# Copyright 2026 Agnes AI. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Image processor for Agnes 3.0 Flash: dynamic-resolution patching.""" import math from collections.abc import Iterable import torch from torchvision.transforms.v2 import functional as tvF from transformers.image_processing_backends import TorchvisionBackend from transformers.image_processing_utils import BatchFeature from transformers.image_transforms import group_images_by_shape, reorder_images from transformers.image_utils import ImageInput, PILImageResampling, SizeDict from transformers.processing_utils import ImagesKwargs, Unpack from transformers.utils import TensorType, auto_docstring class AgnesImageProcessorKwargs(ImagesKwargs, total=False): r""" min_pixels (`int`, *optional*, defaults to `256 * 256`): Lower bound on the pixel count after resizing. max_pixels (`int`, *optional*, defaults to `4096 * 4096`): Upper bound on the pixel count after resizing. patch_size (`int`, *optional*, defaults to 16): Spatial patch size of the vision tower. temporal_patch_size (`int`, *optional*, defaults to 2): Temporal patch size of the vision tower (images are duplicated to fill it). merge_size (`int`, *optional*, defaults to 2): Side of the patch square merged into one language-model token. """ min_pixels: int max_pixels: int patch_size: int temporal_patch_size: int merge_size: int def fit_to_grid(height: int, width: int, factor: int = 32, min_pixels: int = 256 * 256, max_pixels: int = 4096 * 4096): """Pick a (height, width) that is a multiple of `factor` on both sides, keeps the pixel count inside [min_pixels, max_pixels] and stays as close as possible to the original aspect ratio.""" if max(height, width) / min(height, width) > 200: raise ValueError(f"absolute aspect ratio must be smaller than 200, got {max(height, width) / min(height, width)}") h = round(height / factor) * factor w = round(width / factor) * factor if h * w > max_pixels: scale = math.sqrt((height * width) / max_pixels) h = max(factor, math.floor(height / scale / factor) * factor) w = max(factor, math.floor(width / scale / factor) * factor) elif h * w < min_pixels: scale = math.sqrt(min_pixels / (height * width)) h = math.ceil(height * scale / factor) * factor w = math.ceil(width * scale / factor) * factor return h, w @auto_docstring class AgnesImageProcessor(TorchvisionBackend): do_resize = True resample = PILImageResampling.BICUBIC size = {"shortest_edge": 256 * 256, "longest_edge": 4096 * 4096} default_to_square = False do_rescale = True do_normalize = True image_mean = [0.5, 0.5, 0.5] image_std = [0.5, 0.5, 0.5] do_convert_rgb = True patch_size = 16 temporal_patch_size = 2 merge_size = 2 valid_kwargs = AgnesImageProcessorKwargs model_input_names = ["pixel_values", "image_grid_thw"] def __init__(self, **kwargs: Unpack[AgnesImageProcessorKwargs]): size = kwargs.pop("size", None) min_pixels = kwargs.pop("min_pixels", None) max_pixels = kwargs.pop("max_pixels", None) size = self.size if size is None else size # min_pixels / max_pixels are the older spelling of the two size keys if min_pixels is not None: size["shortest_edge"] = min_pixels size.pop("min_pixels", None) if max_pixels is not None: size["longest_edge"] = max_pixels size.pop("max_pixels", None) if "shortest_edge" not in size or "longest_edge" not in size: raise ValueError("size must contain 'shortest_edge' and 'longest_edge' keys.") super().__init__(size=size, **kwargs) def _standardize_kwargs( self, size: int | Iterable[int] | dict[str, int] | SizeDict | None = None, min_pixels: int | None = None, max_pixels: int | None = None, **kwargs, ) -> dict: if min_pixels is not None and max_pixels is not None: size = SizeDict(shortest_edge=min_pixels, longest_edge=max_pixels) kwargs = super()._standardize_kwargs(size=size, **kwargs) size = kwargs.get("size", self.size) if not size.shortest_edge or not size.longest_edge: raise ValueError("size must contain 'shortest_edge' and 'longest_edge' keys.") return kwargs @auto_docstring def preprocess(self, images: ImageInput, **kwargs: Unpack[AgnesImageProcessorKwargs]) -> BatchFeature: return super().preprocess(images, **kwargs) def _preprocess( self, images: list["torch.Tensor"], do_resize: bool, size: SizeDict, resample: "PILImageResampling | tvF.InterpolationMode | int | None", do_rescale: bool, rescale_factor: float, do_normalize: bool, image_mean: float | list[float] | None, image_std: float | list[float] | None, patch_size: int, temporal_patch_size: int, merge_size: int, disable_grouping: bool | None, return_tensors: str | TensorType | None, **kwargs, ) -> BatchFeature: # 1. resize, batched per input shape by_shape, order = group_images_by_shape(images, disable_grouping=disable_grouping) resized = {} for shape, batch in by_shape.items(): height, width = batch.shape[-2:] if do_resize: new_h, new_w = fit_to_grid( height, width, factor=patch_size * merge_size, min_pixels=size.shortest_edge, max_pixels=size.longest_edge, ) batch = self.resize(image=batch, size=SizeDict(height=new_h, width=new_w), resample=resample) resized[shape] = batch images = reorder_images(resized, order) # 2. normalise and cut into merge-ordered patches, batched per resized shape by_shape, order = group_images_by_shape(images, disable_grouping=disable_grouping) flat = {} grids = {} for shape, batch in by_shape.items(): new_h, new_w = batch.shape[-2:] px = self.rescale_and_normalize(batch, do_rescale, rescale_factor, do_normalize, image_mean, image_std) n, c = px.shape[:2] gh, gw = new_h // patch_size, new_w // patch_size px = px.reshape(n, c, gh // merge_size, merge_size, patch_size, gw // merge_size, merge_size, patch_size) # -> [n, gh/merge, gw/merge, merge, merge, c, patch, patch]: patches of one # merge square end up adjacent in the flattened sequence px = px.permute(0, 2, 5, 3, 6, 1, 4, 7) px = ( px.unsqueeze(6) .expand(-1, -1, -1, -1, -1, -1, temporal_patch_size, -1, -1) .reshape(n, gh * gw, c * temporal_patch_size * patch_size * patch_size) ) flat[shape] = px grids[shape] = [[1, gh, gw]] * n pixel_values = torch.cat(reorder_images(flat, order), dim=0) image_grid_thw = torch.tensor(reorder_images(grids, order), dtype=torch.long) return BatchFeature(data={"pixel_values": pixel_values, "image_grid_thw": image_grid_thw}, tensor_type=return_tensors) def get_number_of_image_patches(self, height: int, width: int, images_kwargs=None): """Number of vision patches an image of this size produces (used by serving engines to lay out placeholders without running the processor).""" min_pixels = images_kwargs["min_pixels"] if "min_pixels" in images_kwargs else self.size["shortest_edge"] max_pixels = images_kwargs["max_pixels"] if "max_pixels" in images_kwargs else self.size["longest_edge"] patch_size = images_kwargs.get("patch_size", self.patch_size) merge_size = images_kwargs.get("merge_size", self.merge_size) new_h, new_w = fit_to_grid(height, width, patch_size * merge_size, min_pixels=min_pixels, max_pixels=max_pixels) return (new_h // patch_size) * (new_w // patch_size) __all__ = ["AgnesImageProcessor"]