Validate processing kwargs with @strict from huggingface_hub (#40793)

* initial design draft

* delete

* fix a few tests

* fix

* fix the rest of tests

* common-kwargs

* why the runner complains about typing with "|"?

* revert

* forgot to delete

* update

* fix last issues

* add more detalis in docs

* pin the latest hub release

* fix tests for new models

* also fast image processor

* fix copies

* image processing ast validated

* fix more tests

* typo.and fix copies

* bump

* style

* fix some tests

* fix copies

* pin rc4 and mark all TypedDict as non-total

* delete typed dict adaptor

* address comments

* delete optionals

setup.py

@@ -114,7 +114,7 @@ _deps = [
     "GitPython<3.1.19",
     "hf-doc-builder>=0.3.0",
     "hf_xet",
-    "huggingface-hub==1.0.0.rc2",
+    "huggingface-hub==1.0.0.rc4",
     "importlib_metadata",
     "ipadic>=1.0.0,<2.0",
     "jinja2>=3.1.0",

src/transformers/dependency_versions_table.py

@@ -23,7 +23,7 @@ deps = {
     "GitPython": "GitPython<3.1.19",
     "hf-doc-builder": "hf-doc-builder>=0.3.0",
     "hf_xet": "hf_xet",
-    "huggingface-hub": "huggingface-hub==1.0.0.rc2",
+    "huggingface-hub": "huggingface-hub==1.0.0.rc4",
     "importlib_metadata": "importlib_metadata",
     "ipadic": "ipadic>=1.0.0,<2.0",
     "jinja2": "jinja2>=3.1.0",

src/transformers/image_processing_utils_fast.py

@@ -18,6 +18,7 @@ from functools import lru_cache, partial
 from typing import Any, Optional, Union
 
 import numpy as np
+from huggingface_hub.dataclasses import validate_typed_dict
 
 from .image_processing_utils import BaseImageProcessor, BatchFeature, get_size_dict
 from .image_transforms import (
@@ -710,6 +711,10 @@ class BaseImageProcessorFast(BaseImageProcessor):
     def preprocess(self, images: ImageInput, *args, **kwargs: Unpack[ImagesKwargs]) -> BatchFeature:
         # args are not validated, but their order in the `preprocess` and `_preprocess` signatures must be the same
         validate_kwargs(captured_kwargs=kwargs.keys(), valid_processor_keys=self._valid_kwargs_names)
+
+        # Perform type validation on received kwargs
+        validate_typed_dict(self.valid_kwargs, kwargs)
+
         # Set default kwargs from self. This ensures that if a kwarg is not provided
         # by the user, it gets its default value from the instance, or is set to None.
         for kwarg_name in self._valid_kwargs_names:

src/transformers/models/aria/modular_aria.py

@@ -38,7 +38,7 @@ from ...image_utils import (
 )
 from ...modeling_flash_attention_utils import FlashAttentionKwargs
 from ...modeling_utils import PreTrainedModel
-from ...processing_utils import MultiModalData, ProcessingKwargs, ProcessorMixin, Unpack
+from ...processing_utils import ImagesKwargs, MultiModalData, ProcessingKwargs, ProcessorMixin, Unpack
 from ...tokenization_utils import PreTokenizedInput, TextInput
 from ...utils import TensorType, TransformersKwargs, auto_docstring, can_return_tuple, logging
 from ..auto import CONFIG_MAPPING, AutoConfig, AutoTokenizer
@@ -904,7 +904,15 @@ class AriaImageProcessor(BaseImageProcessor):
         return num_patches
 
 
+class AriaImagesKwargs(ImagesKwargs, total=False):
+    split_image: bool
+    max_image_size: int
+    min_image_size: int
+
+
 class AriaProcessorKwargs(ProcessingKwargs, total=False):
+    images_kwargs: AriaImagesKwargs
+
     _defaults = {
         "text_kwargs": {
             "padding": False,

src/transformers/models/aria/processing_aria.py

@@ -24,13 +24,21 @@ import numpy as np
 
 from ...image_processing_utils import BatchFeature
 from ...image_utils import ImageInput
-from ...processing_utils import MultiModalData, ProcessingKwargs, ProcessorMixin, Unpack
+from ...processing_utils import ImagesKwargs, MultiModalData, ProcessingKwargs, ProcessorMixin, Unpack
 from ...tokenization_utils import PreTokenizedInput, TextInput
 from ...utils import TensorType
 from ..auto import AutoTokenizer
 
 
+class AriaImagesKwargs(ImagesKwargs, total=False):
+    split_image: bool
+    max_image_size: int
+    min_image_size: int
+
+
 class AriaProcessorKwargs(ProcessingKwargs, total=False):
+    images_kwargs: AriaImagesKwargs
+
     _defaults = {
         "text_kwargs": {
             "padding": False,

src/transformers/models/beit/image_processing_beit.py

@@ -55,7 +55,7 @@ if is_torch_available():
 logger = logging.get_logger(__name__)
 
 
-class BeitImageProcessorKwargs(ImagesKwargs):
+class BeitImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     do_reduce_labels (`bool`, *optional*, defaults to `self.do_reduce_labels`):
         Whether or not to reduce all label values of segmentation maps by 1. Usually used for datasets where 0
@@ -63,7 +63,7 @@ class BeitImageProcessorKwargs(ImagesKwargs):
         ADE20k). The background label will be replaced by 255.
     """
 
-    do_reduce_labels: Optional[bool]
+    do_reduce_labels: bool
 
 
 @requires(backends=("vision",))

src/transformers/models/bridgetower/image_processing_bridgetower.py

@@ -123,8 +123,8 @@ def get_resize_output_image_size(
     return new_height, new_width
 
 
-class BridgeTowerImageProcessorKwargs(ImagesKwargs):
-    size_divisor: Optional[int]
+class BridgeTowerImageProcessorKwargs(ImagesKwargs, total=False):
+    size_divisor: int
 
 
 class BridgeTowerImageProcessor(BaseImageProcessor):

src/transformers/models/cohere2_vision/image_processing_cohere2_vision_fast.py

@@ -33,7 +33,7 @@ from ...processing_utils import ImagesKwargs, Unpack
 from ...utils import TensorType, auto_docstring
 
 
-class Cohere2VisionFastImageProcessorKwargs(ImagesKwargs):
+class Cohere2VisionFastImageProcessorKwargs(ImagesKwargs, total=False):
     """
     crop_to_patches (`bool`, *optional*, defaults to `False`):
         Whether to crop the image to patches. Can be overridden by the `crop_to_patches` parameter in the
@@ -46,9 +46,9 @@ class Cohere2VisionFastImageProcessorKwargs(ImagesKwargs):
         set to `True`. Can be overridden by the `max_patches` parameter in the `preprocess` method.
     """
 
-    crop_to_patches: Optional[bool]
-    min_patches: Optional[int]
-    max_patches: Optional[int]
+    crop_to_patches: bool
+    min_patches: int
+    max_patches: int
 
 
 @lru_cache(maxsize=10)

src/transformers/models/cohere2_vision/modular_cohere2_vision.py

@@ -303,7 +303,7 @@ def get_optimal_tiled_canvas(
     return best_grid
 
 
-class Cohere2VisionFastImageProcessorKwargs(ImagesKwargs):
+class Cohere2VisionFastImageProcessorKwargs(ImagesKwargs, total=False):
     """
     crop_to_patches (`bool`, *optional*, defaults to `False`):
         Whether to crop the image to patches. Can be overridden by the `crop_to_patches` parameter in the
@@ -316,9 +316,9 @@ class Cohere2VisionFastImageProcessorKwargs(ImagesKwargs):
         set to `True`. Can be overridden by the `max_patches` parameter in the `preprocess` method.
     """
 
-    crop_to_patches: Optional[bool]
-    min_patches: Optional[int]
-    max_patches: Optional[int]
+    crop_to_patches: bool
+    min_patches: int
+    max_patches: int
 
 
 @auto_docstring

src/transformers/models/conditional_detr/image_processing_conditional_detr.py

@@ -729,7 +729,7 @@ def compute_segments(
     return segmentation, segments
 
 
-class ConditionalDetrImageProcessorKwargs(ImagesKwargs):
+class ConditionalDetrImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     format (`str`, *optional*, defaults to `AnnotationFormat.COCO_DETECTION`):
         Data format of the annotations. One of "coco_detection" or "coco_panoptic".
@@ -745,9 +745,9 @@ class ConditionalDetrImageProcessorKwargs(ImagesKwargs):
         Path to the directory containing the segmentation masks.
     """
 
-    format: Optional[Union[str, AnnotationFormat]]
-    do_convert_annotations: Optional[bool]
-    return_segmentation_masks: Optional[bool]
+    format: Union[str, AnnotationFormat]
+    do_convert_annotations: bool
+    return_segmentation_masks: bool
     annotations: Optional[Union[AnnotationType, list[AnnotationType]]]
     masks_path: Optional[Union[str, pathlib.Path]]
 

src/transformers/models/convnext/image_processing_convnext.py

@@ -50,14 +50,14 @@ if is_vision_available():
 logger = logging.get_logger(__name__)
 
 
-class ConvNextImageProcessorKwargs(ImagesKwargs):
+class ConvNextImageProcessorKwargs(ImagesKwargs, total=False):
     """
     crop_pct (`float`, *optional*):
         Percentage of the image to crop. Only has an effect if size < 384. Can be
         overridden by `crop_pct` in the`preprocess` method.
     """
 
-    crop_pct: Optional[float]
+    crop_pct: float
 
 
 @requires(backends=("vision",))

src/transformers/models/deepseek_vl/image_processing_deepseek_vl.py

@@ -49,7 +49,7 @@ if is_vision_available():
 logger = logging.get_logger(__name__)
 
 
-class DeepseekVLImageProcessorKwargs(ImagesKwargs):
+class DeepseekVLImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     min_size (`int`, *optional*, defaults to 14):
         The minimum allowed size for the resized image. Ensures that neither the height nor width

src/transformers/models/deepseek_vl_hybrid/image_processing_deepseek_vl_hybrid.py

@@ -50,7 +50,7 @@ if is_vision_available():
 logger = logging.get_logger(__name__)
 
 
-class DeepseekVLHybridImageProcessorKwargs(ImagesKwargs):
+class DeepseekVLHybridImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     min_size (`int`, *optional*, defaults to 14):
         The minimum allowed size for the resized image. Ensures that neither the height nor width
@@ -71,9 +71,9 @@ class DeepseekVLHybridImageProcessorKwargs(ImagesKwargs):
 
     min_size: int
     high_res_size: dict
-    high_res_resample: "PILImageResampling"
-    high_res_image_mean: list[float]
-    high_res_image_std: list[float]
+    high_res_resample: Union["PILImageResampling", int]
+    high_res_image_mean: Union[float, list[float], tuple[float, ...]]
+    high_res_image_std: Union[float, list[float], tuple[float, ...]]
 
 
 class DeepseekVLHybridImageProcessor(BaseImageProcessor):

src/transformers/models/deepseek_vl_hybrid/modular_deepseek_vl_hybrid.py

@@ -429,7 +429,7 @@ class DeepseekVLHybridForConditionalGeneration(DeepseekVLForConditionalGeneratio
         return model_inputs
 
 
-class DeepseekVLHybridImageProcessorKwargs(ImagesKwargs):
+class DeepseekVLHybridImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     min_size (`int`, *optional*, defaults to 14):
         The minimum allowed size for the resized image. Ensures that neither the height nor width
@@ -450,9 +450,9 @@ class DeepseekVLHybridImageProcessorKwargs(ImagesKwargs):
 
     min_size: int
     high_res_size: dict
-    high_res_resample: "PILImageResampling"
-    high_res_image_mean: list[float]
-    high_res_image_std: list[float]
+    high_res_resample: Union["PILImageResampling", int]
+    high_res_image_mean: Union[float, list[float], tuple[float, ...]]
+    high_res_image_std: Union[float, list[float], tuple[float, ...]]
 
 
 class DeepseekVLHybridImageProcessor(DeepseekVLImageProcessor):

src/transformers/models/deformable_detr/image_processing_deformable_detr.py

@@ -82,7 +82,7 @@ if is_scipy_available():
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
-class DeformableDetrImageProcessorKwargs(ImagesKwargs):
+class DeformableDetrImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     format (`str`, *optional*, defaults to `AnnotationFormat.COCO_DETECTION`):
         Data format of the annotations. One of "coco_detection" or "coco_panoptic".
@@ -98,9 +98,9 @@ class DeformableDetrImageProcessorKwargs(ImagesKwargs):
         Path to the directory containing the segmentation masks.
     """
 
-    format: Optional[Union[str, AnnotationFormat]]
-    do_convert_annotations: Optional[bool]
-    return_segmentation_masks: Optional[bool]
+    format: Union[str, AnnotationFormat]
+    do_convert_annotations: bool
+    return_segmentation_masks: bool
     annotations: Optional[Union[AnnotationType, list[AnnotationType]]]
     masks_path: Optional[Union[str, pathlib.Path]]
 

src/transformers/models/detr/image_processing_detr.py

@@ -84,7 +84,7 @@ logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 SUPPORTED_ANNOTATION_FORMATS = (AnnotationFormat.COCO_DETECTION, AnnotationFormat.COCO_PANOPTIC)
 
 
-class DetrImageProcessorKwargs(ImagesKwargs):
+class DetrImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     format (`str`, *optional*, defaults to `AnnotationFormat.COCO_DETECTION`):
         Data format of the annotations. One of "coco_detection" or "coco_panoptic".
@@ -100,9 +100,9 @@ class DetrImageProcessorKwargs(ImagesKwargs):
         Path to the directory containing the segmentation masks.
     """
 
-    format: Optional[Union[str, AnnotationFormat]]
-    do_convert_annotations: Optional[bool]
-    return_segmentation_masks: Optional[bool]
+    format: Union[str, AnnotationFormat]
+    do_convert_annotations: bool
+    return_segmentation_masks: bool
     annotations: Optional[Union[AnnotationType, list[AnnotationType]]]
     masks_path: Optional[Union[str, pathlib.Path]]
 

src/transformers/models/dia/processing_dia.py

@@ -55,7 +55,9 @@ class DiaProcessorKwargs(ProcessingKwargs, total=False):
             "generation": True,
             "sampling_rate": 44100,
         },
-        "common_kwargs": {"return_tensors": "pt"},
+        "common_kwargs": {
+            "return_tensors": "pt",
+        },
     }
 
 

src/transformers/models/donut/image_processing_donut.py

@@ -52,7 +52,7 @@ if is_vision_available():
     import PIL
 
 
-class DonutImageProcessorKwargs(ImagesKwargs):
+class DonutImageProcessorKwargs(ImagesKwargs, total=False):
     """
     do_thumbnail (`bool`, *optional*, defaults to `self.do_thumbnail`):
         Whether to resize the image using thumbnail method.
@@ -60,8 +60,8 @@ class DonutImageProcessorKwargs(ImagesKwargs):
         Whether to align the long axis of the image with the long axis of `size` by rotating by 90 degrees.
     """
 
-    do_thumbnail: Optional[bool]
-    do_align_long_axis: Optional[bool]
+    do_thumbnail: bool
+    do_align_long_axis: bool
 
 
 @requires(backends=("vision",))

src/transformers/models/dpt/image_processing_dpt.py

@@ -64,7 +64,7 @@ if is_vision_available():
 logger = logging.get_logger(__name__)
 
 
-class DPTImageProcessorKwargs(ImagesKwargs):
+class DPTImageProcessorKwargs(ImagesKwargs, total=False):
     """
     ensure_multiple_of (`int`, *optional*, defaults to 1):
         If `do_resize` is `True`, the image is resized to a size that is a multiple of this value. Can be overridden
@@ -78,10 +78,10 @@ class DPTImageProcessorKwargs(ImagesKwargs):
         ADE20k). The background label will be replaced by 255.
     """
 
-    ensure_multiple_of: Optional[int]
-    size_divisor: Optional[int]
-    keep_aspect_ratio: Optional[bool]
-    do_reduce_labels: Optional[bool]
+    ensure_multiple_of: int
+    size_divisor: int
+    keep_aspect_ratio: bool
+    do_reduce_labels: bool
 
 
 def get_resize_output_image_size(

src/transformers/models/efficientloftr/image_processing_efficientloftr.py

@@ -50,13 +50,13 @@ if is_vision_available():
 logger = logging.get_logger(__name__)
 
 
-class EfficientLoFTRImageProcessorKwargs(ImagesKwargs):
+class EfficientLoFTRImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     do_grayscale (`bool`, *optional*, defaults to `True`):
         Whether to convert the image to grayscale. Can be overridden by `do_grayscale` in the `preprocess` method.
     """
 
-    do_grayscale: Optional[bool] = True
+    do_grayscale: bool
 
 
 # Copied from transformers.models.superpoint.image_processing_superpoint.is_grayscale

src/transformers/models/efficientnet/image_processing_efficientnet.py

@@ -44,7 +44,7 @@ if is_vision_available():
 logger = logging.get_logger(__name__)
 
 
-class EfficientNetImageProcessorKwargs(ImagesKwargs):
+class EfficientNetImageProcessorKwargs(ImagesKwargs, total=False):
     """
     rescale_offset (`bool`, *optional*, defaults to `self.rescale_offset`):
         Whether to rescale the image between [-max_range/2, scale_range/2] instead of [0, scale_range].

src/transformers/models/emu3/image_processing_emu3.py

@@ -47,9 +47,9 @@ if is_vision_available():
 logger = logging.get_logger(__name__)
 
 
-class Emu3ImageProcessorKwargs(ImagesKwargs):
-    ratio: Optional[str]
-    image_area: Optional[int]
+class Emu3ImageProcessorKwargs(ImagesKwargs, total=False):
+    ratio: str
+    image_area: int
 
 
 def smart_resize(

src/transformers/models/eomt/image_processing_eomt.py

@@ -55,7 +55,7 @@ if is_torch_available():
     import torch.nn.functional as F
 
 
-class EomtImageProcessorKwargs(ImagesKwargs):
+class EomtImageProcessorKwargs(ImagesKwargs, total=False):
     """
     do_split_image (`bool`, *optional*, defaults to `False`):
         Whether to split the input images into overlapping patches for semantic segmentation. If set to `True`, the
@@ -67,7 +67,7 @@ class EomtImageProcessorKwargs(ImagesKwargs):
     """
 
     do_split_image: bool
-    ignore_index: Optional[int] = None
+    ignore_index: Optional[int]
 
 
 # Adapted from transformers.models.maskformer.image_processing_maskformer.convert_segmentation_map_to_binary_masks

src/transformers/models/flava/image_processing_flava.py

@@ -57,7 +57,7 @@ FLAVA_CODEBOOK_STD = [1.0, 1.0, 1.0]
 LOGIT_LAPLACE_EPS: float = 0.1
 
 
-class FlavaImageProcessorKwargs(ImagesKwargs):
+class FlavaImageProcessorKwargs(ImagesKwargs, total=False):
     """
     return_image_mask (`bool`, *optional*, defaults to `False`):
         Whether to return the image mask. Can be overridden by the `return_image_mask` parameter in `preprocess`.
@@ -118,26 +118,26 @@ class FlavaImageProcessorKwargs(ImagesKwargs):
     """
 
     # Mask related params
-    return_image_mask: Optional[bool]
-    input_size_patches: Optional[int]
-    total_mask_patches: Optional[int]
-    mask_group_min_patches: Optional[int]
-    mask_group_max_patches: Optional[int]
-    mask_group_min_aspect_ratio: Optional[float]
-    mask_group_max_aspect_ratio: Optional[float]
+    return_image_mask: bool
+    input_size_patches: int
+    total_mask_patches: int
+    mask_group_min_patches: int
+    mask_group_max_patches: int
+    mask_group_min_aspect_ratio: float
+    mask_group_max_aspect_ratio: float
     # Codebook related params
-    return_codebook_pixels: Optional[bool]
-    codebook_do_resize: Optional[bool]
-    codebook_size: Optional[bool]
-    codebook_resample: Optional[int]
-    codebook_do_center_crop: Optional[bool]
-    codebook_crop_size: Optional[int]
-    codebook_do_rescale: Optional[bool]
-    codebook_rescale_factor: Optional[Union[int, float]]
-    codebook_do_map_pixels: Optional[bool]
-    codebook_do_normalize: Optional[bool]
-    codebook_image_mean: Optional[Union[float, Iterable[float]]]
-    codebook_image_std: Optional[Union[float, Iterable[float]]]
+    return_codebook_pixels: bool
+    codebook_do_resize: bool
+    codebook_size: bool
+    codebook_resample: int
+    codebook_do_center_crop: bool
+    codebook_crop_size: int
+    codebook_do_rescale: bool
+    codebook_rescale_factor: Union[int, float]
+    codebook_do_map_pixels: bool
+    codebook_do_normalize: bool
+    codebook_image_mean: Union[float, Iterable[float]]
+    codebook_image_std: Union[float, Iterable[float]]
 
 
 # Inspired from https://github.com/microsoft/unilm/blob/master/beit/masking_generator.py

src/transformers/models/gemma3/image_processing_gemma3.py

@@ -51,7 +51,7 @@ if is_vision_available():
     import PIL
 
 
-class Gemma3ImageProcessorKwargs(ImagesKwargs):
+class Gemma3ImageProcessorKwargs(ImagesKwargs, total=False):
     """
     do_pan_and_scan (`bool`, *optional*):
         Whether to apply `pan_and_scan` to images.
@@ -63,10 +63,10 @@ class Gemma3ImageProcessorKwargs(ImagesKwargs):
         Minimum aspect ratio to activate pan and scan.
     """
 
-    do_pan_and_scan: Optional[bool]
-    pan_and_scan_min_crop_size: Optional[int]
-    pan_and_scan_max_num_crops: Optional[int]
-    pan_and_scan_min_ratio_to_activate: Optional[float]
+    do_pan_and_scan: bool
+    pan_and_scan_min_crop_size: int
+    pan_and_scan_max_num_crops: int
+    pan_and_scan_min_ratio_to_activate: float
 
 
 class Gemma3ImageProcessor(BaseImageProcessor):

src/transformers/models/glm4v/image_processing_glm4v.py

@@ -47,7 +47,7 @@ from ...video_utils import VideoInput
 logger = logging.get_logger(__name__)
 
 
-class Glm4vImageProcessorKwargs(ImagesKwargs):
+class Glm4vImageProcessorKwargs(ImagesKwargs, total=False):
     """
     patch_size (`int`, *optional*, defaults to 14):
         The spatial patch size of the vision encoder.
@@ -57,9 +57,9 @@ class Glm4vImageProcessorKwargs(ImagesKwargs):
         The merge size of the vision encoder to llm encoder.
     """
 
-    patch_size: Optional[int]
-    temporal_patch_size: Optional[int]
-    merge_size: Optional[int]
+    patch_size: int
+    temporal_patch_size: int
+    merge_size: int
 
 
 def smart_resize(

src/transformers/models/glm4v/video_processing_glm4v.py

@@ -36,12 +36,12 @@ from ...video_utils import VideoMetadata, group_videos_by_shape, reorder_videos
 from .image_processing_glm4v import smart_resize
 
 
-class Glm4vVideoProcessorInitKwargs(VideosKwargs):
-    max_image_size: Optional[dict[str, int]]
-    patch_size: Optional[int]
-    temporal_patch_size: Optional[int]
-    merge_size: Optional[int]
-    max_duration: Optional[int]
+class Glm4vVideoProcessorInitKwargs(VideosKwargs, total=False):
+    max_image_size: dict[str, int]
+    patch_size: int
+    temporal_patch_size: int
+    merge_size: int
+    max_duration: int
 
 
 @add_start_docstrings(

src/transformers/models/got_ocr2/image_processing_got_ocr2.py

@@ -49,7 +49,7 @@ if is_vision_available():
 logger = logging.get_logger(__name__)
 
 
-class GotOcr2ImageProcessorKwargs(ImagesKwargs):
+class GotOcr2ImageProcessorKwargs(ImagesKwargs, total=False):
     """
     crop_to_patches (`bool`, *optional*, defaults to `False`):
         Whether to crop the image to patches. Can be overridden by the `crop_to_patches` parameter in the
@@ -62,9 +62,9 @@ class GotOcr2ImageProcessorKwargs(ImagesKwargs):
         set to `True`. Can be overridden by the `max_patches` parameter in the `preprocess` method.
     """
 
-    crop_to_patches: Optional[bool]
-    min_patches: Optional[int]
-    max_patches: Optional[int]
+    crop_to_patches: bool
+    min_patches: int
+    max_patches: int
 
 
 # Similar to image_processing_mllama.get_all_supported_aspect_ratios

src/transformers/models/got_ocr2/processing_got_ocr2.py

@@ -36,13 +36,13 @@ class GotOcr2TextKwargs(TextKwargs, total=False):
 
 
 class GotOcr2ImagesKwargs(ImagesKwargs, total=False):
-    crop_to_patches: Optional[bool]
-    min_patches: Optional[int]
-    max_patches: Optional[int]
+    crop_to_patches: bool
+    min_patches: int
+    max_patches: int
     box: Optional[Union[list, tuple[float, float], tuple[float, float, float, float]]]
     color: Optional[str]
-    num_image_tokens: Optional[int]
-    multi_page: Optional[bool]
+    num_image_tokens: int
+    multi_page: bool
 
 
 class GotOcr2ProcessorKwargs(ProcessingKwargs, total=False):

src/transformers/models/grounding_dino/image_processing_grounding_dino.py

@@ -93,7 +93,7 @@ class AnnotationFormat(ExplicitEnum):
 SUPPORTED_ANNOTATION_FORMATS = (AnnotationFormat.COCO_DETECTION, AnnotationFormat.COCO_PANOPTIC)
 
 
-class GroundingDinoImageProcessorKwargs(ImagesKwargs):
+class GroundingDinoImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     format (`str`, *optional*, defaults to `AnnotationFormat.COCO_DETECTION`):
         Data format of the annotations. One of "coco_detection" or "coco_panoptic".
@@ -109,9 +109,9 @@ class GroundingDinoImageProcessorKwargs(ImagesKwargs):
         Path to the directory containing the segmentation masks.
     """
 
-    format: Optional[Union[str, AnnotationFormat]]
-    do_convert_annotations: Optional[bool]
-    return_segmentation_masks: Optional[bool]
+    format: Union[str, AnnotationFormat]
+    do_convert_annotations: bool
+    return_segmentation_masks: bool
     annotations: Optional[Union[AnnotationType, list[AnnotationType]]]
     masks_path: Optional[Union[str, pathlib.Path]]
 

src/transformers/models/idefics/image_processing_idefics.py

@@ -36,7 +36,7 @@ IDEFICS_STANDARD_MEAN = [0.48145466, 0.4578275, 0.40821073]
 IDEFICS_STANDARD_STD = [0.26862954, 0.26130258, 0.27577711]
 
 
-class IdeficsImageProcessorKwargs(ImagesKwargs):
+class IdeficsImageProcessorKwargs(ImagesKwargs, total=False):
     """
     transform (`Callable`, *optional*):
         A custom transform function that accepts a single image can be passed for training. For example,
@@ -47,7 +47,7 @@ class IdeficsImageProcessorKwargs(ImagesKwargs):
     """
 
     transform: Optional[Callable]
-    image_size: Optional[dict[str, int]]
+    image_size: dict[str, int]
 
 
 def convert_to_rgb(image):

src/transformers/models/idefics2/image_processing_idefics2.py

@@ -47,13 +47,13 @@ if is_vision_available():
     from PIL import Image
 
 
-class Idefics2ImageProcessorKwargs(ImagesKwargs):
+class Idefics2ImageProcessorKwargs(ImagesKwargs, total=False):
     """
     do_image_splitting (`bool`, *optional*, defaults to `False`):
         Whether to split the image into a sequence 4 equal sub-images concatenated with the original image.
     """
 
-    do_image_splitting: Optional[bool]
+    do_image_splitting: bool
 
 
 def get_resize_output_image_size(image, size, input_data_format) -> tuple[int, int]:

src/transformers/models/idefics3/image_processing_idefics3.py

@@ -48,7 +48,7 @@ if is_vision_available():
     from PIL import Image
 
 
-class Idefics3ImageProcessorKwargs(ImagesKwargs):
+class Idefics3ImageProcessorKwargs(ImagesKwargs, total=False):
     """
     do_image_splitting (`bool`, *optional*, defaults to `True`):
         Whether to split the image into sub-images concatenated with the original image. They are split into patches
@@ -59,9 +59,9 @@ class Idefics3ImageProcessorKwargs(ImagesKwargs):
         Whether to return the row and column information of the images.
     """
 
-    do_image_splitting: Optional[bool]
-    max_image_size: Optional[dict[str, int]]
-    return_row_col_info: Optional[bool]
+    do_image_splitting: bool
+    max_image_size: dict[str, int]
+    return_row_col_info: bool
 
 
 def _resize_output_size_rescale_to_max_len(

src/transformers/models/imagegpt/image_processing_imagegpt.py

@@ -45,7 +45,7 @@ if is_torch_available():
 logger = logging.get_logger(__name__)
 
 
-class ImageGPTImageProcessorKwargs(ImagesKwargs):
+class ImageGPTImageProcessorKwargs(ImagesKwargs, total=False):
     """
     clusters (`np.ndarray` or `list[list[int]]` or `torch.Tensor`, *optional*):
         The color clusters to use, of shape `(n_clusters, 3)` when color quantizing. Can be overridden by `clusters`
@@ -56,7 +56,7 @@ class ImageGPTImageProcessorKwargs(ImagesKwargs):
     """
 
     clusters: Optional[Union[np.ndarray, list[list[int]], "torch.Tensor"]]
-    do_color_quantize: Optional[bool]
+    do_color_quantize: bool
 
 
 def squared_euclidean_distance(a, b):

src/transformers/models/instructblipvideo/video_processing_instructblipvideo.py

@@ -24,15 +24,11 @@ from torchvision.transforms.v2 import functional as F
 
 from ...image_processing_utils import BatchFeature
 from ...image_utils import OPENAI_CLIP_MEAN, OPENAI_CLIP_STD, PILImageResampling, SizeDict
-from ...processing_utils import Unpack, VideosKwargs
 from ...utils import TensorType
 from ...video_processing_utils import BaseVideoProcessor
 from ...video_utils import group_videos_by_shape, reorder_videos
 
 
-class InstructBlipVideoVideoProcessorInitKwargs(VideosKwargs): ...
-
-
 class InstructBlipVideoVideoProcessor(BaseVideoProcessor):
     resample = PILImageResampling.BICUBIC
     image_mean = OPENAI_CLIP_MEAN
@@ -44,12 +40,8 @@ class InstructBlipVideoVideoProcessor(BaseVideoProcessor):
     do_normalize = True
     do_convert_rgb = True
     do_sample_frames = False  # Set to False for BC, recommended to set `True` in new models
-    valid_kwargs = InstructBlipVideoVideoProcessorInitKwargs
     model_input_names = ["pixel_values"]
 
-    def __init__(self, **kwargs: Unpack[InstructBlipVideoVideoProcessorInitKwargs]):
-        super().__init__(**kwargs)
-
     def _preprocess(
         self,
         videos: list["torch.Tensor"],

src/transformers/models/internvl/video_processing_internvl.py

@@ -27,7 +27,7 @@ from ...video_processing_utils import BaseVideoProcessor
 from ...video_utils import VideoMetadata, group_videos_by_shape, reorder_videos
 
 
-class InternVLVideoProcessorInitKwargs(VideosKwargs):
+class InternVLVideoProcessorInitKwargs(VideosKwargs, total=False):
     initial_shift: Union[bool, float, int]
 
 

src/transformers/models/janus/image_processing_janus.py

@@ -51,7 +51,7 @@ if is_vision_available():
 logger = logging.get_logger(__name__)
 
 
-class JanusImageProcessorKwargs(ImagesKwargs):
+class JanusImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     min_size (`int`, *optional*, defaults to 14):
         The minimum allowed size for the resized image. Ensures that neither the height nor width

src/transformers/models/janus/modular_janus.py

@@ -1289,7 +1289,7 @@ class JanusForConditionalGeneration(JanusPreTrainedModel, GenerationMixin):
             return generated_tokens
 
 
-class JanusImageProcessorKwargs(ImagesKwargs):
+class JanusImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     min_size (`int`, *optional*, defaults to 14):
         The minimum allowed size for the resized image. Ensures that neither the height nor width

src/transformers/models/kosmos2/processing_kosmos2.py

@@ -33,15 +33,17 @@ BboxInput = Union[
     list[list[tuple[float, float, float]]],
 ]
 
+NestedList = list[Union[Optional[int], "NestedList"]]
+
 
 class Kosmos2ImagesKwargs(ImagesKwargs, total=False):
-    bboxes: Optional[list[float]]
-    num_image_tokens: Optional[int]
+    bboxes: Optional[NestedList]  # NOTE: hub validators can't accept `Sequence`
+    num_image_tokens: int
     first_image_token_id: Optional[int]
 
 
 class Kosmos2TextKwargs(TextKwargs, total=False):
-    add_eos_token: Optional[bool]
+    add_eos_token: bool
 
 
 class Kosmos2ProcessorKwargs(ProcessingKwargs, total=False):

src/transformers/models/kosmos2_5/image_processing_kosmos2_5.py

@@ -46,7 +46,7 @@ logger = logging.get_logger(__name__)
 DEFAULT_FONT_PATH = "ybelkada/fonts"
 
 
-class Kosmos2_5ImageProcessorKwargs(ImagesKwargs):
+class Kosmos2_5ImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     patch_size (`Dict[str, int]`, *optional*, defaults to `{"height": 16, "width": 16}`):
         The patch size to use for the image. According to Kosmos2_5 paper and code, the patch size is 16x16.
@@ -55,8 +55,8 @@ class Kosmos2_5ImageProcessorKwargs(ImagesKwargs):
         [KOSMOS 2.5 paper](https://huggingface.co/papers/2309.11419).
     """
 
-    patch_size: Optional[dict[str, int]]
-    max_patches: Optional[int]
+    patch_size: dict[str, int]
+    max_patches: int
 
 
 # Copied from transformers.models.pix2struct.image_processing_pix2struct.torch_extract_patches

src/transformers/models/layoutlmv2/image_processing_layoutlmv2.py

@@ -52,7 +52,7 @@ if is_pytesseract_available():
 logger = logging.get_logger(__name__)
 
 
-class LayoutLMv2ImageProcessorKwargs(ImagesKwargs):
+class LayoutLMv2ImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     apply_ocr (`bool`, *optional*, defaults to `True`):
         Whether to apply the Tesseract OCR engine to get words + normalized bounding boxes. Can be overridden by
@@ -66,7 +66,7 @@ class LayoutLMv2ImageProcessorKwargs(ImagesKwargs):
         `preprocess` method.
     """
 
-    apply_ocr: Optional[bool]
+    apply_ocr: bool
     ocr_lang: Optional[str]
     tesseract_config: Optional[str]
 

src/transformers/models/layoutlmv3/image_processing_layoutlmv3.py

@@ -56,7 +56,7 @@ if is_pytesseract_available():
 logger = logging.get_logger(__name__)
 
 
-class LayoutLMv3ImageProcessorKwargs(ImagesKwargs):
+class LayoutLMv3ImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     apply_ocr (`bool`, *optional*, defaults to `True`):
         Whether to apply the Tesseract OCR engine to get words + normalized bounding boxes. Can be overridden by
@@ -70,7 +70,7 @@ class LayoutLMv3ImageProcessorKwargs(ImagesKwargs):
         `preprocess` method.
     """
 
-    apply_ocr: Optional[bool]
+    apply_ocr: bool
     ocr_lang: Optional[str]
     tesseract_config: Optional[str]
 

src/transformers/models/lfm2_vl/image_processing_lfm2_vl_fast.py

@@ -14,7 +14,7 @@
 # limitations under the License.
 import math
 from functools import lru_cache
-from typing import Optional, Union
+from typing import Union
 
 import torch
 from torchvision.transforms.v2 import functional as F
@@ -169,24 +169,24 @@ def pad_along_first_dim(
     return images, pixel_mask
 
 
-class Lfm2VlImageProcessorKwargs(ImagesKwargs):
+class Lfm2VlImageProcessorKwargs(ImagesKwargs, total=False):
     """
     downsample_factor (`int`, *optional*, defaults to `2`):
         The downsampling factor for images used when resizing the image.
     """
 
-    downsample_factor: Optional[int]
-    do_image_splitting: Optional[bool]
-    min_tiles: Optional[int]
-    max_tiles: Optional[int]
-    use_thumbnail: Optional[bool]
-    min_image_tokens: Optional[int]
-    max_image_tokens: Optional[int]
-    encoder_patch_size: Optional[int]
-    tile_size: Optional[int]
-    max_pixels_tolerance: Optional[float]
-    do_pad: Optional[bool]
-    return_row_col_info: Optional[bool]
+    downsample_factor: int
+    do_image_splitting: bool
+    min_tiles: int
+    max_tiles: int
+    use_thumbnail: bool
+    min_image_tokens: int
+    max_image_tokens: int
+    encoder_patch_size: int
+    tile_size: int
+    max_pixels_tolerance: float
+    do_pad: bool
+    return_row_col_info: bool
 
 
 @auto_docstring

src/transformers/models/lfm2_vl/processing_lfm2_vl.py

@@ -18,9 +18,9 @@ from typing import Optional, Union
 from ...feature_extraction_utils import BatchFeature
 from ...image_utils import ImageInput, make_nested_list_of_images
 from ...processing_utils import (
-    ImagesKwargs,
     ProcessingKwargs,
     ProcessorMixin,
+    TextKwargs,
     Unpack,
 )
 from ...tokenization_utils_base import BatchEncoding, TextInput
@@ -30,25 +30,12 @@ from ...utils import logging
 logger = logging.get_logger(__name__)
 
 
-class Lfm2VlImagesKwargs(ImagesKwargs, total=False):
-    downsample_factor: Optional[int]
-    do_image_splitting: Optional[bool]
-    min_tiles: Optional[int]
-    max_tiles: Optional[int]
-    use_thumbnail: Optional[bool]
-    min_image_tokens: Optional[int]
-    max_image_tokens: Optional[int]
-    encoder_patch_size: Optional[int]
-    tile_size: Optional[int]
-    max_pixels_tolerance: Optional[float]
-    patch_size: Optional[int]
-    do_pad: Optional[bool]
-    return_row_col_info: Optional[bool]
+class Lfm2VlTextKwargs(TextKwargs, total=False):
+    use_image_special_tokens: Optional[bool]
 
 
 class Lfm2VlProcessorKwargs(ProcessingKwargs, total=False):
-    images_kwargs: Lfm2VlImagesKwargs
-
+    text_kwargs: Lfm2VlTextKwargs
     _defaults = {
         "images_kwargs": {
             "return_row_col_info": True,
@@ -75,8 +62,6 @@ class Lfm2VlProcessor(ProcessorMixin):
             An instance of [`PreTrainedTokenizerBase`]. This should correspond with the model's text model. The tokenizer is a required input.
         chat_template (`str`, *optional*):
             A Jinja template which will be used to convert lists of messages in a chat into a tokenizable string.
-        use_image_special_tokens (`bool`, *optional*, defaults to `True`):
-            Whether to use image special tokens or not when processing.
     """
 
     attributes = ["image_processor", "tokenizer"]
@@ -88,12 +73,10 @@ class Lfm2VlProcessor(ProcessorMixin):
         image_processor,
         tokenizer,
         chat_template: Optional[str] = None,
-        use_image_special_tokens: Optional[bool] = True,
         **kwargs,
     ):
         self.image_token = tokenizer.image_token
         self.image_token_id = tokenizer.image_token_id
-        self.use_image_special_tokens = use_image_special_tokens
         self.image_start_token = tokenizer.image_start_token
         self.image_end_token = tokenizer.image_end_token
         self.image_thumbnail_token = tokenizer.image_thumbnail

src/transformers/models/llama4/image_processing_llama4_fast.py

@@ -308,7 +308,7 @@ def get_best_fit(
     return optimal_canvas
 
 
-class Llama4ImageProcessorKwargs(ImagesKwargs):
+class Llama4ImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     max_patches (`int`, *optional*, defaults to 16):
         The maximum number of patches to be extracted from the image.
@@ -320,8 +320,8 @@ class Llama4ImageProcessorKwargs(ImagesKwargs):
         but never upsample, unless the image is smaller than the patch size.
     """
 
-    max_patches: Optional[int]
-    resize_to_max_canvas: Optional[bool]
+    max_patches: int
+    resize_to_max_canvas: bool
 
 
 @auto_docstring

src/transformers/models/llava_next/image_processing_llava_next.py

@@ -59,7 +59,7 @@ if is_vision_available():
     from PIL import Image
 
 
-class LlavaNextImageProcessorKwargs(ImagesKwargs):
+class LlavaNextImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     image_grid_pinpoints (`list[list[int]]`, *optional*):
         A list of possible resolutions to use for processing high resolution images. The best resolution is selected
@@ -67,7 +67,7 @@ class LlavaNextImageProcessorKwargs(ImagesKwargs):
         method.
     """
 
-    image_grid_pinpoints: Optional[list[list[int]]]
+    image_grid_pinpoints: list[list[int]]
 
 
 def divide_to_patches(image: np.ndarray, patch_size: int, input_data_format) -> list[np.ndarray]:

src/transformers/models/llava_onevision/image_processing_llava_onevision.py

@@ -58,7 +58,7 @@ if is_vision_available():
     from PIL import Image
 
 
-class LlavaOnevisionImageProcessorKwargs(ImagesKwargs):
+class LlavaOnevisionImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     image_grid_pinpoints (`list[list[int]]`, *optional*):
         A list of possible resolutions to use for processing high resolution images. The best resolution is selected
@@ -66,7 +66,7 @@ class LlavaOnevisionImageProcessorKwargs(ImagesKwargs):
         method.
     """
 
-    image_grid_pinpoints: Optional[list[list[int]]]
+    image_grid_pinpoints: list[list[int]]
 
 
 # Copied from transformers.models.llava_next.image_processing_llava_next.divide_to_patches

src/transformers/models/llava_onevision/image_processing_llava_onevision_fast.py

@@ -76,8 +76,7 @@ class LlavaOnevisionImageProcessorFast(BaseImageProcessorFast):
             batch_num_images = [1] * len(images)
         else:
             batch_num_images = [1]
-        kwargs["batch_num_images"] = batch_num_images
-        return super().preprocess(images, **kwargs)
+        return super().preprocess(images, batch_num_images, **kwargs)
 
     def _resize_for_patching(
         self,
@@ -202,6 +201,7 @@ class LlavaOnevisionImageProcessorFast(BaseImageProcessorFast):
     def _preprocess(
         self,
         images: list["torch.Tensor"],
+        batch_num_images: list[int],
         do_resize: bool,
         size: SizeDict,
         image_grid_pinpoints: list[list[int]],
@@ -214,7 +214,6 @@ class LlavaOnevisionImageProcessorFast(BaseImageProcessorFast):
         image_mean: Optional[Union[float, list[float]]],
         image_std: Optional[Union[float, list[float]]],
         do_pad: bool,
-        batch_num_images: list[int],
         disable_grouping: Optional[bool],
         return_tensors: Optional[Union[str, TensorType]],
         **kwargs,

src/transformers/models/llava_onevision/modular_llava_onevision.py

@@ -35,7 +35,7 @@ from transformers.models.llava_next_video.modeling_llava_next_video import (
 
 from ...cache_utils import Cache
 from ...image_processing_utils import BatchFeature
-from ...image_processing_utils_fast import group_images_by_shape, reorder_images
+from ...image_processing_utils_fast import BaseImageProcessorFast, group_images_by_shape, reorder_images
 from ...image_utils import (
     OPENAI_CLIP_MEAN,
     OPENAI_CLIP_STD,
@@ -128,12 +128,12 @@ class LlavaOnevisionImageProcessorFast(LlavaNextImageProcessorFast):
             batch_num_images = [1] * len(images)
         else:
             batch_num_images = [1]
-        kwargs["batch_num_images"] = batch_num_images
-        return super().preprocess(images, **kwargs)
+        return BaseImageProcessorFast.preprocess(images, batch_num_images, **kwargs)
 
     def _preprocess(
         self,
         images: list["torch.Tensor"],
+        batch_num_images: list[int],
         do_resize: bool,
         size: SizeDict,
         image_grid_pinpoints: list[list[int]],
@@ -146,7 +146,6 @@ class LlavaOnevisionImageProcessorFast(LlavaNextImageProcessorFast):
         image_mean: Optional[Union[float, list[float]]],
         image_std: Optional[Union[float, list[float]]],
         do_pad: bool,
-        batch_num_images: list[int],
         disable_grouping: Optional[bool],
         return_tensors: Optional[Union[str, TensorType]],
         **kwargs,

src/transformers/models/mask2former/image_processing_mask2former.py

@@ -61,7 +61,7 @@ if is_torch_available():
     from torch import nn
 
 
-class Mask2FormerImageProcessorKwargs(ImagesKwargs):
+class Mask2FormerImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     ignore_index (`int`, *optional*):
         Label to be assigned to background pixels in segmentation maps. If provided, segmentation map pixels
@@ -74,9 +74,9 @@ class Mask2FormerImageProcessorKwargs(ImagesKwargs):
         The number of labels in the segmentation map.
     """
 
-    size_divisor: Optional[int]
+    size_divisor: int
     ignore_index: Optional[int]
-    do_reduce_labels: Optional[bool]
+    do_reduce_labels: bool
     num_labels: Optional[int]
 
 

src/transformers/models/maskformer/image_processing_maskformer.py

@@ -67,7 +67,7 @@ if is_torch_available():
     from torch import nn
 
 
-class MaskFormerImageProcessorKwargs(ImagesKwargs):
+class MaskFormerImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     ignore_index (`int`, *optional*):
         Label to be assigned to background pixels in segmentation maps. If provided, segmentation map pixels
@@ -80,9 +80,9 @@ class MaskFormerImageProcessorKwargs(ImagesKwargs):
         The number of labels in the segmentation map.
     """
 
-    size_divisor: Optional[int]
+    size_divisor: int
     ignore_index: Optional[int]
-    do_reduce_labels: Optional[bool]
+    do_reduce_labels: bool
     num_labels: Optional[int]
 
 

src/transformers/models/mllama/image_processing_mllama.py

@@ -50,13 +50,13 @@ if is_vision_available():
 logger = logging.get_logger(__name__)
 
 
-class MllamaImageProcessorKwargs(ImagesKwargs):
+class MllamaImageProcessorKwargs(ImagesKwargs, total=False):
     """
     max_image_tiles (`int`, *optional*):
         The maximum number of tiles allowed.
     """
 
-    max_image_tiles: Optional[int]
+    max_image_tiles: int
 
 
 @lru_cache(maxsize=10)

src/transformers/models/mllama/processing_mllama.py

@@ -258,9 +258,7 @@ class MllamaProcessor(ProcessorMixin):
             tokenizer_init_kwargs=self.tokenizer.init_kwargs,
             **kwargs,
         )
-
         return_tensors = output_kwargs["text_kwargs"].pop("return_tensors", None)
-        images_kwargs = output_kwargs["images_kwargs"]
 
         data = {}
         if text is not None:
@@ -306,7 +304,7 @@ class MllamaProcessor(ProcessorMixin):
                     )
 
         if images is not None:
-            image_features = self.image_processor(images, **images_kwargs)
+            image_features = self.image_processor(images, **output_kwargs["images_kwargs"])
             num_tiles = image_features.pop("num_tiles")
             data.update(image_features)
 

src/transformers/models/mobilenet_v2/image_processing_mobilenet_v2.py

@@ -51,7 +51,7 @@ from ...utils.import_utils import requires
 logger = logging.get_logger(__name__)
 
 
-class MobileNetV2ImageProcessorKwargs(ImagesKwargs):
+class MobileNetV2ImageProcessorKwargs(ImagesKwargs, total=False):
     """
     do_reduce_labels (`bool`, *optional*, defaults to `self.do_reduce_labels`):
         Whether or not to reduce all label values of segmentation maps by 1. Usually used for datasets where 0
@@ -59,7 +59,7 @@ class MobileNetV2ImageProcessorKwargs(ImagesKwargs):
         ADE20k). The background label will be replaced by 255.
     """
 
-    do_reduce_labels: Optional[bool]
+    do_reduce_labels: bool
 
 
 @requires(backends=("vision",))

src/transformers/models/mobilevit/image_processing_mobilevit.py

@@ -53,7 +53,7 @@ if is_torch_available():
 logger = logging.get_logger(__name__)
 
 
-class MobileVitImageProcessorKwargs(ImagesKwargs):
+class MobileVitImageProcessorKwargs(ImagesKwargs, total=False):
     """
     do_flip_channel_order (`bool`, *optional*, defaults to `self.do_flip_channel_order`):
         Whether to flip the color channels from RGB to BGR or vice versa.
@@ -63,8 +63,8 @@ class MobileVitImageProcessorKwargs(ImagesKwargs):
         ADE20k). The background label will be replaced by 255.
     """
 
-    do_flip_channel_order: Optional[bool]
-    do_reduce_labels: Optional[bool]
+    do_flip_channel_order: bool
+    do_reduce_labels: bool
 
 
 @requires(backends=("vision",))

src/transformers/models/nougat/image_processing_nougat.py

@@ -52,7 +52,7 @@ if is_vision_available():
     import PIL
 
 
-class NougatImageProcessorKwargs(ImagesKwargs):
+class NougatImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     do_crop_margin (`bool`, *optional*, defaults to `True`):
         Whether to crop the image margins.
@@ -62,9 +62,9 @@ class NougatImageProcessorKwargs(ImagesKwargs):
         Whether to align the long axis of the image with the long axis of `size` by rotating by 90 degrees.
     """
 
-    do_crop_margin: Optional[bool]
-    do_thumbnail: Optional[bool]
-    do_align_long_axis: Optional[bool]
+    do_crop_margin: bool
+    do_thumbnail: bool
+    do_align_long_axis: bool
 
 
 class NougatImageProcessor(BaseImageProcessor):

src/transformers/models/oneformer/image_processing_oneformer.py

@@ -64,7 +64,7 @@ if is_torch_available():
     from torch import nn
 
 
-class OneFormerImageProcessorKwargs(ImagesKwargs):
+class OneFormerImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     repo_path (`str`, *optional*, defaults to `shi-labs/oneformer_demo`):
         Path to a local directory or Hugging Face Hub repository containing model metadata.
@@ -85,7 +85,7 @@ class OneFormerImageProcessorKwargs(ImagesKwargs):
     num_text: Optional[int]
     num_labels: Optional[int]
     ignore_index: Optional[int]
-    do_reduce_labels: Optional[bool]
+    do_reduce_labels: bool
 
 
 # Copied from transformers.models.detr.image_processing_detr.max_across_indices

src/transformers/models/ovis2/image_processing_ovis2.py

@@ -44,7 +44,7 @@ if is_vision_available():
 logger = logging.get_logger(__name__)
 
 
-class Ovis2ImageProcessorKwargs(ImagesKwargs):
+class Ovis2ImageProcessorKwargs(ImagesKwargs, total=False):
     """
     crop_to_patches (`bool`, *optional*, defaults to `False`):
         Whether to crop the image to patches. Can be overridden by the `crop_to_patches` parameter in the
@@ -61,10 +61,10 @@ class Ovis2ImageProcessorKwargs(ImagesKwargs):
         `preprocess` method.
     """
 
-    crop_to_patches: Optional[bool]
-    min_patches: Optional[int]
-    max_patches: Optional[int]
-    use_covering_area_grid: Optional[bool]
+    crop_to_patches: bool
+    min_patches: int
+    max_patches: int
+    use_covering_area_grid: bool
 
 
 # Similar to image_processing_mllama.get_all_supported_aspect_ratios

src/transformers/models/perception_lm/image_processing_perception_lm_fast.py

@@ -42,7 +42,7 @@ from ...utils import (
 )
 
 
-class PerceptionLMImageProcessorKwargs(ImagesKwargs):
+class PerceptionLMImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     vision_input_type (`str`, *optional*, defaults to `"thumb+tile"`):
         Vision processing strategy. `"thumb+tile"` uses both thumbnails and multiple tiles for
@@ -54,8 +54,8 @@ class PerceptionLMImageProcessorKwargs(ImagesKwargs):
     """
 
     vision_input_type: Optional[str]
-    tile_size: Optional[int]
-    max_num_tiles: Optional[int]
+    tile_size: int
+    max_num_tiles: int
 
 
 @auto_docstring
@@ -68,7 +68,7 @@ class PerceptionLMImageProcessorFast(BaseImageProcessorFast):
     do_rescale = True
     do_normalize = True
     do_convert_rgb = True
-    vision_input_type = "thumb+tail"
+    vision_input_type = "thumb+tile"
     tile_size = 448
     max_num_tiles = 36
     size = {"width": 448, "height": 448}  # for backward compatibility in tests

src/transformers/models/phi4_multimodal/image_processing_phi4_multimodal_fast.py

@@ -35,7 +35,7 @@ from ...utils import (
 logger = logging.get_logger(__name__)
 
 
-class Phi4MultimodalImageProcessorKwargs(ImagesKwargs):
+class Phi4MultimodalImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     patch_size (`int`, *optional*):
         The size of the patch.
@@ -43,8 +43,8 @@ class Phi4MultimodalImageProcessorKwargs(ImagesKwargs):
         The maximum number of crops per image.
     """
 
-    patch_size: Optional[int]
-    dynamic_hd: Optional[int]
+    patch_size: int
+    dynamic_hd: int
 
 
 @auto_docstring

src/transformers/models/pix2struct/image_processing_pix2struct.py

@@ -49,7 +49,7 @@ logger = logging.get_logger(__name__)
 DEFAULT_FONT_PATH = "ybelkada/fonts"
 
 
-class Pix2StructImageProcessorKwargs(ImagesKwargs):
+class Pix2StructImageProcessorKwargs(ImagesKwargs, total=False):
     """
     max_patches (`int`, *optional*):
         Maximum number of patches to extract.
@@ -57,7 +57,7 @@ class Pix2StructImageProcessorKwargs(ImagesKwargs):
         Text to render as a header. Only has an effect if `image_processor.is_vqa` is `True`.
     """
 
-    max_patches: Optional[int]
+    max_patches: int
     header_text: Optional[Union[list[str], str]]
 
 

src/transformers/models/pixtral/image_processing_pixtral.py

@@ -50,13 +50,13 @@ if is_vision_available():
     import PIL
 
 
-class PixtralImageProcessorKwargs(ImagesKwargs):
+class PixtralImageProcessorKwargs(ImagesKwargs, total=False):
     """
-    patch_size (`dict[str, int]` *optional*, defaults to `{"height": 16, "width": 16}`):
+    patch_size (`Union[dict[str, int], int]` *optional*, defaults to `{"height": 16, "width": 16}`):
         Size of the patches in the model, used to calculate the output image size. Can be overridden by `patch_size` in the `preprocess` method.
     """
 
-    patch_size: Optional[dict[str, int]]
+    patch_size: Union[dict[str, int], int]
 
 
 # Adapted from function in image_transforms.py to ensure any transparent pixels are converted to white.

src/transformers/models/poolformer/image_processing_poolformer.py

@@ -48,13 +48,13 @@ if is_vision_available():
 logger = logging.get_logger(__name__)
 
 
-class PoolFormerImageProcessorKwargs(ImagesKwargs):
+class PoolFormerImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     crop_pct (`float`, *optional*, defaults to `self.crop_pct`):
         Percentage of the image to crop. Only has an effect if `do_resize` is set to `True`.
     """
 
-    crop_pct: Optional[float]
+    crop_pct: float
 
 
 class PoolFormerImageProcessor(BaseImageProcessor):

src/transformers/models/prompt_depth_anything/image_processing_prompt_depth_anything.py

@@ -54,7 +54,7 @@ if is_torch_available():
 logger = logging.get_logger(__name__)
 
 
-class PromptDepthAnythingImageProcessorKwargs(ImagesKwargs):
+class PromptDepthAnythingImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     keep_aspect_ratio (`bool`, *optional*):
         If `True`, the image is resized to the largest possible size such that the aspect ratio is preserved.
@@ -64,10 +64,10 @@ class PromptDepthAnythingImageProcessorKwargs(ImagesKwargs):
         Scale factor to convert the prompt depth to meters.
     """
 
-    keep_aspect_ratio: Optional[bool]
-    ensure_multiple_of: Optional[int]
-    size_divisor: Optional[int]
-    prompt_scale_to_meter: Optional[float]
+    keep_aspect_ratio: bool
+    ensure_multiple_of: int
+    size_divisor: int
+    prompt_scale_to_meter: float
 
 
 def _constrain_to_multiple_of(val, multiple, min_val=0, max_val=None):

src/transformers/models/qwen2_5_omni/processing_qwen2_5_omni.py

@@ -32,17 +32,17 @@ from ...video_utils import VideoInput
 
 # Redefine kwargs for videos because Qwen-Omni uses some kwargs for processing omni
 # and does not use them in video processor class
-class Qwen2_5_OmniVideosKwargs(VideosKwargs):
-    min_pixels: Optional[int]
-    max_pixels: Optional[int]
-    patch_size: Optional[int]
-    temporal_patch_size: Optional[int]
-    merge_size: Optional[int]
-    min_frames: Optional[int]
-    max_frames: Optional[int]
-    use_audio_in_video: Optional[bool]
-    seconds_per_chunk: Optional[float]
-    position_id_per_seconds: Optional[int]
+class Qwen2_5_OmniVideosKwargs(VideosKwargs, total=False):
+    min_pixels: int
+    max_pixels: int
+    patch_size: int
+    temporal_patch_size: int
+    merge_size: int
+    min_frames: int
+    max_frames: int
+    use_audio_in_video: bool
+    seconds_per_chunk: float
+    position_id_per_seconds: Union[int, float]
 
 
 class Qwen2_5OmniProcessorKwargs(ProcessingKwargs, total=False):

src/transformers/models/qwen2_vl/image_processing_qwen2_vl.py

@@ -52,7 +52,7 @@ from ...video_utils import VideoInput, make_batched_videos
 logger = logging.get_logger(__name__)
 
 
-class Qwen2VLImageProcessorKwargs(ImagesKwargs):
+class Qwen2VLImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     min_pixels (`int`, *optional*, defaults to `56 * 56`):
         The min pixels of the image to resize the image.
@@ -66,11 +66,11 @@ class Qwen2VLImageProcessorKwargs(ImagesKwargs):
         The merge size of the vision encoder to llm encoder.
     """
 
-    min_pixels: Optional[int]
-    max_pixels: Optional[int]
-    patch_size: Optional[int]
-    temporal_patch_size: Optional[int]
-    merge_size: Optional[int]
+    min_pixels: int
+    max_pixels: int
+    patch_size: int
+    temporal_patch_size: int
+    merge_size: int
 
 
 def smart_resize(

src/transformers/models/qwen2_vl/video_processing_qwen2_vl.py

@@ -41,14 +41,14 @@ from ...video_utils import VideoMetadata, group_videos_by_shape, reorder_videos
 from .image_processing_qwen2_vl import smart_resize
 
 
-class Qwen2VLVideoProcessorInitKwargs(VideosKwargs):
-    min_pixels: Optional[int]
-    max_pixels: Optional[int]
-    patch_size: Optional[int]
-    temporal_patch_size: Optional[int]
-    merge_size: Optional[int]
-    min_frames: Optional[int]
-    max_frames: Optional[int]
+class Qwen2VLVideoProcessorInitKwargs(VideosKwargs, total=False):
+    min_pixels: int
+    max_pixels: int
+    patch_size: int
+    temporal_patch_size: int
+    merge_size: int
+    min_frames: int
+    max_frames: int
 
 
 @add_start_docstrings(

src/transformers/models/qwen3_omni_moe/processing_qwen3_omni_moe.py

@@ -20,7 +20,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import re
-from typing import Optional
+from typing import Union
 
 import numpy as np
 
@@ -34,17 +34,17 @@ from ...video_utils import VideoInput, make_batched_videos
 
 # Redefine kwargs for videos because Qwen-Omni uses some kwargs for processing omni
 # and does not use them in video processor class
-class Qwen3OmniMoeVideosKwargs(VideosKwargs):
-    min_pixels: Optional[int]
-    max_pixels: Optional[int]
-    patch_size: Optional[int]
-    temporal_patch_size: Optional[int]
-    merge_size: Optional[int]
-    min_frames: Optional[int]
-    max_frames: Optional[int]
-    use_audio_in_video: Optional[bool]
-    seconds_per_chunk: Optional[float]
-    position_id_per_seconds: Optional[int]
+class Qwen3OmniMoeVideosKwargs(VideosKwargs, total=False):
+    min_pixels: int
+    max_pixels: int
+    patch_size: int
+    temporal_patch_size: int
+    merge_size: int
+    min_frames: int
+    max_frames: int
+    use_audio_in_video: bool
+    seconds_per_chunk: float
+    position_id_per_seconds: Union[int, float]
 
 
 class Qwen3OmniMoeProcessorKwargs(ProcessingKwargs, total=False):

src/transformers/models/qwen3_vl/video_processing_qwen3_vl.py

@@ -64,12 +64,12 @@ def smart_resize(
     return h_bar, w_bar
 
 
-class Qwen3VLVideoProcessorInitKwargs(VideosKwargs):
-    patch_size: Optional[int]
-    temporal_patch_size: Optional[int]
-    merge_size: Optional[int]
-    min_frames: Optional[int]
-    max_frames: Optional[int]
+class Qwen3VLVideoProcessorInitKwargs(VideosKwargs, total=False):
+    patch_size: int
+    temporal_patch_size: int
+    merge_size: int
+    min_frames: int
+    max_frames: int
 
 
 @add_start_docstrings(

src/transformers/models/rt_detr/image_processing_rt_detr.py

@@ -68,7 +68,7 @@ logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 SUPPORTED_ANNOTATION_FORMATS = (AnnotationFormat.COCO_DETECTION,)
 
 
-class RTDetrImageProcessorKwargs(ImagesKwargs):
+class RTDetrImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     format (`str`, *optional*, defaults to `AnnotationFormat.COCO_DETECTION`):
         Data format of the annotations. One of "coco_detection" or "coco_panoptic".
@@ -84,9 +84,9 @@ class RTDetrImageProcessorKwargs(ImagesKwargs):
         Path to the directory containing the segmentation masks.
     """
 
-    format: Optional[Union[str, AnnotationFormat]]
-    do_convert_annotations: Optional[bool]
-    return_segmentation_masks: Optional[bool]
+    format: Union[str, AnnotationFormat]
+    do_convert_annotations: bool
+    return_segmentation_masks: bool
     annotations: Optional[Union[AnnotationType, list[AnnotationType]]]
     masks_path: Optional[Union[str, pathlib.Path]]
 

src/transformers/models/sam/image_processing_sam.py

@@ -58,7 +58,7 @@ if is_torchvision_available():
 logger = logging.get_logger(__name__)
 
 
-class SamImageProcessorKwargs(ImagesKwargs):
+class SamImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     mask_size (`dict[str, int]`, *optional*):
         The size `{"longest_edge": int}` to resize the segmentation maps to.
@@ -67,8 +67,8 @@ class SamImageProcessorKwargs(ImagesKwargs):
         map size provided for preprocessing.
     """
 
-    mask_size: Optional[dict[str, int]]
-    mask_pad_size: Optional[dict[str, int]]
+    mask_size: dict[str, int]
+    mask_pad_size: dict[str, int]
 
 
 class SamImageProcessor(BaseImageProcessor):

src/transformers/models/sam/processing_sam.py

@@ -31,14 +31,14 @@ if is_torch_available():
     import torch
 
 
-class SamImagesKwargs(ImagesKwargs):
+class SamImagesKwargs(ImagesKwargs, total=False):
     segmentation_maps: Optional[ImageInput]
     input_points: Optional[list[list[float]]]
     input_labels: Optional[list[list[int]]]
     input_boxes: Optional[list[list[list[float]]]]
-    point_pad_value: Optional[int]
-    mask_size: Optional[dict[str, int]]
-    mask_pad_size: Optional[dict[str, int]]
+    point_pad_value: int
+    mask_size: dict[str, int]
+    mask_pad_size: dict[str, int]
 
 
 class SamProcessorKwargs(ProcessingKwargs, total=False):

src/transformers/models/sam2/image_processing_sam2_fast.py

@@ -43,13 +43,13 @@ from ...processing_utils import ImagesKwargs, Unpack
 from ...utils import TensorType, auto_docstring
 
 
-class Sam2FastImageProcessorKwargs(ImagesKwargs):
+class Sam2FastImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     mask_size (`dict[str, int]`, *optional*):
         The size `{"height": int, "width": int}` to resize the segmentation maps to.
     """
 
-    mask_size: Optional[dict[str, int]]
+    mask_size: dict[str, int]
 
 
 def _compute_stability_score(masks: "torch.Tensor", mask_threshold: float, stability_score_offset: int):

src/transformers/models/sam2/modular_sam2.py

@@ -70,13 +70,13 @@ from .configuration_sam2 import (
 logger = logging.get_logger(__name__)
 
 
-class Sam2FastImageProcessorKwargs(ImagesKwargs):
+class Sam2FastImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     mask_size (`dict[str, int]`, *optional*):
         The size `{"height": int, "width": int}` to resize the segmentation maps to.
     """
 
-    mask_size: Optional[dict[str, int]]
+    mask_size: dict[str, int]
 
 
 @auto_docstring

src/transformers/models/sam_hq/processing_samhq.py

@@ -31,14 +31,14 @@ if is_torch_available():
     import torch
 
 
-class SamHQImagesKwargs(ImagesKwargs):
+class SamHQImagesKwargs(ImagesKwargs, total=False):
     segmentation_maps: Optional[ImageInput]
     input_points: Optional[list[list[float]]]
     input_labels: Optional[list[list[int]]]
     input_boxes: Optional[list[list[list[float]]]]
     point_pad_value: Optional[int]
-    mask_size: Optional[dict[str, int]]
-    mask_pad_size: Optional[dict[str, int]]
+    mask_size: dict[str, int]
+    mask_pad_size: dict[str, int]
 
 
 class SamHQProcessorKwargs(ProcessingKwargs, total=False):

src/transformers/models/segformer/image_processing_segformer.py

@@ -55,7 +55,7 @@ if is_torch_available():
 logger = logging.get_logger(__name__)
 
 
-class SegformerImageProcessorKwargs(ImagesKwargs):
+class SegformerImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     do_reduce_labels (`bool`, *optional*, defaults to `self.do_reduce_labels`):
         Whether or not to reduce all label values of segmentation maps by 1. Usually used for datasets where 0
@@ -63,7 +63,7 @@ class SegformerImageProcessorKwargs(ImagesKwargs):
         ADE20k). The background label will be replaced by 255.
     """
 
-    do_reduce_labels: Optional[bool]
+    do_reduce_labels: bool
 
 
 @requires(backends=("vision",))

src/transformers/models/siglip2/image_processing_siglip2.py

@@ -48,7 +48,7 @@ if is_vision_available():
     from PIL import Image
 
 
-class Siglip2ImageProcessorKwargs(ImagesKwargs):
+class Siglip2ImageProcessorKwargs(ImagesKwargs, total=False):
     """
     patch_size (`int`, *optional*, defaults to 16):
         The size (resolution) of each patch the image will be split to.
@@ -57,8 +57,8 @@ class Siglip2ImageProcessorKwargs(ImagesKwargs):
         and then padded in "patch" dimension to match this number exactly.
     """
 
-    patch_size: Optional[int]
-    max_num_patches: Optional[int]
+    patch_size: int
+    max_num_patches: int
 
 
 @lru_cache(maxsize=256)

src/transformers/models/smolvlm/image_processing_smolvlm.py

@@ -53,7 +53,7 @@ if is_vision_available():
 logger = logging.get_logger(__name__)
 
 
-class SmolVLMImageProcessorKwargs(ImagesKwargs):
+class SmolVLMImageProcessorKwargs(ImagesKwargs, total=False):
     """
     do_image_splitting (`bool`, *optional*, defaults to `True`):
         Whether to split the image into sub-images concatenated with the original image. They are split into patches
@@ -64,9 +64,9 @@ class SmolVLMImageProcessorKwargs(ImagesKwargs):
         Whether to return the row and column information of the images.
     """
 
-    do_image_splitting: Optional[bool]
-    max_image_size: Optional[dict[str, int]]
-    return_row_col_info: Optional[bool]
+    do_image_splitting: bool
+    max_image_size: dict[str, int]
+    return_row_col_info: bool
 
 
 MAX_IMAGE_SIZE = 4096  # 4k resolution as absolute maximum

src/transformers/models/smolvlm/video_processing_smolvlm.py

@@ -90,8 +90,8 @@ def get_resize_output_image_size(
     return height, width
 
 
-class SmolVLMVideoProcessorInitKwargs(VideosKwargs):
-    max_image_size: Optional[dict[str, int]]
+class SmolVLMVideoProcessorInitKwargs(VideosKwargs, total=False):
+    max_image_size: dict[str, int]
 
 
 class SmolVLMVideoProcessor(BaseVideoProcessor):

src/transformers/models/superpoint/image_processing_superpoint.py

@@ -46,13 +46,13 @@ if is_vision_available():
 logger = logging.get_logger(__name__)
 
 
-class SuperPointImageProcessorKwargs(ImagesKwargs):
+class SuperPointImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     do_grayscale (`bool`, *optional*, defaults to `True`):
         Whether to convert the image to grayscale. Can be overridden by `do_grayscale` in the `preprocess` method.
     """
 
-    do_grayscale: Optional[bool] = True
+    do_grayscale: bool
 
 
 def is_grayscale(

src/transformers/models/swin2sr/image_processing_swin2sr.py

@@ -38,8 +38,8 @@ from ...utils.deprecation import deprecate_kwarg
 logger = logging.get_logger(__name__)
 
 
-class Swin2SRImageProcessorKwargs(ImagesKwargs):
-    size_divisor: Optional[int]
+class Swin2SRImageProcessorKwargs(ImagesKwargs, total=False):
+    size_divisor: int
 
 
 class Swin2SRImageProcessor(BaseImageProcessor):

src/transformers/models/textnet/image_processing_textnet.py

@@ -49,8 +49,8 @@ if is_vision_available():
     import PIL
 
 
-class TextNetImageProcessorKwargs(ImagesKwargs):
-    size_divisor: Optional[int]
+class TextNetImageProcessorKwargs(ImagesKwargs, total=False):
+    size_divisor: int
 
 
 class TextNetImageProcessor(BaseImageProcessor):

src/transformers/models/tvp/image_processing_tvp.py

@@ -50,7 +50,7 @@ if is_vision_available():
 logger = logging.get_logger(__name__)
 
 
-class TvpImageProcessorKwargs(ImagesKwargs):
+class TvpImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     do_flip_channel_order (`bool`, *optional*):
         Whether to flip the channel order of the image from RGB to BGR.
@@ -60,7 +60,7 @@ class TvpImageProcessorKwargs(ImagesKwargs):
         Padding mode to use — `'constant'`, `'edge'`, `'reflect'`, or `'symmetric'`.
     """
 
-    do_flip_channel_order: Optional[bool]
+    do_flip_channel_order: bool
     constant_values: Optional[Union[float, list[float]]]
     pad_mode: Optional[str]
 

src/transformers/models/udop/processing_udop.py

@@ -31,7 +31,7 @@ logger = logging.get_logger(__name__)
 
 class UdopTextKwargs(TextKwargs, total=False):
     word_labels: Optional[Union[list[int], list[list[int]]]]
-    boxes: Union[list[list[int]], list[list[list[int]]]]
+    boxes: Optional[Union[list[list[int]], list[list[list[int]]]]]
 
 
 class UdopProcessorKwargs(ProcessingKwargs, total=False):

src/transformers/models/vilt/image_processing_vilt.py

@@ -47,8 +47,8 @@ if is_vision_available():
 logger = logging.get_logger(__name__)
 
 
-class ViltImageProcessorKwargs(ImagesKwargs):
-    size_divisor: Optional[int]
+class ViltImageProcessorKwargs(ImagesKwargs, total=False):
+    size_divisor: int
 
 
 def max_across_indices(values: Iterable[Any]) -> list[Any]:

src/transformers/models/vitmatte/image_processing_vitmatte.py

@@ -41,8 +41,8 @@ from ...utils.deprecation import deprecate_kwarg
 logger = logging.get_logger(__name__)
 
 
-class VitMatteImageProcessorKwargs(ImagesKwargs):
-    size_divisor: Optional[int]
+class VitMatteImageProcessorKwargs(ImagesKwargs, total=False):
+    size_divisor: int
 
 
 class VitMatteImageProcessor(BaseImageProcessor):

src/transformers/models/yolos/image_processing_yolos.py

@@ -81,7 +81,7 @@ logger = logging.get_logger(__name__)
 SUPPORTED_ANNOTATION_FORMATS = (AnnotationFormat.COCO_DETECTION, AnnotationFormat.COCO_PANOPTIC)
 
 
-class YolosImageProcessorKwargs(ImagesKwargs):
+class YolosImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     format (`str`, *optional*, defaults to `AnnotationFormat.COCO_DETECTION`):
         Data format of the annotations. One of "coco_detection" or "coco_panoptic".
@@ -97,9 +97,9 @@ class YolosImageProcessorKwargs(ImagesKwargs):
         Path to the directory containing the segmentation masks.
     """
 
-    format: Optional[Union[str, AnnotationFormat]]
-    do_convert_annotations: Optional[bool]
-    return_segmentation_masks: Optional[bool]
+    format: Union[str, AnnotationFormat]
+    do_convert_annotations: bool
+    return_segmentation_masks: bool
     annotations: Optional[Union[AnnotationType, list[AnnotationType]]]
     masks_path: Optional[Union[str, pathlib.Path]]
 

src/transformers/models/zoedepth/image_processing_zoedepth.py

@@ -62,7 +62,7 @@ if is_torch_available():
 logger = logging.get_logger(__name__)
 
 
-class ZoeDepthImageProcessorKwargs(ImagesKwargs):
+class ZoeDepthImageProcessorKwargs(ImagesKwargs, total=False):
     """
     keep_aspect_ratio (`bool`, *optional*, defaults to `True`):
         If `True`, the image is resized by choosing the smaller of the height and width scaling factors and using it
@@ -77,8 +77,8 @@ class ZoeDepthImageProcessorKwargs(ImagesKwargs):
         Can be overridden by `ensure_multiple_of` in `preprocess`.
     """
 
-    keep_aspect_ratio: Optional[bool]
-    ensure_multiple_of: Optional[int]
+    keep_aspect_ratio: bool
+    ensure_multiple_of: int
 
 
 def get_resize_output_image_size(

src/transformers/processing_utils.py

@@ -25,10 +25,11 @@ import typing
 import warnings
 from dataclasses import dataclass
 from pathlib import Path
-from typing import Any, Optional, TypedDict, TypeVar, Union
+from typing import Annotated, Any, Literal, Optional, TypedDict, TypeVar, Union
 
 import numpy as np
 import typing_extensions
+from huggingface_hub.dataclasses import validate_typed_dict
 from huggingface_hub.errors import EntryNotFoundError
 
 from .audio_utils import AudioInput, load_audio
@@ -36,13 +37,23 @@ from .dynamic_module_utils import custom_object_save
 from .feature_extraction_utils import BatchFeature
 from .image_utils import ChannelDimension, ImageInput, is_vision_available
 from .utils.chat_template_utils import render_jinja_template
-from .video_utils import VideoInput, VideoMetadata
+from .utils.type_validators import (
+    device_validator,
+    image_size_validator,
+    padding_validator,
+    positive_any_number,
+    positive_int,
+    resampling_validator,
+    tensor_type_validator,
+    truncation_validator,
+    video_metadata_validator,
+)
+from .video_utils import VideoInput, VideoMetadataType
 
 
 if is_vision_available():
     from .image_utils import PILImageResampling
 
-
 from .tokenization_utils_base import (
     PaddingStrategy,
     PreTokenizedInput,
@@ -72,8 +83,6 @@ from .utils.deprecation import deprecate_kwarg
 
 
 if is_torch_available():
-    import torch
-
     from .modeling_utils import PreTrainedAudioTokenizerBase
 
 
@@ -137,18 +146,22 @@ class TextKwargs(TypedDict, total=False):
             The side on which padding will be applied.
         return_mm_token_type_ids (`bool`, *optional*):
             Whether to return multimodal token type ids indicating mm placeholder token positions.
+        return_tensors (`str` or [`~utils.TensorType`], *optional*):
+            If set, will return tensors of a particular framework. Acceptable values are:
+            - `'pt'`: Return PyTorch `torch.Tensor` objects.
+            - `'np'`: Return NumPy `np.ndarray` objects.
     """
 
     text_pair: Optional[Union[TextInput, PreTokenizedInput, list[TextInput], list[PreTokenizedInput]]]
-    text_target: Union[TextInput, PreTokenizedInput, list[TextInput], list[PreTokenizedInput]]
+    text_target: Optional[Union[TextInput, PreTokenizedInput, list[TextInput], list[PreTokenizedInput]]]
     text_pair_target: Optional[Union[TextInput, PreTokenizedInput, list[TextInput], list[PreTokenizedInput]]]
     add_special_tokens: Optional[bool]
-    padding: Union[bool, str, PaddingStrategy]
-    truncation: Union[bool, str, TruncationStrategy]
-    max_length: Optional[int]
-    stride: Optional[int]
+    padding: Annotated[Optional[Union[bool, str, PaddingStrategy]], padding_validator()]
+    truncation: Annotated[Optional[Union[bool, str, TruncationStrategy]], truncation_validator()]
+    max_length: Annotated[Optional[int], positive_int()]
+    stride: Annotated[Optional[int], positive_int()]
     is_split_into_words: Optional[bool]
-    pad_to_multiple_of: Optional[int]
+    pad_to_multiple_of: Annotated[Optional[int], positive_int()]
     return_token_type_ids: Optional[bool]
     return_attention_mask: Optional[bool]
     return_overflowing_tokens: Optional[bool]
@@ -156,9 +169,9 @@ class TextKwargs(TypedDict, total=False):
     return_offsets_mapping: Optional[bool]
     return_length: Optional[bool]
     verbose: Optional[bool]
-    padding_side: Optional[str]
+    padding_side: Optional[Literal["left", "right"]]
     return_mm_token_type_ids: Optional[bool]
-    return_tensors: Optional[Union[str, TensorType]]
+    return_tensors: Annotated[Optional[Union[str, TensorType]], tensor_type_validator()]
 
 
 class ImagesKwargs(TypedDict, total=False):
@@ -175,6 +188,8 @@ class ImagesKwargs(TypedDict, total=False):
             Resize the shorter side of the input to `size["shortest_edge"]`.
         crop_size (`dict[str, int]`, *optional*):
             Desired output size when applying center-cropping.
+        do_convert_rgb (`bool`):
+            Whether to convert the video to RGB format.
         resample (`PILImageResampling`, *optional*):
             Resampling filter to use if resizing the image.
         do_rescale (`bool`, *optional*):
@@ -183,9 +198,9 @@ class ImagesKwargs(TypedDict, total=False):
             Scale factor to use if rescaling the image.
         do_normalize (`bool`, *optional*):
             Whether to normalize the image.
-        image_mean (`float` or `list[float]`, *optional*):
+        image_mean (`float` or `list[float] or tuple[float, float, float]`, *optional*):
             Mean to use if normalizing the image.
-        image_std (`float` or `list[float]`, *optional*):
+        image_std (`float` or `list[float] or tuple[float, float, float]`, *optional*):
             Standard deviation to use if normalizing the image.
         do_pad (`bool`, *optional*):
             Whether to pad the images in the batch.
@@ -199,28 +214,32 @@ class ImagesKwargs(TypedDict, total=False):
             The channel dimension format for the input image.
         device (`Union[str, torch.Tensor]`, *optional*):
             The device to use for processing (e.g. "cpu", "cuda"), only relevant for fast image processing.
+        return_tensors (`str` or [`~utils.TensorType`], *optional*):
+            If set, will return tensors of a particular framework. Acceptable values are:
+            - `'pt'`: Return PyTorch `torch.Tensor` objects.
+            - `'np'`: Return NumPy `np.ndarray` objects.
         disable_grouping (`bool`, *optional*):
             Whether to group images by shapes when processing or not, only relevant for fast image processing.
     """
 
     do_convert_rgb: Optional[bool]
     do_resize: Optional[bool]
-    size: Optional[dict[str, int]]
-    crop_size: Optional[dict[str, int]]
-    resample: Optional[Union["PILImageResampling", int]]
+    size: Annotated[Optional[Union[int, list[int], tuple[int, ...], dict[str, int]]], image_size_validator()]
+    crop_size: Annotated[Optional[Union[int, list[int], tuple[int, ...], dict[str, int]]], image_size_validator()]
+    resample: Annotated[Optional[Union["PILImageResampling", int]], resampling_validator()]
     do_rescale: Optional[bool]
     rescale_factor: Optional[float]
     do_normalize: Optional[bool]
-    image_mean: Optional[Union[float, list[float]]]
-    image_std: Optional[Union[float, list[float]]]
+    image_mean: Optional[Union[float, list[float], tuple[float, ...]]]
+    image_std: Optional[Union[float, list[float], tuple[float, ...]]]
     do_pad: Optional[bool]
-    pad_size: Optional[dict[str, int]]
+    pad_size: Annotated[Optional[Union[int, list[int], tuple[int, ...], dict[str, int]]], image_size_validator()]
     do_center_crop: Optional[bool]
-    data_format: Optional[ChannelDimension]
+    data_format: Optional[Union[str, ChannelDimension]]
     input_data_format: Optional[Union[str, ChannelDimension]]
-    device: Optional[Union[str, "torch.device"]]
+    device: Annotated[Optional[str], device_validator()]
+    return_tensors: Annotated[Optional[Union[str, TensorType]], tensor_type_validator()]
     disable_grouping: Optional[bool]
-    return_tensors: Optional[Union[str, TensorType]]
 
 
 class VideosKwargs(TypedDict, total=False):
@@ -244,9 +263,9 @@ class VideosKwargs(TypedDict, total=False):
             Scale factor to use if rescaling the video.
         do_normalize (`bool`, *optional*):
             Whether to normalize the video.
-        image_mean (`float` or `list[float]`, *optional*):
+        image_mean (`float` or `list[float] or tuple[float, float, float]`, *optional*):
             Mean to use if normalizing the video.
-        image_std (`float` or `list[float]`, *optional*):
+        image_std (`float` or `list[float] or tuple[float, float, float]`, *optional*):
             Standard deviation to use if normalizing the video.
         do_center_crop (`bool`, *optional*):
             Whether to center crop the video.
@@ -268,32 +287,36 @@ class VideosKwargs(TypedDict, total=False):
             The channel dimension format for the input video.
         device (`Union[str, torch.Tensor]`, *optional*):
             The device to use for processing (e.g. "cpu", "cuda"), only relevant for fast image processing.
-        return_metadata (`ChannelDimension` or `str`, *optional*):
+        return_metadata (`bool`, *optional*):
             Whether to return video metadata or not.
+        return_tensors (`str` or [`~utils.TensorType`], *optional*):
+            If set, will return tensors of a particular framework. Acceptable values are:
+            - `'pt'`: Return PyTorch `torch.Tensor` objects.
+            - `'np'`: Return NumPy `np.ndarray` objects.
     """
 
     do_convert_rgb: Optional[bool]
     do_resize: Optional[bool]
-    size: Optional[dict[str, int]]
+    size: Annotated[Optional[Union[int, list[int], tuple[int, ...], dict[str, int]]], image_size_validator()]
     default_to_square: Optional[bool]
-    resample: Optional["PILImageResampling"]
+    resample: Annotated[Optional[Union["PILImageResampling", int]], resampling_validator()]
     do_rescale: Optional[bool]
     rescale_factor: Optional[float]
     do_normalize: Optional[bool]
-    image_mean: Optional[Union[float, list[float]]]
-    image_std: Optional[Union[float, list[float]]]
+    image_mean: Optional[Union[float, list[float], tuple[float, ...]]]
+    image_std: Optional[Union[float, list[float], tuple[float, ...]]]
     do_center_crop: Optional[bool]
     do_pad: Optional[bool]
-    crop_size: Optional[dict[str, int]]
-    data_format: Optional[ChannelDimension]
+    crop_size: Annotated[Optional[Union[int, list[int], tuple[int, ...], dict[str, int]]], image_size_validator()]
+    data_format: Optional[Union[str, ChannelDimension]]
     input_data_format: Optional[Union[str, ChannelDimension]]
-    device: Optional[Union[str, "torch.device"]]
+    device: Annotated[Optional[str], device_validator()]
     do_sample_frames: Optional[bool]
-    video_metadata: Optional[Union[VideoMetadata, dict]]
-    fps: Optional[Union[int, float]]
-    num_frames: Optional[int]
+    video_metadata: Annotated[Optional[VideoMetadataType], video_metadata_validator()]
+    fps: Annotated[Optional[Union[int, float]], positive_any_number()]
+    num_frames: Annotated[Optional[int], positive_int()]
     return_metadata: Optional[bool]
-    return_tensors: Optional[Union[str, TensorType]]
+    return_tensors: Annotated[Optional[Union[str, TensorType]], tensor_type_validator()]
 
 
 class AudioKwargs(TypedDict, total=False):
@@ -324,16 +347,20 @@ class AudioKwargs(TypedDict, total=False):
             If set, will pad the sequence to a multiple of the provided value.
         return_attention_mask (`bool`, *optional*):
             Whether or not [`~ASTFeatureExtractor.__call__`] should return `attention_mask`.
+        return_tensors (`str` or [`~utils.TensorType`], *optional*):
+            If set, will return tensors of a particular framework. Acceptable values are:
+            - `'pt'`: Return PyTorch `torch.Tensor` objects.
+            - `'np'`: Return NumPy `np.ndarray` objects.
     """
 
-    sampling_rate: Optional[int]
-    raw_speech: Optional[Union[np.ndarray, list[float], list[np.ndarray], list[list[float]]]]
-    padding: Optional[Union[bool, str, PaddingStrategy]]
-    max_length: Optional[int]
-    truncation: Optional[bool]
-    pad_to_multiple_of: Optional[int]
+    sampling_rate: Annotated[Optional[int], positive_int()]
+    raw_speech: Optional[Union["np.ndarray", list[float], list["np.ndarray"], list[list[float]]]]
+    padding: Annotated[Optional[Union[bool, str, PaddingStrategy]], padding_validator()]
+    max_length: Annotated[Optional[int], positive_int()]
+    truncation: Annotated[Optional[Union[bool, str, TruncationStrategy]], truncation_validator()]
+    pad_to_multiple_of: Annotated[Optional[int], positive_int()]
     return_attention_mask: Optional[bool]
-    return_tensors: Optional[Union[str, TensorType]]
+    return_tensors: Annotated[Optional[Union[str, TensorType]], tensor_type_validator()]
 
 
 class ProcessingKwargs(TypedDict, total=False):
@@ -1361,6 +1388,18 @@ class ProcessorMixin(PushToHubMixin):
                         f"Keyword argument `{key}` is not a valid argument for this processor and will be ignored."
                     )
 
+        for key, typed_dict_obj in ModelProcessorKwargs.__annotations__.items():
+            if key in map_preprocessor_kwargs:
+                preprocessor = getattr(self, map_preprocessor_kwargs[key], None)
+                if preprocessor is None or getattr(preprocessor, "valid_kwargs", None) is None:
+                    continue
+                preprocessor_typed_dict_obj = getattr(preprocessor, "valid_kwargs")
+                typed_dict_obj = TypedDict(
+                    "merged_typed_dict",
+                    {**preprocessor_typed_dict_obj.__annotations__, **typed_dict_obj.__annotations__},
+                    total=False,
+                )
+            validate_typed_dict(typed_dict_obj, output_kwargs[key])
         return output_kwargs
 
     @classmethod

src/transformers/utils/type_validators.py

@@ -0,0 +1,115 @@
+from collections.abc import Sequence
+from typing import Optional, Union
+
+from ..tokenization_utils_base import PaddingStrategy, TruncationStrategy
+from ..video_utils import VideoMetadataType
+from .generic import TensorType
+from .import_utils import is_vision_available
+
+
+if is_vision_available():
+    from ..image_utils import PILImageResampling
+
+
+def positive_any_number(value: Optional[Union[int, float]] = None):
+    if value is not None and (not isinstance(value, (int, float)) or not value >= 0):
+        raise ValueError(f"Value must be a positive integer or floating number, got {value}")
+
+
+def positive_int(value: Optional[int] = None):
+    if value is not None and (not isinstance(value, int) or not value >= 0):
+        raise ValueError(f"Value must be a positive integer, got {value}")
+
+
+def padding_validator(value: Optional[Union[bool, str, PaddingStrategy]] = None):
+    possible_names = ["longest", "max_length", "do_not_pad"]
+    if value is None:
+        pass
+    elif not isinstance(value, (bool, str, PaddingStrategy)):
+        raise ValueError("Value for padding must be either a boolean, a string or a `PaddingStrategy`")
+    elif isinstance(value, str) and value not in possible_names:
+        raise ValueError(f"If padding is a string, the value must be one of {possible_names}")
+
+
+def truncation_validator(value: Optional[Union[bool, str, TruncationStrategy]] = None):
+    possible_names = ["only_first", "only_second", "longest_first", "do_not_truncate"]
+    if value is None:
+        pass
+    elif not isinstance(value, (bool, str, TruncationStrategy)):
+        raise ValueError("Value for truncation must be either a boolean, a string or a `TruncationStrategy`")
+    elif isinstance(value, str) and value not in possible_names:
+        raise ValueError(f"If truncation is a string, value must be one of {possible_names}")
+
+
+def image_size_validator(value: Optional[Union[int, Sequence[int], dict[str, int]]] = None):
+    possible_keys = ["height", "width", "longest_edge", "shortest_edge", "max_height", "max_width"]
+    if value is None:
+        pass
+    elif isinstance(value, dict) and any(k not in possible_keys for k in value.keys()):
+        raise ValueError(f"Value for size must be a dict with keys {possible_keys} but got size={value}")
+
+
+def device_validator(value: Optional[Union[str, int]] = None):
+    possible_names = ["cpu", "cuda", "xla", "xpu", "mps", "meta"]
+    if value is None:
+        pass
+    elif isinstance(value, int) and value < 0:
+        raise ValueError(
+            f"If device is an integer, the value must be a strictly positive integer but got device={value}"
+        )
+    elif isinstance(value, str) and value.split(":")[0] not in possible_names:
+        raise ValueError(f"If device is an string, the value must be one of {possible_names} but got device={value}")
+    elif not isinstance(value, (int, str)):
+        raise ValueError(
+            f"Device must be either an integer device ID or a string (e.g., 'cpu', 'cuda:0'), but got device={value}"
+        )
+
+
+def resampling_validator(value: Optional[Union[int, "PILImageResampling"]] = None):
+    if value is None:
+        pass
+    elif isinstance(value, int) and value not in list(range(6)):
+        raise ValueError(
+            f"The resampling should be one of {list(range(6))} when provided as integer, but got resampling={value}"
+        )
+    elif is_vision_available() and not isinstance(value, (PILImageResampling, int)):
+        raise ValueError(f"The resampling should an integer or `PIL.Image.Resampling`, but got resampling={value}")
+
+
+def video_metadata_validator(value: Optional[VideoMetadataType] = None):
+    if value is None:
+        return
+
+    valid_keys = ["total_num_frames", "fps", "width", "height", "duration", "video_backend", "frames_indices"]
+
+    def check_dict_keys(d: dict) -> bool:
+        return all(key in valid_keys for key in d.keys())
+
+    if isinstance(value, Sequence) and isinstance(value[0], Sequence) and isinstance(value[0][0], dict):
+        for sublist in value:
+            for item in sublist:
+                if not check_dict_keys(item):
+                    raise ValueError(
+                        f"Invalid keys found in video metadata. Valid keys: {valid_keys} got: {list(item.keys())}"
+                    )
+
+    elif isinstance(value, Sequence) and isinstance(value[0], dict):
+        for item in value:
+            if not check_dict_keys(item):
+                raise ValueError(
+                    f"Invalid keys found in video metadata. Valid keys: {valid_keys} got: {list(item.keys())}"
+                )
+
+    elif isinstance(value, dict):
+        if not check_dict_keys(value):
+            raise ValueError(
+                f"Invalid keys found in video metadata. Valid keys: {valid_keys}, got: {list(value.keys())}"
+            )
+
+
+def tensor_type_validator(value: Optional[Union[str, TensorType]] = None):
+    possible_names = ["pt", "np", "mlx"]
+    if value is None:
+        pass
+    elif not isinstance(value, str) or value not in possible_names:
+        raise ValueError(f"The tensor type should be one of {possible_names} but got tensor_type={value}")

src/transformers/video_processing_utils.py

@@ -21,6 +21,7 @@ from functools import partial
 from typing import Any, Callable, Optional, Union
 
 import numpy as np
+from huggingface_hub.dataclasses import validate_typed_dict
 
 from .dynamic_module_utils import custom_object_save
 from .image_processing_utils import (
@@ -358,6 +359,10 @@ class BaseVideoProcessor(BaseImageProcessorFast):
             captured_kwargs=kwargs.keys(),
             valid_processor_keys=list(self.valid_kwargs.__annotations__.keys()) + ["return_tensors"],
         )
+
+        # Perform type validation on received kwargs
+        validate_typed_dict(self.valid_kwargs, kwargs)
+
         # Set default kwargs from self. This ensures that if a kwarg is not provided
         # by the user, it gets its default value from the instance, or is set to None.
         for kwarg_name in self.valid_kwargs.__annotations__:

src/transformers/video_utils.py

@@ -112,6 +112,11 @@ class VideoMetadata(Mapping):
                 setattr(self, key, value)
 
 
+VideoMetadataType = Union[
+    VideoMetadata, dict, list[Union[dict, VideoMetadata]], list[list[Union[dict, VideoMetadata]]]
+]
+
+
 def is_valid_video_frame(frame):
     return isinstance(frame, PIL.Image.Image) or (
         (is_numpy_array(frame) or is_torch_tensor(frame)) and frame.ndim == 3
@@ -217,7 +222,7 @@ def make_batched_videos(videos) -> list[Union[np.ndarray, "torch.Tensor", "URL",
     return flat_videos_list
 
 
-def make_batched_metadata(videos: VideoInput, video_metadata: Union[VideoMetadata, dict]):
+def make_batched_metadata(videos: VideoInput, video_metadata: VideoMetadataType) -> list[VideoMetadata]:
     if video_metadata is None:
         # Create default metadata and fill attributes we can infer from given video
         video_metadata = [

tests/models/cohere2_vision/test_image_processing_cohere2_vision.py

@@ -176,8 +176,8 @@ class Cohere2VisionProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
                 image_inputs[0],
                 return_tensors="pt",
                 input_data_format="channels_last",
-                image_mean=0,
-                image_std=1,
+                image_mean=(0.0, 0.0, 0.0, 0.0),
+                image_std=(1.0, 1.0, 1.0, 1.0),
             ).pixel_values
             self.assertEqual(tuple(encoded_images.shape), (10, 4, 30, 30))
 
@@ -186,7 +186,7 @@ class Cohere2VisionProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
                 image_inputs,
                 return_tensors="pt",
                 input_data_format="channels_last",
-                image_mean=0,
-                image_std=1,
+                image_mean=(0.0, 0.0, 0.0, 0.0),
+                image_std=(1.0, 1.0, 1.0, 1.0),
             ).pixel_values
             self.assertEqual(tuple(encoded_images.shape), (70, 4, 30, 30))

tests/models/colpali/test_processing_colpali.py

@@ -133,7 +133,7 @@ class ColPaliProcessorTest(ProcessorTesterMixin, unittest.TestCase):
 
     def test_image_processor_defaults_preserved_by_image_kwargs(self):
         """
-        We use do_rescale=True, rescale_factor=-1 to ensure that image_processor kwargs are preserved in the processor.
+        We use do_rescale=True, rescale_factor=-1.0 to ensure that image_processor kwargs are preserved in the processor.
         We then check that the mean of the pixel_values is less than or equal to 0 after processing.
         Since the original pixel_values are in [0, 255], this is a good indicator that the rescale_factor is indeed applied.
         """
@@ -141,7 +141,7 @@ class ColPaliProcessorTest(ProcessorTesterMixin, unittest.TestCase):
             self.skipTest(f"image_processor attribute not present in {self.processor_class}")
         processor_components = self.prepare_components()
         processor_components["image_processor"] = self.get_component(
-            "image_processor", do_rescale=True, rescale_factor=-1
+            "image_processor", do_rescale=True, rescale_factor=-1.0
         )
         processor_components["tokenizer"] = self.get_component("tokenizer", max_length=117, padding="max_length")
 
@@ -179,7 +179,7 @@ class ColPaliProcessorTest(ProcessorTesterMixin, unittest.TestCase):
 
         image_input = self.prepare_image_inputs()
 
-        inputs = processor(images=image_input, do_rescale=True, rescale_factor=-1, return_tensors="pt")
+        inputs = processor(images=image_input, do_rescale=True, rescale_factor=-1.0, return_tensors="pt")
         self.assertLessEqual(inputs[self.images_input_name][0][0].mean(), 0)
 
     def test_unstructured_kwargs(self):
@@ -194,7 +194,7 @@ class ColPaliProcessorTest(ProcessorTesterMixin, unittest.TestCase):
             text=input_str,
             return_tensors="pt",
             do_rescale=True,
-            rescale_factor=-1,
+            rescale_factor=-1.0,
             padding="max_length",
             max_length=76,
         )
@@ -213,7 +213,7 @@ class ColPaliProcessorTest(ProcessorTesterMixin, unittest.TestCase):
             images=image_input,
             return_tensors="pt",
             do_rescale=True,
-            rescale_factor=-1,
+            rescale_factor=-1.0,
             padding="longest",
             max_length=76,
         )
@@ -231,7 +231,7 @@ class ColPaliProcessorTest(ProcessorTesterMixin, unittest.TestCase):
         with self.assertRaises(ValueError):
             _ = processor(
                 images=image_input,
-                images_kwargs={"do_rescale": True, "rescale_factor": -1},
+                images_kwargs={"do_rescale": True, "rescale_factor": -1.0},
                 do_rescale=True,
                 return_tensors="pt",
             )
@@ -248,7 +248,7 @@ class ColPaliProcessorTest(ProcessorTesterMixin, unittest.TestCase):
         # Define the kwargs for each modality
         all_kwargs = {
             "common_kwargs": {"return_tensors": "pt"},
-            "images_kwargs": {"do_rescale": True, "rescale_factor": -1},
+            "images_kwargs": {"do_rescale": True, "rescale_factor": -1.0},
             "text_kwargs": {"padding": "max_length", "max_length": 76},
         }
 
@@ -268,7 +268,7 @@ class ColPaliProcessorTest(ProcessorTesterMixin, unittest.TestCase):
         # Define the kwargs for each modality
         all_kwargs = {
             "common_kwargs": {"return_tensors": "pt"},
-            "images_kwargs": {"do_rescale": True, "rescale_factor": -1},
+            "images_kwargs": {"do_rescale": True, "rescale_factor": -1.0},
             "text_kwargs": {"padding": "max_length", "max_length": 76},
         }
 

tests/models/colqwen2/test_processing_colqwen2.py

@@ -132,7 +132,7 @@ class ColQwen2ProcessorTest(ProcessorTesterMixin, unittest.TestCase):
 
     def test_image_processor_defaults_preserved_by_image_kwargs(self):
         """
-        We use do_rescale=True, rescale_factor=-1 to ensure that image_processor kwargs are preserved in the processor.
+        We use do_rescale=True, rescale_factor=-1.0 to ensure that image_processor kwargs are preserved in the processor.
         We then check that the mean of the pixel_values is less than or equal to 0 after processing.
         Since the original pixel_values are in [0, 255], this is a good indicator that the rescale_factor is indeed applied.
         """
@@ -140,7 +140,7 @@ class ColQwen2ProcessorTest(ProcessorTesterMixin, unittest.TestCase):
             self.skipTest(f"image_processor attribute not present in {self.processor_class}")
         processor_components = self.prepare_components()
         processor_components["image_processor"] = self.get_component(
-            "image_processor", do_rescale=True, rescale_factor=-1
+            "image_processor", do_rescale=True, rescale_factor=-1.0
         )
         processor_components["tokenizer"] = self.get_component("tokenizer", max_length=117, padding="max_length")
 
@@ -178,7 +178,7 @@ class ColQwen2ProcessorTest(ProcessorTesterMixin, unittest.TestCase):
 
         image_input = self.prepare_image_inputs()
 
-        inputs = processor(images=image_input, do_rescale=True, rescale_factor=-1, return_tensors="pt")
+        inputs = processor(images=image_input, do_rescale=True, rescale_factor=-1.0, return_tensors="pt")
         self.assertLessEqual(inputs[self.images_input_name][0][0].mean(), 0)
 
     def test_unstructured_kwargs(self):
@@ -193,7 +193,7 @@ class ColQwen2ProcessorTest(ProcessorTesterMixin, unittest.TestCase):
             text=input_str,
             return_tensors="pt",
             do_rescale=True,
-            rescale_factor=-1,
+            rescale_factor=-1.0,
             padding="max_length",
             max_length=76,
         )
@@ -212,7 +212,7 @@ class ColQwen2ProcessorTest(ProcessorTesterMixin, unittest.TestCase):
             images=image_input,
             return_tensors="pt",
             do_rescale=True,
-            rescale_factor=-1,
+            rescale_factor=-1.0,
             padding="longest",
             max_length=76,
         )
@@ -230,7 +230,7 @@ class ColQwen2ProcessorTest(ProcessorTesterMixin, unittest.TestCase):
         with self.assertRaises(ValueError):
             _ = processor(
                 images=image_input,
-                images_kwargs={"do_rescale": True, "rescale_factor": -1},
+                images_kwargs={"do_rescale": True, "rescale_factor": -1.0},
                 do_rescale=True,
                 return_tensors="pt",
             )
@@ -247,7 +247,7 @@ class ColQwen2ProcessorTest(ProcessorTesterMixin, unittest.TestCase):
         # Define the kwargs for each modality
         all_kwargs = {
             "common_kwargs": {"return_tensors": "pt"},
-            "images_kwargs": {"do_rescale": True, "rescale_factor": -1},
+            "images_kwargs": {"do_rescale": True, "rescale_factor": -1.0},
             "text_kwargs": {"padding": "max_length", "max_length": 76},
         }
 
@@ -267,7 +267,7 @@ class ColQwen2ProcessorTest(ProcessorTesterMixin, unittest.TestCase):
         # Define the kwargs for each modality
         all_kwargs = {
             "common_kwargs": {"return_tensors": "pt"},
-            "images_kwargs": {"do_rescale": True, "rescale_factor": -1},
+            "images_kwargs": {"do_rescale": True, "rescale_factor": -1.0},
             "text_kwargs": {"padding": "max_length", "max_length": 76},
         }
 

tests/models/glm4v/test_image_processing_glm4v.py

@@ -236,8 +236,8 @@ class ViTImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
                 image_inputs[0],
                 return_tensors="pt",
                 input_data_format="channels_last",
-                image_mean=0,
-                image_std=1,
+                image_mean=(0.0, 0.0, 0.0, 0.0),
+                image_std=(1.0, 1.0, 1.0, 1.0),
             ).pixel_values
             expected_output_image_shape = self.image_processor_tester.expected_output_image_shape([image_inputs[0]])
             self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)
@@ -247,8 +247,8 @@ class ViTImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
                 image_inputs,
                 return_tensors="pt",
                 input_data_format="channels_last",
-                image_mean=0,
-                image_std=1,
+                image_mean=(0.0, 0.0, 0.0, 0.0),
+                image_std=(1.0, 1.0, 1.0, 1.0),
             ).pixel_values
             expected_output_image_shape = self.image_processor_tester.expected_output_image_shape(image_inputs)
             self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)

tests/models/glm4v/test_video_processing_glm4v.py

@@ -250,8 +250,8 @@ class Glm4vVideoProcessingTest(VideoProcessingTestMixin, unittest.TestCase):
                 video_inputs[0],
                 return_tensors="pt",
                 input_data_format="channels_last",
-                image_mean=0,
-                image_std=1,
+                image_mean=(0.0, 0.0, 0.0, 0.0),
+                image_std=(1.0, 1.0, 1.0, 1.0),
             )[self.input_name]
             expected_output_video_shape = self.video_processor_tester.expected_output_video_shape([video_inputs[0]])
             self.assertEqual(list(encoded_videos.shape), expected_output_video_shape)
@@ -261,8 +261,8 @@ class Glm4vVideoProcessingTest(VideoProcessingTestMixin, unittest.TestCase):
                 video_inputs,
                 return_tensors="pt",
                 input_data_format="channels_last",
-                image_mean=0,
-                image_std=1,
+                image_mean=(0.0, 0.0, 0.0, 0.0),
+                image_std=(1.0, 1.0, 1.0, 1.0),
             )[self.input_name]
             expected_output_video_shape = self.video_processor_tester.expected_output_video_shape(video_inputs)
             self.assertEqual(list(encoded_videos.shape), expected_output_video_shape)

tests/models/janus/test_processing_janus.py

@@ -444,7 +444,7 @@ class JanusProcessorTest(ProcessorTesterMixin, unittest.TestCase):
             tokenize=True,
             return_dict=True,
             do_rescale=True,
-            rescale_factor=-1,
+            rescale_factor=-1.0,
             return_tensors="np",
         )
         self.assertLessEqual(out_dict[self.images_input_name][0][0].mean(), 0)

tests/models/lfm2_vl/test_processing_lfm2_vl.py

@@ -100,7 +100,7 @@ class Lfm2VlProcessorTest(ProcessorTesterMixin, unittest.TestCase):
             "{{'<|im_start|>assistant\n' }}"
             "{% endif %}"
         )
-        return {"chat_template": chat_template, "use_image_special_tokens": True}
+        return {"chat_template": chat_template}
 
     # Override as Lfm2VL needs images/video to be an explicitly nested batch
     def prepare_image_inputs(self, batch_size=None):

tests/models/mllama/test_processing_mllama.py

@@ -386,7 +386,7 @@ class MllamaProcessorTest(ProcessorTesterMixin, unittest.TestCase):
             images=image_input,
             return_tensors="pt",
             do_rescale=True,
-            rescale_factor=-1,
+            rescale_factor=-1.0,
             padding="longest",
             max_length=76,
         )