[Bugfix][Multimodal] Normalize direct PIL image inputs (#47566)
Signed-off-by: Ting Sun <suntcrick@gmail.com> Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>
Ting SUN committed
Jul 4, 2026 at 04:27 UTC
379950191f9d229a44c75d0256d95677fa34519c
1 file changed
+8
vllm/multimodal/parse.py
+8
@@ -24,6 +24,7 @@ from vllm.utils.collection_utils import is_list_of
24
from vllm.utils.import_utils import LazyLoader
25
26
from .audio import AudioResampler, AudioSpec, normalize_audio
27
+from .image import convert_image_mode, normalize_image
28
from .inputs import (
29
AudioItem,
30
HfAudioItem,
@@ -621,6 +622,13 @@ class MultiModalDataParser:
622
else:
623
data_items = data
624
625
+ data_items = [
626
+ convert_image_mode(normalize_image(item), "RGB")
627
+ if isinstance(item, PILImage.Image)
628
+ else item
629
+ for item in data_items
630
+ ]
631
+
632
return ImageProcessorItems(data_items)
633
634
def _parse_video_data(