Add Fast Image Processor for Donut

docs/source/en/model_doc/donut.md

@@ -197,6 +197,11 @@ We refer to the [tutorial notebooks](https://github.com/NielsRogge/Transformers-
 [[autodoc]] DonutImageProcessor
     - preprocess
 
+## DonutImageProcessorFast
+
+[[autodoc]] DonutImageProcessorFast
+    - preprocess
+
 ## DonutFeatureExtractor
 
 [[autodoc]] DonutFeatureExtractor

src/transformers/__init__.py

@@ -1350,6 +1350,7 @@
     _import_structure["models.deit"].append("DeiTImageProcessorFast")
     _import_structure["models.depth_pro"].append("DepthProImageProcessorFast")
     _import_structure["models.detr"].append("DetrImageProcessorFast")
+    _import_structure["models.donut"].append("DonutImageProcessorFast")
     _import_structure["models.gemma3"].append("Gemma3ImageProcessorFast")
     _import_structure["models.got_ocr2"].append("GotOcr2ImageProcessorFast")
     _import_structure["models.llava"].append("LlavaImageProcessorFast")
@@ -6609,6 +6610,7 @@
         from .models.deit import DeiTImageProcessorFast
         from .models.depth_pro import DepthProImageProcessorFast
         from .models.detr import DetrImageProcessorFast
+        from .models.donut import DonutImageProcessorFast
         from .models.gemma3 import Gemma3ImageProcessorFast
         from .models.got_ocr2 import GotOcr2ImageProcessorFast
         from .models.llava import LlavaImageProcessorFast

src/transformers/models/auto/image_processing_auto.py

@@ -79,7 +79,7 @@
             ("detr", ("DetrImageProcessor", "DetrImageProcessorFast")),
             ("dinat", ("ViTImageProcessor", "ViTImageProcessorFast")),
             ("dinov2", ("BitImageProcessor",)),
-            ("donut-swin", ("DonutImageProcessor",)),
+            ("donut-swin", ("DonutImageProcessor", "DonutImageProcessorFast")),
             ("dpt", ("DPTImageProcessor",)),
             ("efficientformer", ("EfficientFormerImageProcessor",)),
             ("efficientnet", ("EfficientNetImageProcessor",)),

src/transformers/models/donut/__init__.py

@@ -21,6 +21,7 @@
     from .configuration_donut_swin import *
     from .feature_extraction_donut import *
     from .image_processing_donut import *
+    from .image_processing_donut_fast import *
     from .modeling_donut_swin import *
     from .processing_donut import *
 else:

src/transformers/models/donut/image_processing_donut_fast.py

@@ -0,0 +1,297 @@
+# coding=utf-8
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Fast Image processor class for Donut."""
+
+from typing import Optional, Union
+
+import numpy as np
+
+from ...image_processing_utils_fast import (
+    BASE_IMAGE_PROCESSOR_FAST_DOCSTRING,
+    BASE_IMAGE_PROCESSOR_FAST_DOCSTRING_PREPROCESS,
+    BaseImageProcessorFast,
+    BatchFeature,
+    DefaultFastImageProcessorKwargs,
+)
+from ...image_transforms import ChannelDimension, group_images_by_shape, reorder_images
+from ...image_utils import (
+    IMAGENET_STANDARD_MEAN,
+    IMAGENET_STANDARD_STD,
+    ImageInput,
+    PILImageResampling,
+    SizeDict,
+    get_image_size,
+)
+from ...processing_utils import Unpack
+from ...utils import (
+    TensorType,
+    add_start_docstrings,
+    is_torch_available,
+    is_torchvision_available,
+    is_torchvision_v2_available,
+    logging,
+)
+
+
+logger = logging.get_logger(__name__)
+
+if is_torch_available():
+    import torch
+
+if is_torchvision_available():
+    if is_torchvision_v2_available():
+        from torchvision.transforms.v2 import functional as F
+    else:
+        from torchvision.transforms import functional as F
+
+
+class DonutFastImageProcessorKwargs(DefaultFastImageProcessorKwargs):
+    do_thumbnail: Optional[bool]
+    do_align_long_axis: Optional[bool]
+    do_pad: Optional[bool]
+
+
+@add_start_docstrings(
+    "Constructs a fast Donut image processor.",
+    BASE_IMAGE_PROCESSOR_FAST_DOCSTRING,
+    """
+        do_thumbnail (`bool`, *optional*, defaults to `self.do_thumbnail`):
+            Whether to resize the image using thumbnail method.
+        do_align_long_axis (`bool`, *optional*, defaults to `self.do_align_long_axis`):
+            Whether to align the long axis of the image with the long axis of `size` by rotating by 90 degrees.
+        do_pad (`bool`, *optional*, defaults to `self.do_pad`):
+            Whether to pad the image. If `random_padding` is set to `True`, each image is padded with a random
+            amount of padding on each size, up to the largest image size in the batch. Otherwise, all images are
+            padded to the largest image size in the batch.
+    """,
+)
+class DonutImageProcessorFast(BaseImageProcessorFast):
+    # This generated class can be used as a starting point for the fast image processor.
+    # if the image processor is only used for simple augmentations, such as resizing, center cropping, rescaling, or normalizing,
+    # only the default values should be set in the class.
+    # If the image processor requires more complex augmentations, methods from BaseImageProcessorFast can be overridden.
+    # In most cases, only the `_preprocess` method should be overridden.
+
+    # For an example of a fast image processor requiring more complex augmentations, see `LlavaNextImageProcessorFast`.
+
+    # Default values should be checked against the slow image processor
+    # None values left after checking can be removed
+    resample = PILImageResampling.BILINEAR
+    image_mean = IMAGENET_STANDARD_MEAN
+    image_std = IMAGENET_STANDARD_STD
+    size = {"height": 2560, "width": 1920}
+    default_to_square = None
+    crop_size = None
+    do_resize = True
+    do_center_crop = None
+    do_rescale = True
+    do_normalize = True
+    do_convert_rgb = None
+    do_thumbnail = True
+    do_align_long_axis = False
+    do_pad = True
+    valid_kwargs = DonutFastImageProcessorKwargs
+
+    def __init__(self, **kwargs: Unpack[DonutFastImageProcessorKwargs]):
+        super().__init__(**kwargs)
+
+    @add_start_docstrings(
+        BASE_IMAGE_PROCESSOR_FAST_DOCSTRING_PREPROCESS,
+        """
+            do_thumbnail (`bool`, *optional*, defaults to `self.do_thumbnail`):
+                Whether to resize the image using thumbnail method.
+            do_align_long_axis (`bool`, *optional*, defaults to `self.do_align_long_axis`):
+                Whether to align the long axis of the image with the long axis of `size` by rotating by 90 degrees.
+            do_pad (`bool`, *optional*, defaults to `self.do_pad`):
+                Whether to pad the image. If `random_padding` is set to `True`, each image is padded with a random
+                amount of padding on each size, up to the largest image size in the batch. Otherwise, all images are
+                padded to the largest image size in the batch.
+        """,
+    )
+    def preprocess(self, images: ImageInput, **kwargs: Unpack[DonutFastImageProcessorKwargs]) -> BatchFeature:
+        return super().preprocess(images, **kwargs)
+
+    def align_long_axis(
+        self,
+        image: "torch.Tensor",
+        size: SizeDict,
+    ) -> "torch.Tensor":
+        """
+        Align the long axis of the image to the longest axis of the specified size.
+
+        Args:
+            image (`torch.Tensor`):
+                The image to be aligned.
+            size (`Dict[str, int]`):
+                The size `{"height": h, "width": w}` to align the long axis to.
+
+        Returns:
+            `torch.Tensor`: The aligned image.
+        """
+        input_height, input_width = get_image_size(image, channel_dim=ChannelDimension.FIRST)
+        output_height, output_width = size.height, size.width
+
+        if (output_width < output_height and input_width > input_height) or (
+            output_width > output_height and input_width < input_height
+        ):
+            height_dim, width_dim = image.dim() - 2, image.dim() - 1
+            image = torch.rot90(image, 3, dims=[height_dim, width_dim])
+
+        return image
+
+    def pad_image(
+        self,
+        image: "torch.Tensor",
+        size: SizeDict,
+        random_padding: bool = False,
+    ) -> "torch.Tensor":
+        """
+        Pad the image to the specified size.
+
+        Args:
+            image (`torch.Tensor`):
+                The image to be padded.
+            size (`Dict[str, int]`):
+                The size `{"height": h, "width": w}` to pad the image to.
+            random_padding (`bool`, *optional*, defaults to `False`):
+                Whether to use random padding or not.
+            data_format (`str` or `ChannelDimension`, *optional*):
+                The data format of the output image. If unset, the same format as the input image is used.
+            input_data_format (`ChannelDimension` or `str`, *optional*):
+                The channel dimension format of the input image. If not provided, it will be inferred.
+        """
+        output_height, output_width = size.height, size.width
+        input_height, input_width = get_image_size(image, channel_dim=ChannelDimension.FIRST)
+
+        delta_width = output_width - input_width
+        delta_height = output_height - input_height
+
+        if random_padding:
+            pad_top = np.random.randint(low=0, high=delta_height + 1)
+            pad_left = np.random.randint(low=0, high=delta_width + 1)
+        else:
+            pad_top = delta_height // 2
+            pad_left = delta_width // 2
+
+        pad_bottom = delta_height - pad_top
+        pad_right = delta_width - pad_left
+
+        padding = (pad_left, pad_top, pad_right, pad_bottom)
+        return F.pad(image, padding)
+
+    def pad(self, *args, **kwargs):
+        logger.info("pad is deprecated and will be removed in version 4.27. Please use pad_image instead.")
+        return self.pad_image(*args, **kwargs)
+
+    def thumbnail(
+        self,
+        image: "torch.Tensor",
+        size: SizeDict,
+    ) -> "torch.Tensor":
+        """
+        Resize the image to make a thumbnail. The image is resized so that no dimension is larger than any
+        corresponding dimension of the specified size.
+
+        Args:
+            image (`np.ndarray`):
+                The image to be resized.
+            size (`Dict[str, int]`):
+                The size `{"height": h, "width": w}` to resize the image to.
+            resample (`PILImageResampling`, *optional*, defaults to `PILImageResampling.BICUBIC`):
+                The resampling filter to use.
+            data_format (`Optional[Union[str, ChannelDimension]]`, *optional*):
+                The data format of the output image. If unset, the same format as the input image is used.
+            input_data_format (`ChannelDimension` or `str`, *optional*):
+                The channel dimension format of the input image. If not provided, it will be inferred.
+        """
+        input_height, input_width = get_image_size(image, channel_dim=ChannelDimension.FIRST)
+        output_height, output_width = size.height, size.width
+
+        # We always resize to the smallest of either the input or output size.
+        height = min(input_height, output_height)
+        width = min(input_width, output_width)
+
+        if height == input_height and width == input_width:
+            return image
+
+        if input_height > input_width:
+            width = int(input_width * height / input_height)
+        elif input_width > input_height:
+            height = int(input_height * width / input_width)
+
+        return self.resize(
+            image,
+            size=SizeDict(width=width, height=height),
+            interpolation=F.InterpolationMode.BICUBIC,
+        )
+
+    def _preprocess(
+        self,
+        images: list["torch.Tensor"],
+        do_resize: bool,
+        do_thumbnail: bool,
+        do_align_long_axis: bool,
+        do_pad: bool,
+        size: SizeDict,
+        interpolation: Optional["F.InterpolationMode"],
+        do_center_crop: bool,
+        crop_size: SizeDict,
+        do_rescale: bool,
+        rescale_factor: float,
+        do_normalize: bool,
+        image_mean: Optional[Union[float, list[float]]],
+        image_std: Optional[Union[float, list[float]]],
+        return_tensors: Optional[Union[str, TensorType]],
+        **kwargs,
+    ) -> BatchFeature:
+        # Group images by size for batched resizing
+        grouped_images, grouped_images_index = group_images_by_shape(images)
+        resized_images_grouped = {}
+        for shape, stacked_images in grouped_images.items():
+            if do_align_long_axis:
+                stacked_images = self.align_long_axis(image=stacked_images, size=size)
+            if do_resize:
+                shortest_edge = min(size.height, size.width)
+                stacked_images = self.resize(
+                    image=stacked_images, size=SizeDict(shortest_edge=shortest_edge), interpolation=interpolation
+                )
+            if do_thumbnail:
+                stacked_images = self.thumbnail(image=stacked_images, size=size)
+            if do_pad:
+                stacked_images = self.pad_image(image=stacked_images, size=size, random_padding=False)
+
+            resized_images_grouped[shape] = stacked_images
+        resized_images = reorder_images(resized_images_grouped, grouped_images_index)
+
+        # Group images by size for further processing
+        # Needed in case do_resize is False, or resize returns images with different sizes
+        grouped_images, grouped_images_index = group_images_by_shape(resized_images)
+        processed_images_grouped = {}
+        for shape, stacked_images in grouped_images.items():
+            if do_center_crop:
+                stacked_images = self.center_crop(stacked_images, crop_size)
+            # Fused rescale and normalize
+            stacked_images = self.rescale_and_normalize(
+                stacked_images, do_rescale, rescale_factor, do_normalize, image_mean, image_std
+            )
+            processed_images_grouped[shape] = stacked_images
+
+        processed_images = reorder_images(processed_images_grouped, grouped_images_index)
+        processed_images = torch.stack(processed_images, dim=0) if return_tensors else processed_images
+
+        return BatchFeature(data={"pixel_values": processed_images}, tensor_type=return_tensors)
+
+
+__all__ = ["DonutImageProcessorFast"]

src/transformers/utils/dummy_torchvision_objects.py

@@ -58,6 +58,13 @@ def __init__(self, *args, **kwargs):
         requires_backends(self, ["torchvision"])
 
 
+class DonutImageProcessorFast(metaclass=DummyObject):
+    _backends = ["torchvision"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torchvision"])
+
+
 class Gemma3ImageProcessorFast(metaclass=DummyObject):
     _backends = ["torchvision"]