diff --git a/tests/inputs/test_diffusion_sampling_params.py b/tests/inputs/test_diffusion_sampling_params.py
new file mode 100644
index 000000000..5daa1fd49
--- /dev/null
+++ b/tests/inputs/test_diffusion_sampling_params.py
@@ -0,0 +1,65 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Regression tests for diffusion sampling params coercion."""
+
+from __future__ import annotations
+
+from typing import Any
+
+import pytest
+from vllm.sampling_params import SamplingParams
+
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+
+pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
+
+
+def _sampling_params(*, seed: int | None = None, extra_args: dict[str, Any] | None = None) -> SamplingParams:
+    return SamplingParams(max_tokens=1, seed=seed, extra_args=extra_args)
+
+
+def test_from_params_returns_existing_omni_params() -> None:
+    params = OmniDiffusionSamplingParams(num_inference_steps=20)
+
+    converted = OmniDiffusionSamplingParams.from_params(params)
+
+    assert converted is params
+
+
+def test_from_params_converts_sampling_params_seed_and_known_extra_args() -> None:
+    params = _sampling_params(
+        seed=1234,
+        extra_args={
+            "num_inference_steps": 50,
+            "height": 1024,
+            "width": 768,
+            "guidance_scale": 7.5,
+        },
+    )
+
+    converted = OmniDiffusionSamplingParams.from_params(params)
+
+    assert converted.seed == 1234
+    assert converted.num_inference_steps == 50
+    assert converted.height == 1024
+    assert converted.width == 768
+    assert converted.guidance_scale == 7.5
+    assert converted.extra_args == {}
+
+
+def test_from_params_preserves_unknown_extra_args() -> None:
+    params = _sampling_params(
+        extra_args={
+            "height": 512,
+            "pipeline_specific": "kept",
+        },
+    )
+
+    converted = OmniDiffusionSamplingParams.from_params(params)
+
+    assert converted.height == 512
+    assert converted.extra_args == {"pipeline_specific": "kept"}
+
+
+def test_from_params_rejects_unsupported_types() -> None:
+    with pytest.raises(TypeError, match="Diffusion stage requires OmniDiffusionSamplingParams"):
+        OmniDiffusionSamplingParams.from_params({"height": 512})
diff --git a/vllm_omni/engine/stage_pool.py b/vllm_omni/engine/stage_pool.py
index 487134db1..820b28510 100644
--- a/vllm_omni/engine/stage_pool.py
+++ b/vllm_omni/engine/stage_pool.py
@@ -937,10 +937,9 @@ class StagePool:
     ) -> int:
         """Submit a stage-entry request into this pool."""
         params = params_override if params_override is not None else req_state.sampling_params_list[self.stage_id]
-        # Convert plain vllm SamplingParams for single-stage diffusion models
-        # that receive sampling params from the user/caller directly.
-        if self.stage_type == "diffusion" and not isinstance(params, OmniDiffusionSamplingParams):
-            params = OmniDiffusionSamplingParams()
+        # Direct engine callers may provide plain vLLM SamplingParams.
+        if self.stage_type == "diffusion":
+            params = OmniDiffusionSamplingParams.from_params(params)
         submit_kwargs = dict(submit_kwargs or {})
         if self.stage_type == "diffusion":
             if isinstance(request, list):
@@ -1003,8 +1002,8 @@ class StagePool:
     ) -> int:
         """Submit a streaming update to an already admitted request."""
         params = req_state.sampling_params_list[self.stage_id]
-        if self.stage_type == "diffusion" and not isinstance(params, OmniDiffusionSamplingParams):
-            params = OmniDiffusionSamplingParams()
+        if self.stage_type == "diffusion":
+            params = OmniDiffusionSamplingParams.from_params(params)
         replica_id = self.get_bound_replica_id(request_id)
         if replica_id is None or self.clients[replica_id] is None:
             replica_id = await self._pick_or_select(request_id)
diff --git a/vllm_omni/inputs/data.py b/vllm_omni/inputs/data.py
index 6626176b6..61d3d12a3 100644
--- a/vllm_omni/inputs/data.py
+++ b/vllm_omni/inputs/data.py
@@ -380,5 +380,30 @@ class OmniDiffusionSamplingParams:
     def clone(self) -> "OmniDiffusionSamplingParams":
         return copy.deepcopy(self)
 
+    @classmethod
+    def from_params(cls, params: Any) -> "OmniDiffusionSamplingParams":
+        """Normalize caller params for diffusion stages.
+
+        Existing Omni params are returned unchanged. Plain vLLM SamplingParams
+        copy seed and matching extra_args fields; remaining extra_args stay nested.
+        Unsupported types raise TypeError.
+        """
+        if isinstance(params, cls):
+            return params
+        if isinstance(params, SamplingParams):
+            extra = dict(getattr(params, "extra_args", None) or {})
+            known = set(cls.__dataclass_fields__)
+            mapped = {k: extra.pop(k) for k in list(extra) if k in known}
+            if extra:
+                mapped["extra_args"] = extra
+            seed = getattr(params, "seed", None)
+            if seed is not None:
+                mapped.setdefault("seed", seed)
+            return cls(**mapped)
+        raise TypeError(
+            "Diffusion stage requires OmniDiffusionSamplingParams or vllm.SamplingParams, "
+            f"got {type(params).__name__!r}."
+        )
+
 
 OmniSamplingParams: TypeAlias = SamplingParams | OmniDiffusionSamplingParams