fix: normalize NumPy image outputs

Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com>

examples/offline_inference/text_to_image/text_to_image.py

@@ -604,18 +604,7 @@ def main():
         raise ValueError("No output generated from omni.generate()")
     logger.info(f"Outputs: {outputs}")
 
-    images = None
-    for output in outputs:
-        images = getattr(output, "images", None)
-        if images:
-            break
-        req_out = getattr(output, "request_output", None)
-        images = getattr(req_out, "images", None) if req_out is not None else None
-        if images:
-            break
-
-    if not images:
-        images = extract_images_from_outputs(outputs)
+    images = extract_images_from_outputs(outputs)
 
     if not images:
         raise ValueError("No images found in request_output")

tests/diffusion/test_image_output.py

@@ -0,0 +1,37 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import numpy as np
+import pytest
+from PIL import Image
+
+from vllm_omni.diffusion.utils.image_output import extract_images_from_outputs
+from vllm_omni.outputs import OmniRequestOutput
+
+pytestmark = [pytest.mark.diffusion, pytest.mark.core_model, pytest.mark.cpu]
+
+
+@pytest.mark.parametrize(
+    ("payload", "expected_pixel"),
+    [
+        (np.array([[[-1.0, 0.0, 1.0]]], dtype=np.float32), (0, 127, 255)),
+        (np.array([[[0, 128, 255]]], dtype=np.uint8), (0, 128, 255)),
+    ],
+)
+def test_extract_images_normalizes_numpy_payload(
+    payload: np.ndarray,
+    expected_pixel: tuple[int, int, int],
+    tmp_path,
+) -> None:
+    output = OmniRequestOutput(images=[payload])
+
+    [image] = extract_images_from_outputs(output)
+
+    assert isinstance(image, Image.Image)
+    assert image.mode == "RGB"
+    assert image.size == (1, 1)
+    assert image.getpixel((0, 0)) == expected_pixel
+
+    output_path = tmp_path / "image.png"
+    image.save(output_path)
+    assert output_path.is_file()

vllm_omni/diffusion/utils/image_output.py

@@ -4,6 +4,7 @@
 from collections.abc import Iterable, Mapping, Sequence
 from typing import Any
 
+import numpy as np
 import torch
 from PIL import Image
 from vllm.outputs import RequestOutput
@@ -20,6 +21,8 @@ def extract_images_from_outputs(outputs: ImageOutputs) -> list[Image.Image]:
     Supported inputs:
     - OmniRequestOutput / list[OmniRequestOutput]
     - RequestOutput / list[RequestOutput] (compatibility fallback)
+
+    PIL images, NumPy arrays, and PyTorch tensors are normalized to PIL images.
     """
     for output in _iter_known_outputs(outputs):
         if isinstance(output, OmniRequestOutput):
@@ -58,6 +61,9 @@ def _iter_multimodal_image_payloads(output: OmniRequestOutput | RequestOutput) -
 
 
 def _iter_request_output_payloads(request_output: RequestOutput) -> Iterable[Any]:
+    images = getattr(request_output, "images", None)
+    if images is not None:
+        yield images
     mm = getattr(request_output, "multimodal_output", None)
     if mm is not None:
         yield from _image_values_from_mapping_like(mm)
@@ -82,6 +88,8 @@ def _coerce_images(payload: Any) -> list[Image.Image]:
         return []
     if isinstance(payload, Image.Image):
         return [payload]
+    if isinstance(payload, np.ndarray):
+        return _ndarray_to_images(payload)
     if isinstance(payload, torch.Tensor):
         return _tensor_to_images(payload)
     if isinstance(payload, list | tuple):
@@ -92,6 +100,37 @@ def _coerce_images(payload: Any) -> list[Image.Image]:
     return []
 
 
+def _ndarray_to_images(array: np.ndarray) -> list[Image.Image]:
+    if array.ndim == 4:
+        images: list[Image.Image] = []
+        for single in array:
+            images.extend(_ndarray_to_images(single))
+        return images
+    if array.ndim != 3 or array.size == 0:
+        return []
+
+    image = array
+    if image.shape[-1] not in (1, 3, 4):
+        if image.shape[0] not in (1, 3, 4):
+            return []
+        image = np.moveaxis(image, 0, -1)
+
+    if np.issubdtype(image.dtype, np.floating):
+        image = image.astype(np.float32, copy=False)
+        if image.min() < 0.0:
+            image = image / 2 + 0.5
+        image = np.clip(image, 0.0, 1.0) * 255
+    elif np.issubdtype(image.dtype, np.integer) or np.issubdtype(image.dtype, np.bool_):
+        image = np.clip(image, 0, 255)
+    else:
+        return []
+
+    image = image.astype(np.uint8)
+    if image.shape[-1] == 1:
+        image = image[..., 0]
+    return [Image.fromarray(np.ascontiguousarray(image))]
+
+
 def _tensor_to_images(tensor: torch.Tensor) -> list[Image.Image]:
     img = tensor.detach().to("cpu", dtype=torch.float32)
     if img.ndim == 4: