diff --git a/comfy/sd.py b/comfy/sd.py
index 5fed4ca9a..52be28137 100644
--- a/comfy/sd.py
+++ b/comfy/sd.py
@@ -1215,6 +1215,9 @@ class VAE:
                     if self.handles_tiling:
                         tile = 256 // self.spacial_compression_decode()
                         overlap = tile // 4
+                        # decode_tiled is identical to decode for these VAEs, so freeing the
+                        # memory other models hold onto is the only thing that can make the retry succeed.
+                        model_management.free_memory(1e30, self.device, keep_loaded=[model_management.LoadedModel(self.patcher)])
                         pixel_samples = self._decode_tiled_owned(samples_in, tile_x=tile, tile_y=tile, overlap=overlap)
                     else:
                         pixel_samples = self.decode_tiled_(samples_in)
@@ -1222,6 +1225,9 @@ class VAE:
                     tile = 256 // self.spacial_compression_decode()
                     overlap = tile // 4
                     if self.handles_tiling:
+                        # decode_tiled is identical to decode for these VAEs, so freeing the
+                        # memory other models hold onto is the only thing that can make the retry succeed.
+                        model_management.free_memory(1e30, self.device, keep_loaded=[model_management.LoadedModel(self.patcher)])
                         pixel_samples = self._decode_tiled_owned(samples_in, tile_x=tile, tile_y=tile, overlap=overlap)
                     else:
                         pixel_samples = self.decode_tiled_3d(samples_in, tile_x=tile, tile_y=tile, overlap=(1, overlap, overlap))
diff --git a/tests-unit/comfy_test/test_seedvr2_vae_tiled.py b/tests-unit/comfy_test/test_seedvr2_vae_tiled.py
index d64f51918..26141c644 100644
--- a/tests-unit/comfy_test/test_seedvr2_vae_tiled.py
+++ b/tests-unit/comfy_test/test_seedvr2_vae_tiled.py
@@ -392,6 +392,35 @@ def test_seedvr2_3d_routes_to_owned_encode_tiled_on_oom():
     )
 
 
+def test_handles_tiling_decode_oom_frees_memory_before_retry():
+    wrapper = seedvr_vae_mod.VideoAutoencoderKLWrapper.__new__(
+        seedvr_vae_mod.VideoAutoencoderKLWrapper)
+    vae = _make_vae(wrapper, latent_channels=_LATENT_CHANNELS, latent_dim=3)
+
+    call_order = []
+    free_memory_call = MagicMock(side_effect=lambda *a, **k: call_order.append("free_memory"))
+    seedvr2_call = MagicMock(
+        side_effect=lambda *a, **k: call_order.append("decode_tiled_owned") or torch.zeros(1, 3, 9, 64, 64))
+
+    mm = sd_mod.model_management
+    with ExitStack() as stack:
+        stack.enter_context(patch.object(mm, "raise_non_oom", lambda e: None))
+        stack.enter_context(patch.object(mm, "load_models_gpu", lambda *a, **k: None))
+        stack.enter_context(patch.object(mm, "soft_empty_cache", lambda: None))
+        stack.enter_context(patch.object(mm, "free_memory", free_memory_call))
+        stack.enter_context(patch.object(sd_mod.VAE, "_decode_tiled_owned", seedvr2_call))
+        stack.enter_context(patch.object(
+            seedvr_vae_mod.VideoAutoencoderKLWrapper, "decode",
+            side_effect=_force_oom))
+        vae.decode(torch.zeros(1, _LATENT_CHANNELS * 3, 8, 8))
+
+    assert call_order == ["free_memory", "decode_tiled_owned"], (
+        "decode_tiled retry for a handles_tiling VAE must free other models' "
+        f"memory before retrying; call order was {call_order}"
+    )
+    assert free_memory_call.call_args.kwargs["keep_loaded"][0].model is vae.patcher
+
+
 def test_non_seedvr2_encode_tiled_3d_default_overlap_is_concrete():
     vae = _make_non_seedvr2_vae_fallback()
     vae.downscale_ratio = (lambda a: max(1, a // 4), 8, 8)