diff --git a/vllm_omni/engine/output_processor.py b/vllm_omni/engine/output_processor.py
index 369fe3c7b..5c0b752b3 100644
--- a/vllm_omni/engine/output_processor.py
+++ b/vllm_omni/engine/output_processor.py
@@ -269,6 +269,10 @@ class OmniRequestState(RequestState):
         # tokens, so we always store a cumulative copy here.
         base_output.cumulative_token_ids = list(self.detokenizer.output_token_ids)
 
+        # Attach cumulative_text only at the final step for inter-stage use.
+        if finish_reason is not None and hasattr(self.detokenizer, "output_text"):
+            base_output.cumulative_text = self.detokenizer.output_text
+
         if not hasattr(base_output, "multimodal_output"):
             setattr(base_output, "multimodal_output", {})
         if self.mm_accumulated:
diff --git a/vllm_omni/model_executor/stage_input_processors/hunyuan_image3.py b/vllm_omni/model_executor/stage_input_processors/hunyuan_image3.py
index 749e213e0..365cf0641 100644
--- a/vllm_omni/model_executor/stage_input_processors/hunyuan_image3.py
+++ b/vllm_omni/model_executor/stage_input_processors/hunyuan_image3.py
@@ -185,7 +185,8 @@ def ar2diffusion(
     for i, ar_output in enumerate(ar_outputs):
         output = ar_output.outputs[0]
         generated_token_ids = output.cumulative_token_ids
-        generated_text = getattr(output, "text", "") or ""
+        # Prefer cumulative_text, fallback to text if aggregation dropped it
+        generated_text = getattr(output, "cumulative_text", None) or getattr(output, "text", "") or ""
 
         # Get original prompt info
         original_prompt = prompt[i] if i < len(prompt) else {}