[Security] Fix image EXIF orientation and tRNS transparency handling (#44974)

Signed-off-by: jperezde <jperezde@redhat.com>
This commit is contained in:
Juan Pérez de Algaba
2026-06-09 09:34:44 -07:00
committed by GitHub
parent 766ce2bb6b
commit cf1c906724
3 changed files with 121 additions and 9 deletions
+86 -1
View File
@@ -1,12 +1,18 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import struct
from io import BytesIO
from pathlib import Path
import numpy as np
import pytest
from PIL import Image, ImageChops
from vllm.multimodal.image import convert_image_mode
from vllm.multimodal.image import (
_has_transparency,
convert_image_mode,
normalize_image,
)
pytestmark = pytest.mark.cpu_test
@@ -38,3 +44,82 @@ def test_rgba_to_rgb():
assert converted_image_numpy[i][j][0] == 255
assert converted_image_numpy[i][j][1] == 255
assert converted_image_numpy[i][j][2] == 255
def test_palette_with_trns_to_rgb():
"""P-mode PNG with tRNS: transparent index should become white."""
# Built synthetically because the existing assets are RGBA PNGs;
# this vulnerability only affects P/L/RGB images carrying a tRNS
# chunk, which uses a different transparency mechanism.
img = Image.new("P", (4, 4))
palette = [0] * 768
palette[0:3] = [255, 0, 0]
palette[3:6] = [0, 0, 255]
img.putpalette(palette)
img.putpixel((0, 0), 0)
img.putpixel((1, 0), 1)
img.info["transparency"] = 1
assert _has_transparency(img)
converted = convert_image_mode(img, "RGB")
assert converted.mode == "RGB"
r, g, b = converted.getpixel((1, 0))
assert (r, g, b) == (255, 255, 255)
r, g, b = converted.getpixel((0, 0))
assert (r, g, b) == (255, 0, 0)
def test_l_mode_no_trns_to_rgb():
"""L-mode without transparency should convert directly."""
img = Image.new("L", (4, 4), 128)
assert not _has_transparency(img)
converted = convert_image_mode(img, "RGB")
assert converted.mode == "RGB"
assert converted.getpixel((0, 0)) == (128, 128, 128)
def test_exif_transpose_normalizes_orientation():
"""Image with EXIF orientation 3 (180-degree rotation) should be
normalized so pixel data matches visual display."""
# Built synthetically because the existing assets are PNGs which
# don't carry EXIF orientation metadata; we need a JPEG with an
# injected EXIF orientation=3 tag.
img = Image.new("RGB", (2, 1))
img.putpixel((0, 0), (255, 0, 0))
img.putpixel((1, 0), (0, 0, 255))
buf = BytesIO()
img.save(buf, format="JPEG")
jpeg_bytes = buf.getvalue()
exif_orientation_3 = (
b"\xff\xe1"
+ struct.pack(">H", 26)
+ b"Exif\x00\x00"
+ b"MM"
+ b"\x00\x2a"
+ b"\x00\x00\x00\x08"
+ b"\x00\x01"
+ b"\x01\x12"
+ b"\x00\x03"
+ b"\x00\x00\x00\x01"
+ b"\x00\x03"
+ b"\x00\x00"
+ b"\x00\x00\x00\x00"
)
soi = jpeg_bytes[:2]
rest = jpeg_bytes[2:]
patched = soi + exif_orientation_3 + rest
rotated = Image.open(BytesIO(patched))
normalized = normalize_image(rotated)
assert normalized.size == (2, 1)
def test_normalize_image_no_exif():
"""Images without EXIF should pass through unchanged."""
img = Image.new("RGB", (4, 4), (100, 100, 100))
result = normalize_image(img)
assert result.size == img.size
assert result.mode == img.mode
+30 -6
View File
@@ -1,7 +1,9 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from PIL import Image
import contextlib
from PIL import Image, ImageOps
def rescale_image_size(
@@ -16,6 +18,13 @@ def rescale_image_size(
return image
def normalize_image(image: Image.Image) -> Image.Image:
"""Normalize EXIF orientation so the pixel data matches visual display."""
with contextlib.suppress(Exception):
image = ImageOps.exif_transpose(image)
return image
def rgba_to_rgb(
image: Image.Image,
background_color: tuple[int, int, int] | list[int] = (255, 255, 255),
@@ -27,10 +36,25 @@ def rgba_to_rgb(
return converted
def convert_image_mode(image: Image.Image, to_mode: str):
def _has_transparency(image: Image.Image) -> bool:
"""Detect whether an image carries transparency data (RGBA, LA, PA,
or tRNS chunk in P/L/RGB PNGs)."""
if image.mode in ("RGBA", "LA", "PA"):
return True
return "transparency" in getattr(image, "info", {})
def convert_image_mode(
image: Image.Image,
to_mode: str,
background_color: tuple[int, int, int] | list[int] = (255, 255, 255),
) -> Image.Image:
if image.mode == to_mode:
return image
elif image.mode == "RGBA" and to_mode == "RGB":
return rgba_to_rgb(image)
else:
return image.convert(to_mode)
if to_mode == "RGB" and _has_transparency(image):
if image.mode != "RGBA":
image = image.convert("RGBA")
return rgba_to_rgb(image, background_color)
return image.convert(to_mode)
+5 -2
View File
@@ -11,7 +11,7 @@ from PIL import Image
from vllm.utils.serial_utils import tensor2base64
from ..image import convert_image_mode, rgba_to_rgb
from ..image import convert_image_mode, normalize_image, rgba_to_rgb
from .base import MediaIO, MediaWithBytes
MAGIC_NUMPY_PREFIX = b"\x93NUMPY" # https://numpy.org/devdocs/reference/generated/numpy.lib.format.html#format-version-1-0
@@ -65,11 +65,14 @@ class ImageMediaIO(MediaIO[Image.Image]):
elif image.mode == "RGBA" and self.image_mode == "RGB":
return rgba_to_rgb(image, self.rgba_background_color)
else:
return convert_image_mode(image, self.image_mode)
return convert_image_mode(
image, self.image_mode, self.rgba_background_color
)
def load_bytes(self, data: bytes) -> MediaWithBytes[Image.Image]:
try:
image = Image.open(BytesIO(data))
image = normalize_image(image)
image.load()
image = self._convert_image_mode(image)
except (OSError, Image.UnidentifiedImageError) as e: