# Copyright (c) 2023, Teriks
#
# dgenerate is distributed under the following BSD 3-Clause License
#
# Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:
#
# 1. Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer.
#
# 2. Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in
# the documentation and/or other materials provided with the distribution.
#
# 3. Neither the name of the copyright holder nor the names of its contributors may be used to endorse or promote products derived
# from this software without specific prior written permission.
#
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
# HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
# ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
import PIL.Image
import dgenerate.extras.controlnet_aux as _cna
import dgenerate.extras.controlnet_aux.util as _cna_util
import cv2
import einops
import numpy
import torch
import dgenerate.hfhub as _hfhub
import dgenerate.image as _image
import dgenerate.textprocessing as _textprocessing
import dgenerate.types as _types
from dgenerate.imageprocessors import imageprocessor as _imageprocessor
[docs]
class HEDProcessor(_imageprocessor.ImageProcessor):
"""
HED detection (holistically-nested edge detection), this is an edge detection algorithm
that can produced something akin to thick lineart.
The "scribble" argument determines whether scribble mode is enabled, this produces thicker lines.
The "safe" argument enables or disables numerically safe / more precise stepping.
The "detect-resolution" argument is the resolution the image is resized to internal to the processor before
detection is run on it. It should be a single dimension for example: "detect-resolution=512" or the X/Y dimensions
seperated by an "x" character, like so: "detect-resolution=1024x512". If you do not specify this argument,
the detector runs on the input image at its full resolution. After processing the image will be resized to
whatever you have requested dgenerate resize it to via --output-size or --resize/--align in the case of the
image-process sub-command, if you have not requested any resizing the output will be resized back to the original
size of the input image.
The "detect-aspect" argument determines if the image resize requested by "detect_resolution" before
detection runs is aspect correct, this defaults to true.
The "detect-align" argument determines the pixel alignment of the image resize requested by
"detect_resolution", it defaults to 1 indicating no requested alignment.
The "pre-resize" argument determines if the processing occurs before or after dgenerate resizes the image.
This defaults to False, meaning the image is processed after dgenerate is done resizing it.
"""
NAMES = ['hed']
[docs]
def __init__(self,
scribble: bool = False,
safe: bool = False,
detect_resolution: str | None = None,
detect_aspect: bool = True,
detect_align: int = 1,
pre_resize: bool = False,
**kwargs):
"""
:param scribble: determines whether or not scribble mode is enabled, this produces thicker lines
:param safe: enables numerically safe / more precise stepping
:param detect_resolution: the input image is resized to this dimension before being processed,
providing ``None`` indicates it is not to be resized. If there is no resize requested during
the processing action via ``resize_resolution`` it will be resized back to its original size.
:param detect_aspect: if the input image is resized by ``detect_resolution`` or ``detect_align``
before processing, will it be an aspect correct resize?
:param detect_align: the input image is forcefully aligned to this amount of pixels
before being processed.
:param pre_resize: process the image before it is resized, or after? default is ``False`` (after).
:param kwargs: forwarded to base class
"""
super().__init__(**kwargs)
if detect_align < 1:
raise self.argument_error('Argument "detect-align" may not be less than 1.')
self._detect_aspect = detect_aspect
self._detect_align = detect_align
self._safe = safe
self._scribble = scribble
self._pre_resize = pre_resize
if detect_resolution is not None:
try:
self._detect_resolution = _textprocessing.parse_image_size(detect_resolution)
except ValueError as e:
raise self.argument_error(
f'Could not parse the "detect-resolution" argument as an image dimension: {e}') from e
else:
self._detect_resolution = None
self.set_size_estimate(29.4 * (1000 ** 2)) # 29.4 MB -> bytes
with (_hfhub.with_hf_errors_as_model_not_found(),
_hfhub.offline_mode_context(self.local_files_only)):
self._hed = self.load_object_cached(
tag="lllyasviel/Annotators",
estimated_size=self.size_estimate,
method=lambda: _cna.HEDdetector.from_pretrained("lllyasviel/Annotators")
)
self.register_module(self._hed)
@torch.inference_mode()
def _process(self, image):
original_size = image.size
with image:
resized = _image.resize_image(
image,
self._detect_resolution,
aspect_correct=self._detect_aspect,
align=self._detect_align
)
image = resized
input_image = numpy.array(image, dtype=numpy.uint8)
input_image = _cna_util.HWC3(input_image)
H, W, C = input_image.shape
image_hed = torch.from_numpy(input_image.copy()).float().to(self.modules_device)
image_hed = einops.rearrange(image_hed, 'h w c -> 1 c h w')
edges = self._hed.netNetwork(image_hed)
image_hed.cpu()
del image_hed
edges = [e.detach().cpu().numpy().astype(numpy.float32)[0, 0] for e in edges]
edges = [_image.cv2_resize_image(e, (W, H)) for e in edges]
edges = numpy.stack(edges, axis=2)
edge = 1 / (1 + numpy.exp(-numpy.mean(edges, axis=2).astype(numpy.float64)))
if self._safe:
edge = _cna_util.safe_step(edge)
edge = (edge * 255.0).clip(0, 255).astype(numpy.uint8)
detected_map = edge
detected_map = _cna_util.HWC3(detected_map)
detected_map = _image.cv2_resize_image(detected_map, original_size)
if self._scribble:
detected_map = _cna_util.nms(detected_map, 127, 3.0)
detected_map = cv2.GaussianBlur(detected_map, (0, 0), 3.0)
detected_map[detected_map > 4] = 255
detected_map[detected_map < 255] = 0
return PIL.Image.fromarray(detected_map)
[docs]
def impl_pre_resize(self, image: PIL.Image.Image, resize_resolution: _types.OptionalSize):
"""
Pre resize.
:param image: image to process
:param resize_resolution: resize resolution
:return: possibly an HED depth detected image, or the input image
"""
if self._pre_resize:
return self._process(image)
return image
[docs]
def impl_post_resize(self, image: PIL.Image.Image):
"""
Post resize.
:param image: image
:return: possibly an HED depth detected image, or the input image
"""
if not self._pre_resize:
return self._process(image)
return image
__all__ = _types.module_all()