diff --git a/vllm_omni/metrics/stats.py b/vllm_omni/metrics/stats.py
index 48e394add..46fec6fcc 100644
--- a/vllm_omni/metrics/stats.py
+++ b/vllm_omni/metrics/stats.py
@@ -226,8 +226,10 @@ class OrchestratorAggregator:
                 and (multimodal_output := output_to_yield.multimodal_output.get("audio")) is not None
                 and len(multimodal_output) > 0
             ):
-                last = multimodal_output[-1]
-                nframes = int(last.shape[0]) if last.ndim > 0 else 1
+                nframes = sum(
+                    int(t.shape[0]) if t.ndim > 0 else 1
+                    for t in (multimodal_output if isinstance(multimodal_output, list) else [multimodal_output])
+                )
                 stage_events_for_req = self.stage_events.get(request_id, [])
                 if stage_events_for_req:
                     for stage_event in stage_events_for_req:
diff --git a/vllm_omni/outputs.py b/vllm_omni/outputs.py
index 3708bc8f9..5bbd27a1d 100644
--- a/vllm_omni/outputs.py
+++ b/vllm_omni/outputs.py
@@ -127,15 +127,17 @@ class OmniRequestOutput:
         For pipeline outputs, this checks completion outputs first, then request_output.
         For diffusion outputs, this returns the local _multimodal_output field.
         """
-        if self.request_output is not None:
-            # Check completion outputs first (where multimodal_output is attached).
-            outputs = getattr(self.request_output, "outputs", None)
-            if isinstance(outputs, list) and outputs:
-                for output in outputs:
-                    mm = getattr(output, "multimodal_output", None)
-                    if mm:
-                        return mm
-            return getattr(self.request_output, "multimodal_output", {})
+        if self.request_output is None:
+            return self._multimodal_output
+
+        request_outputs = self.request_output if isinstance(self.request_output, list) else [self.request_output]
+        for req_out in request_outputs:
+            # Check completion outputs first (where multimodal_output is attached)
+            for output in getattr(req_out, "outputs", []):
+                if mm := getattr(output, "multimodal_output", None):
+                    return mm
+            if mm := getattr(req_out, "multimodal_output", None):
+                return mm
         return self._multimodal_output
 
     @property