["overlay_mask_b64"])
+            assert len(sent) == bw * bh, (
+                f"a {shape} mask shipped {len(sent)} bytes; the renderer "
+                f"expects {bw * bh} and drops anything else silently")
+            assert any(sent), "the reduction emptied the mask"
+
+    def test_overlay_mask_reduction_is_block_ANY_not_a_subsample(self):
+        """A single-pixel object 4 blocks in must survive the reduction.
+
+        Objects here are often a few pixels across; a strided sample of a 2048²
+        mask at 512² keeps one pixel in sixteen and drops them at random.
+        """
+        import base64
+
+        plot = apl.subplots(1, 1)[1].imshow(
+            np.zeros((2048, 2048), np.uint8), tile="auto")
+        st = plot._state
+        bw, bh = st["base_width"], st["base_height"]
+        mask = np.zeros((st["image_height"], st["image_width"]), bool)
+        # Deliberately NOT on a block boundary — a subsample would miss it.
+        mask[537, 921] = True
+        plot.set_overlay_mask(mask)
+        sent = np.frombuffer(base64.b64decode(st["overlay_mask_b64"]), np.uint8)
+        # Bytes are 0/255, so COUNT the lit ones rather than summing them.
+        assert np.count_nonzero(sent.reshape(bh, bw)) == 1, (
+            "the single lit pixel did not survive the reduction — this is a "
+            "subsample, not a block ANY")
+
     def test_set_overlay_mask_alpha_boundary(self):
         plot = _img(n=16)
         mask = np.zeros((16, 16), dtype=bool)
diff --git a/upcoming_changes/+overlay_mask_tile.bugfix.rst b/upcoming_changes/+overlay_mask_tile.bugfix.rst
new file mode 100644
index 00000000..7f06a1e4
--- /dev/null
+++ b/upcoming_changes/+overlay_mask_tile.bugfix.rst
@@ -0,0 +1,10 @@
+``Plot2D.set_overlay_mask`` now works in TILE mode. The renderer sizes the mask
+against ``base_width || image_width`` -- the tile overview grid -- but tile mode
+sets ``image_width`` to the full native frame, so the shape check accepted only
+the one shape the renderer silently discards (``maskCache = null``, no error) and
+rejected the one that actually renders. On a 4096x4096 tiled plot neither a
+1024x1024 nor a 4096x4096 mask could be drawn: the first raised ``ValueError``,
+the second encoded 22.4 MB the renderer dropped. Both shapes are now accepted and
+a full-resolution mask is reduced to the overview grid with a block ANY -- never
+a subsample, so an object a few pixels across cannot vanish into a skipped
+sample.