Spaces:
Runtime error
Runtime error
Commit
·
1f54d7b
1
Parent(s):
aca81a2
Create src/pipeline_stable_diffusion_controlnet_inpaint.py
Browse files
src/ControlNetInpaint/src
DELETED
|
File without changes
|
src/ControlNetInpaint/src/pipeline_stable_diffusion_controlnet_inpaint.py
ADDED
|
@@ -0,0 +1,507 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
import PIL.Image
|
| 3 |
+
import numpy as np
|
| 4 |
+
|
| 5 |
+
from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_controlnet import *
|
| 6 |
+
|
| 7 |
+
EXAMPLE_DOC_STRING = """
|
| 8 |
+
Examples:
|
| 9 |
+
```py
|
| 10 |
+
>>> # !pip install opencv-python transformers accelerate
|
| 11 |
+
>>> from diffusers import StableDiffusionControlNetInpaintPipeline, ControlNetModel, UniPCMultistepScheduler
|
| 12 |
+
>>> from diffusers.utils import load_image
|
| 13 |
+
>>> import numpy as np
|
| 14 |
+
>>> import torch
|
| 15 |
+
|
| 16 |
+
>>> import cv2
|
| 17 |
+
>>> from PIL import Image
|
| 18 |
+
>>> # download an image
|
| 19 |
+
>>> image = load_image(
|
| 20 |
+
... "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png"
|
| 21 |
+
... )
|
| 22 |
+
>>> image = np.array(image)
|
| 23 |
+
>>> mask_image = load_image(
|
| 24 |
+
... "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"
|
| 25 |
+
... )
|
| 26 |
+
>>> mask_image = np.array(mask_image)
|
| 27 |
+
>>> # get canny image
|
| 28 |
+
>>> canny_image = cv2.Canny(image, 100, 200)
|
| 29 |
+
>>> canny_image = canny_image[:, :, None]
|
| 30 |
+
>>> canny_image = np.concatenate([canny_image, canny_image, canny_image], axis=2)
|
| 31 |
+
>>> canny_image = Image.fromarray(canny_image)
|
| 32 |
+
|
| 33 |
+
>>> # load control net and stable diffusion v1-5
|
| 34 |
+
>>> controlnet = ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-canny", torch_dtype=torch.float16)
|
| 35 |
+
>>> pipe = StableDiffusionControlNetInpaintPipeline.from_pretrained(
|
| 36 |
+
... "runwayml/stable-diffusion-inpainting", controlnet=controlnet, torch_dtype=torch.float16
|
| 37 |
+
... )
|
| 38 |
+
|
| 39 |
+
>>> # speed up diffusion process with faster scheduler and memory optimization
|
| 40 |
+
>>> pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
|
| 41 |
+
>>> # remove following line if xformers is not installed
|
| 42 |
+
>>> pipe.enable_xformers_memory_efficient_attention()
|
| 43 |
+
|
| 44 |
+
>>> pipe.enable_model_cpu_offload()
|
| 45 |
+
|
| 46 |
+
>>> # generate image
|
| 47 |
+
>>> generator = torch.manual_seed(0)
|
| 48 |
+
>>> image = pipe(
|
| 49 |
+
... "futuristic-looking doggo",
|
| 50 |
+
... num_inference_steps=20,
|
| 51 |
+
... generator=generator,
|
| 52 |
+
... image=image,
|
| 53 |
+
... control_image=canny_image,
|
| 54 |
+
... mask_image=mask_image
|
| 55 |
+
... ).images[0]
|
| 56 |
+
```
|
| 57 |
+
"""
|
| 58 |
+
|
| 59 |
+
|
| 60 |
+
def prepare_mask_and_masked_image(image, mask):
|
| 61 |
+
"""
|
| 62 |
+
Prepares a pair (image, mask) to be consumed by the Stable Diffusion pipeline. This means that those inputs will be
|
| 63 |
+
converted to ``torch.Tensor`` with shapes ``batch x channels x height x width`` where ``channels`` is ``3`` for the
|
| 64 |
+
``image`` and ``1`` for the ``mask``.
|
| 65 |
+
The ``image`` will be converted to ``torch.float32`` and normalized to be in ``[-1, 1]``. The ``mask`` will be
|
| 66 |
+
binarized (``mask > 0.5``) and cast to ``torch.float32`` too.
|
| 67 |
+
Args:
|
| 68 |
+
image (Union[np.array, PIL.Image, torch.Tensor]): The image to inpaint.
|
| 69 |
+
It can be a ``PIL.Image``, or a ``height x width x 3`` ``np.array`` or a ``channels x height x width``
|
| 70 |
+
``torch.Tensor`` or a ``batch x channels x height x width`` ``torch.Tensor``.
|
| 71 |
+
mask (_type_): The mask to apply to the image, i.e. regions to inpaint.
|
| 72 |
+
It can be a ``PIL.Image``, or a ``height x width`` ``np.array`` or a ``1 x height x width``
|
| 73 |
+
``torch.Tensor`` or a ``batch x 1 x height x width`` ``torch.Tensor``.
|
| 74 |
+
Raises:
|
| 75 |
+
ValueError: ``torch.Tensor`` images should be in the ``[-1, 1]`` range. ValueError: ``torch.Tensor`` mask
|
| 76 |
+
should be in the ``[0, 1]`` range. ValueError: ``mask`` and ``image`` should have the same spatial dimensions.
|
| 77 |
+
TypeError: ``mask`` is a ``torch.Tensor`` but ``image`` is not
|
| 78 |
+
(ot the other way around).
|
| 79 |
+
Returns:
|
| 80 |
+
tuple[torch.Tensor]: The pair (mask, masked_image) as ``torch.Tensor`` with 4
|
| 81 |
+
dimensions: ``batch x channels x height x width``.
|
| 82 |
+
"""
|
| 83 |
+
if isinstance(image, torch.Tensor):
|
| 84 |
+
if not isinstance(mask, torch.Tensor):
|
| 85 |
+
raise TypeError(f"`image` is a torch.Tensor but `mask` (type: {type(mask)} is not")
|
| 86 |
+
|
| 87 |
+
# Batch single image
|
| 88 |
+
if image.ndim == 3:
|
| 89 |
+
assert image.shape[0] == 3, "Image outside a batch should be of shape (3, H, W)"
|
| 90 |
+
image = image.unsqueeze(0)
|
| 91 |
+
|
| 92 |
+
# Batch and add channel dim for single mask
|
| 93 |
+
if mask.ndim == 2:
|
| 94 |
+
mask = mask.unsqueeze(0).unsqueeze(0)
|
| 95 |
+
|
| 96 |
+
# Batch single mask or add channel dim
|
| 97 |
+
if mask.ndim == 3:
|
| 98 |
+
# Single batched mask, no channel dim or single mask not batched but channel dim
|
| 99 |
+
if mask.shape[0] == 1:
|
| 100 |
+
mask = mask.unsqueeze(0)
|
| 101 |
+
|
| 102 |
+
# Batched masks no channel dim
|
| 103 |
+
else:
|
| 104 |
+
mask = mask.unsqueeze(1)
|
| 105 |
+
|
| 106 |
+
assert image.ndim == 4 and mask.ndim == 4, "Image and Mask must have 4 dimensions"
|
| 107 |
+
assert image.shape[-2:] == mask.shape[-2:], "Image and Mask must have the same spatial dimensions"
|
| 108 |
+
assert image.shape[0] == mask.shape[0], "Image and Mask must have the same batch size"
|
| 109 |
+
|
| 110 |
+
# Check image is in [-1, 1]
|
| 111 |
+
if image.min() < -1 or image.max() > 1:
|
| 112 |
+
raise ValueError("Image should be in [-1, 1] range")
|
| 113 |
+
|
| 114 |
+
# Check mask is in [0, 1]
|
| 115 |
+
if mask.min() < 0 or mask.max() > 1:
|
| 116 |
+
raise ValueError("Mask should be in [0, 1] range")
|
| 117 |
+
|
| 118 |
+
# Binarize mask
|
| 119 |
+
mask[mask < 0.5] = 0
|
| 120 |
+
mask[mask >= 0.5] = 1
|
| 121 |
+
|
| 122 |
+
# Image as float32
|
| 123 |
+
image = image.to(dtype=torch.float32)
|
| 124 |
+
elif isinstance(mask, torch.Tensor):
|
| 125 |
+
raise TypeError(f"`mask` is a torch.Tensor but `image` (type: {type(image)} is not")
|
| 126 |
+
else:
|
| 127 |
+
# preprocess image
|
| 128 |
+
if isinstance(image, (PIL.Image.Image, np.ndarray)):
|
| 129 |
+
image = [image]
|
| 130 |
+
|
| 131 |
+
if isinstance(image, list) and isinstance(image[0], PIL.Image.Image):
|
| 132 |
+
image = [np.array(i.convert("RGB"))[None, :] for i in image]
|
| 133 |
+
image = np.concatenate(image, axis=0)
|
| 134 |
+
elif isinstance(image, list) and isinstance(image[0], np.ndarray):
|
| 135 |
+
image = np.concatenate([i[None, :] for i in image], axis=0)
|
| 136 |
+
|
| 137 |
+
image = image.transpose(0, 3, 1, 2)
|
| 138 |
+
image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0
|
| 139 |
+
|
| 140 |
+
# preprocess mask
|
| 141 |
+
if isinstance(mask, (PIL.Image.Image, np.ndarray)):
|
| 142 |
+
mask = [mask]
|
| 143 |
+
|
| 144 |
+
if isinstance(mask, list) and isinstance(mask[0], PIL.Image.Image):
|
| 145 |
+
mask = np.concatenate([np.array(m.convert("L"))[None, None, :] for m in mask], axis=0)
|
| 146 |
+
mask = mask.astype(np.float32) / 255.0
|
| 147 |
+
elif isinstance(mask, list) and isinstance(mask[0], np.ndarray):
|
| 148 |
+
mask = np.concatenate([m[None, None, :] for m in mask], axis=0)
|
| 149 |
+
|
| 150 |
+
mask[mask < 0.5] = 0
|
| 151 |
+
mask[mask >= 0.5] = 1
|
| 152 |
+
mask = torch.from_numpy(mask)
|
| 153 |
+
|
| 154 |
+
masked_image = image * (mask < 0.5)
|
| 155 |
+
|
| 156 |
+
return mask, masked_image
|
| 157 |
+
|
| 158 |
+
class StableDiffusionControlNetInpaintPipeline(StableDiffusionControlNetPipeline):
|
| 159 |
+
r"""
|
| 160 |
+
Pipeline for text-guided image inpainting using Stable Diffusion with ControlNet guidance.
|
| 161 |
+
|
| 162 |
+
This model inherits from [`StableDiffusionControlNetPipeline`]. Check the superclass documentation for the generic methods the
|
| 163 |
+
library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
|
| 164 |
+
|
| 165 |
+
Args:
|
| 166 |
+
vae ([`AutoencoderKL`]):
|
| 167 |
+
Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
|
| 168 |
+
text_encoder ([`CLIPTextModel`]):
|
| 169 |
+
Frozen text-encoder. Stable Diffusion uses the text portion of
|
| 170 |
+
[CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
|
| 171 |
+
the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
|
| 172 |
+
tokenizer (`CLIPTokenizer`):
|
| 173 |
+
Tokenizer of class
|
| 174 |
+
[CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
|
| 175 |
+
unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
|
| 176 |
+
controlnet ([`ControlNetModel`]):
|
| 177 |
+
Provides additional conditioning to the unet during the denoising process
|
| 178 |
+
scheduler ([`SchedulerMixin`]):
|
| 179 |
+
A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
|
| 180 |
+
[`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
|
| 181 |
+
safety_checker ([`StableDiffusionSafetyChecker`]):
|
| 182 |
+
Classification module that estimates whether generated images could be considered offensive or harmful.
|
| 183 |
+
Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
|
| 184 |
+
feature_extractor ([`CLIPFeatureExtractor`]):
|
| 185 |
+
Model that extracts features from generated images to be used as inputs for the `safety_checker`.
|
| 186 |
+
"""
|
| 187 |
+
|
| 188 |
+
def prepare_mask_latents(
|
| 189 |
+
self, mask, masked_image, batch_size, height, width, dtype, device, generator, do_classifier_free_guidance
|
| 190 |
+
):
|
| 191 |
+
# resize the mask to latents shape as we concatenate the mask to the latents
|
| 192 |
+
# we do that before converting to dtype to avoid breaking in case we're using cpu_offload
|
| 193 |
+
# and half precision
|
| 194 |
+
mask = torch.nn.functional.interpolate(
|
| 195 |
+
mask, size=(height // self.vae_scale_factor, width // self.vae_scale_factor)
|
| 196 |
+
)
|
| 197 |
+
mask = mask.to(device=device, dtype=dtype)
|
| 198 |
+
|
| 199 |
+
masked_image = masked_image.to(device=device, dtype=dtype)
|
| 200 |
+
|
| 201 |
+
# encode the mask image into latents space so we can concatenate it to the latents
|
| 202 |
+
if isinstance(generator, list):
|
| 203 |
+
masked_image_latents = [
|
| 204 |
+
self.vae.encode(masked_image[i : i + 1]).latent_dist.sample(generator=generator[i])
|
| 205 |
+
for i in range(batch_size)
|
| 206 |
+
]
|
| 207 |
+
masked_image_latents = torch.cat(masked_image_latents, dim=0)
|
| 208 |
+
else:
|
| 209 |
+
masked_image_latents = self.vae.encode(masked_image).latent_dist.sample(generator=generator)
|
| 210 |
+
masked_image_latents = self.vae.config.scaling_factor * masked_image_latents
|
| 211 |
+
|
| 212 |
+
# duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method
|
| 213 |
+
if mask.shape[0] < batch_size:
|
| 214 |
+
if not batch_size % mask.shape[0] == 0:
|
| 215 |
+
raise ValueError(
|
| 216 |
+
"The passed mask and the required batch size don't match. Masks are supposed to be duplicated to"
|
| 217 |
+
f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number"
|
| 218 |
+
" of masks that you pass is divisible by the total requested batch size."
|
| 219 |
+
)
|
| 220 |
+
mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1)
|
| 221 |
+
if masked_image_latents.shape[0] < batch_size:
|
| 222 |
+
if not batch_size % masked_image_latents.shape[0] == 0:
|
| 223 |
+
raise ValueError(
|
| 224 |
+
"The passed images and the required batch size don't match. Images are supposed to be duplicated"
|
| 225 |
+
f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed."
|
| 226 |
+
" Make sure the number of images that you pass is divisible by the total requested batch size."
|
| 227 |
+
)
|
| 228 |
+
masked_image_latents = masked_image_latents.repeat(batch_size // masked_image_latents.shape[0], 1, 1, 1)
|
| 229 |
+
|
| 230 |
+
mask = torch.cat([mask] * 2) if do_classifier_free_guidance else mask
|
| 231 |
+
masked_image_latents = (
|
| 232 |
+
torch.cat([masked_image_latents] * 2) if do_classifier_free_guidance else masked_image_latents
|
| 233 |
+
)
|
| 234 |
+
|
| 235 |
+
# aligning device to prevent device errors when concating it with the latent model input
|
| 236 |
+
masked_image_latents = masked_image_latents.to(device=device, dtype=dtype)
|
| 237 |
+
return mask, masked_image_latents
|
| 238 |
+
|
| 239 |
+
@torch.no_grad()
|
| 240 |
+
@replace_example_docstring(EXAMPLE_DOC_STRING)
|
| 241 |
+
def __call__(
|
| 242 |
+
self,
|
| 243 |
+
prompt: Union[str, List[str]] = None,
|
| 244 |
+
image: Union[torch.FloatTensor, PIL.Image.Image] = None,
|
| 245 |
+
control_image: Union[torch.FloatTensor, PIL.Image.Image, List[torch.FloatTensor], List[PIL.Image.Image]] = None,
|
| 246 |
+
mask_image: Union[torch.FloatTensor, PIL.Image.Image] = None,
|
| 247 |
+
height: Optional[int] = None,
|
| 248 |
+
width: Optional[int] = None,
|
| 249 |
+
num_inference_steps: int = 50,
|
| 250 |
+
guidance_scale: float = 7.5,
|
| 251 |
+
negative_prompt: Optional[Union[str, List[str]]] = None,
|
| 252 |
+
num_images_per_prompt: Optional[int] = 1,
|
| 253 |
+
eta: float = 0.0,
|
| 254 |
+
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
|
| 255 |
+
latents: Optional[torch.FloatTensor] = None,
|
| 256 |
+
prompt_embeds: Optional[torch.FloatTensor] = None,
|
| 257 |
+
negative_prompt_embeds: Optional[torch.FloatTensor] = None,
|
| 258 |
+
output_type: Optional[str] = "pil",
|
| 259 |
+
return_dict: bool = True,
|
| 260 |
+
callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
|
| 261 |
+
callback_steps: int = 1,
|
| 262 |
+
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
|
| 263 |
+
controlnet_conditioning_scale: float = 1.0,
|
| 264 |
+
):
|
| 265 |
+
r"""
|
| 266 |
+
Function invoked when calling the pipeline for generation.
|
| 267 |
+
Args:
|
| 268 |
+
prompt (`str` or `List[str]`, *optional*):
|
| 269 |
+
The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
|
| 270 |
+
instead.
|
| 271 |
+
image (`PIL.Image.Image`):
|
| 272 |
+
`Image`, or tensor representing an image batch which will be inpainted, *i.e.* parts of the image will
|
| 273 |
+
be masked out with `mask_image` and repainted according to `prompt`.
|
| 274 |
+
control_image (`torch.FloatTensor`, `PIL.Image.Image`, `List[torch.FloatTensor]` or `List[PIL.Image.Image]`):
|
| 275 |
+
The ControlNet input condition. ControlNet uses this input condition to generate guidance to Unet. If
|
| 276 |
+
the type is specified as `Torch.FloatTensor`, it is passed to ControlNet as is. PIL.Image.Image` can
|
| 277 |
+
also be accepted as an image. The control image is automatically resized to fit the output image.
|
| 278 |
+
mask_image (`PIL.Image.Image`):
|
| 279 |
+
`Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be
|
| 280 |
+
repainted, while black pixels will be preserved. If `mask_image` is a PIL image, it will be converted
|
| 281 |
+
to a single channel (luminance) before use. If it's a tensor, it should contain one color channel (L)
|
| 282 |
+
instead of 3, so the expected shape would be `(B, H, W, 1)`.
|
| 283 |
+
height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
|
| 284 |
+
The height in pixels of the generated image.
|
| 285 |
+
width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
|
| 286 |
+
The width in pixels of the generated image.
|
| 287 |
+
num_inference_steps (`int`, *optional*, defaults to 50):
|
| 288 |
+
The number of denoising steps. More denoising steps usually lead to a higher quality image at the
|
| 289 |
+
expense of slower inference.
|
| 290 |
+
guidance_scale (`float`, *optional*, defaults to 7.5):
|
| 291 |
+
Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
|
| 292 |
+
`guidance_scale` is defined as `w` of equation 2. of [Imagen
|
| 293 |
+
Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
|
| 294 |
+
1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
|
| 295 |
+
usually at the expense of lower image quality.
|
| 296 |
+
negative_prompt (`str` or `List[str]`, *optional*):
|
| 297 |
+
The prompt or prompts not to guide the image generation. If not defined, one has to pass
|
| 298 |
+
`negative_prompt_embeds`. instead. If not defined, one has to pass `negative_prompt_embeds`. instead.
|
| 299 |
+
Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`).
|
| 300 |
+
num_images_per_prompt (`int`, *optional*, defaults to 1):
|
| 301 |
+
The number of images to generate per prompt.
|
| 302 |
+
eta (`float`, *optional*, defaults to 0.0):
|
| 303 |
+
Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
|
| 304 |
+
[`schedulers.DDIMScheduler`], will be ignored for others.
|
| 305 |
+
generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
|
| 306 |
+
One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
|
| 307 |
+
to make generation deterministic.
|
| 308 |
+
latents (`torch.FloatTensor`, *optional*):
|
| 309 |
+
Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
|
| 310 |
+
generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
|
| 311 |
+
tensor will ge generated by sampling using the supplied random `generator`.
|
| 312 |
+
prompt_embeds (`torch.FloatTensor`, *optional*):
|
| 313 |
+
Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
|
| 314 |
+
provided, text embeddings will be generated from `prompt` input argument.
|
| 315 |
+
negative_prompt_embeds (`torch.FloatTensor`, *optional*):
|
| 316 |
+
Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
|
| 317 |
+
weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
|
| 318 |
+
argument.
|
| 319 |
+
output_type (`str`, *optional*, defaults to `"pil"`):
|
| 320 |
+
The output format of the generate image. Choose between
|
| 321 |
+
[PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
|
| 322 |
+
return_dict (`bool`, *optional*, defaults to `True`):
|
| 323 |
+
Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
|
| 324 |
+
plain tuple.
|
| 325 |
+
callback (`Callable`, *optional*):
|
| 326 |
+
A function that will be called every `callback_steps` steps during inference. The function will be
|
| 327 |
+
called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
|
| 328 |
+
callback_steps (`int`, *optional*, defaults to 1):
|
| 329 |
+
The frequency at which the `callback` function will be called. If not specified, the callback will be
|
| 330 |
+
called at every step.
|
| 331 |
+
cross_attention_kwargs (`dict`, *optional*):
|
| 332 |
+
A kwargs dictionary that if specified is passed along to the `AttnProcessor` as defined under
|
| 333 |
+
`self.processor` in
|
| 334 |
+
[diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py).
|
| 335 |
+
controlnet_conditioning_scale (`float`, *optional*, defaults to 1.0):
|
| 336 |
+
The outputs of the controlnet are multiplied by `controlnet_conditioning_scale` before they are added
|
| 337 |
+
to the residual in the original unet.
|
| 338 |
+
Examples:
|
| 339 |
+
Returns:
|
| 340 |
+
[`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
|
| 341 |
+
[`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
|
| 342 |
+
When returning a tuple, the first element is a list with the generated images, and the second element is a
|
| 343 |
+
list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
|
| 344 |
+
(nsfw) content, according to the `safety_checker`.
|
| 345 |
+
"""
|
| 346 |
+
# 0. Default height and width to unet
|
| 347 |
+
height, width = self._default_height_width(height, width, control_image)
|
| 348 |
+
|
| 349 |
+
# 1. Check inputs. Raise error if not correct
|
| 350 |
+
self.check_inputs(
|
| 351 |
+
prompt, control_image, height, width, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds
|
| 352 |
+
)
|
| 353 |
+
|
| 354 |
+
# 2. Define call parameters
|
| 355 |
+
if prompt is not None and isinstance(prompt, str):
|
| 356 |
+
batch_size = 1
|
| 357 |
+
elif prompt is not None and isinstance(prompt, list):
|
| 358 |
+
batch_size = len(prompt)
|
| 359 |
+
else:
|
| 360 |
+
batch_size = prompt_embeds.shape[0]
|
| 361 |
+
|
| 362 |
+
device = self._execution_device
|
| 363 |
+
# here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
|
| 364 |
+
# of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
|
| 365 |
+
# corresponds to doing no classifier free guidance.
|
| 366 |
+
do_classifier_free_guidance = guidance_scale > 1.0
|
| 367 |
+
|
| 368 |
+
# 3. Encode input prompt
|
| 369 |
+
prompt_embeds = self._encode_prompt(
|
| 370 |
+
prompt,
|
| 371 |
+
device,
|
| 372 |
+
num_images_per_prompt,
|
| 373 |
+
do_classifier_free_guidance,
|
| 374 |
+
negative_prompt,
|
| 375 |
+
prompt_embeds=prompt_embeds,
|
| 376 |
+
negative_prompt_embeds=negative_prompt_embeds,
|
| 377 |
+
)
|
| 378 |
+
|
| 379 |
+
# 4. Prepare image
|
| 380 |
+
control_image = self.prepare_image(
|
| 381 |
+
control_image,
|
| 382 |
+
width,
|
| 383 |
+
height,
|
| 384 |
+
batch_size * num_images_per_prompt,
|
| 385 |
+
num_images_per_prompt,
|
| 386 |
+
device,
|
| 387 |
+
self.controlnet.dtype,
|
| 388 |
+
)
|
| 389 |
+
|
| 390 |
+
if do_classifier_free_guidance:
|
| 391 |
+
control_image = torch.cat([control_image] * 2)
|
| 392 |
+
|
| 393 |
+
# 5. Prepare timesteps
|
| 394 |
+
self.scheduler.set_timesteps(num_inference_steps, device=device)
|
| 395 |
+
timesteps = self.scheduler.timesteps
|
| 396 |
+
|
| 397 |
+
# 6. Prepare latent variables
|
| 398 |
+
num_channels_latents = self.controlnet.config.in_channels
|
| 399 |
+
latents = self.prepare_latents(
|
| 400 |
+
batch_size * num_images_per_prompt,
|
| 401 |
+
num_channels_latents,
|
| 402 |
+
height,
|
| 403 |
+
width,
|
| 404 |
+
prompt_embeds.dtype,
|
| 405 |
+
device,
|
| 406 |
+
generator,
|
| 407 |
+
latents,
|
| 408 |
+
)
|
| 409 |
+
|
| 410 |
+
# EXTRA: prepare mask latents
|
| 411 |
+
mask, masked_image = prepare_mask_and_masked_image(image, mask_image)
|
| 412 |
+
mask, masked_image_latents = self.prepare_mask_latents(
|
| 413 |
+
mask,
|
| 414 |
+
masked_image,
|
| 415 |
+
batch_size * num_images_per_prompt,
|
| 416 |
+
height,
|
| 417 |
+
width,
|
| 418 |
+
prompt_embeds.dtype,
|
| 419 |
+
device,
|
| 420 |
+
generator,
|
| 421 |
+
do_classifier_free_guidance,
|
| 422 |
+
)
|
| 423 |
+
|
| 424 |
+
# 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
|
| 425 |
+
extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
|
| 426 |
+
|
| 427 |
+
# 8. Denoising loop
|
| 428 |
+
num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
|
| 429 |
+
with self.progress_bar(total=num_inference_steps) as progress_bar:
|
| 430 |
+
for i, t in enumerate(timesteps):
|
| 431 |
+
# expand the latents if we are doing classifier free guidance
|
| 432 |
+
latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
|
| 433 |
+
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
|
| 434 |
+
|
| 435 |
+
down_block_res_samples, mid_block_res_sample = self.controlnet(
|
| 436 |
+
latent_model_input,
|
| 437 |
+
t,
|
| 438 |
+
encoder_hidden_states=prompt_embeds,
|
| 439 |
+
controlnet_cond=control_image,
|
| 440 |
+
return_dict=False,
|
| 441 |
+
)
|
| 442 |
+
|
| 443 |
+
down_block_res_samples = [
|
| 444 |
+
down_block_res_sample * controlnet_conditioning_scale
|
| 445 |
+
for down_block_res_sample in down_block_res_samples
|
| 446 |
+
]
|
| 447 |
+
mid_block_res_sample *= controlnet_conditioning_scale
|
| 448 |
+
|
| 449 |
+
# predict the noise residual
|
| 450 |
+
latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents], dim=1)
|
| 451 |
+
noise_pred = self.unet(
|
| 452 |
+
latent_model_input,
|
| 453 |
+
t,
|
| 454 |
+
encoder_hidden_states=prompt_embeds,
|
| 455 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
| 456 |
+
down_block_additional_residuals=down_block_res_samples,
|
| 457 |
+
mid_block_additional_residual=mid_block_res_sample,
|
| 458 |
+
).sample
|
| 459 |
+
|
| 460 |
+
# perform guidance
|
| 461 |
+
if do_classifier_free_guidance:
|
| 462 |
+
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
|
| 463 |
+
noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
|
| 464 |
+
|
| 465 |
+
# compute the previous noisy sample x_t -> x_t-1
|
| 466 |
+
latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
|
| 467 |
+
|
| 468 |
+
# call the callback, if provided
|
| 469 |
+
if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
|
| 470 |
+
progress_bar.update()
|
| 471 |
+
if callback is not None and i % callback_steps == 0:
|
| 472 |
+
callback(i, t, latents)
|
| 473 |
+
|
| 474 |
+
# If we do sequential model offloading, let's offload unet and controlnet
|
| 475 |
+
# manually for max memory savings
|
| 476 |
+
if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
|
| 477 |
+
self.unet.to("cpu")
|
| 478 |
+
self.controlnet.to("cpu")
|
| 479 |
+
torch.cuda.empty_cache()
|
| 480 |
+
|
| 481 |
+
if output_type == "latent":
|
| 482 |
+
image = latents
|
| 483 |
+
has_nsfw_concept = None
|
| 484 |
+
elif output_type == "pil":
|
| 485 |
+
# 8. Post-processing
|
| 486 |
+
image = self.decode_latents(latents)
|
| 487 |
+
|
| 488 |
+
# 9. Run safety checker
|
| 489 |
+
image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
|
| 490 |
+
|
| 491 |
+
# 10. Convert to PIL
|
| 492 |
+
image = self.numpy_to_pil(image)
|
| 493 |
+
else:
|
| 494 |
+
# 8. Post-processing
|
| 495 |
+
image = self.decode_latents(latents)
|
| 496 |
+
|
| 497 |
+
# 9. Run safety checker
|
| 498 |
+
image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
|
| 499 |
+
|
| 500 |
+
# Offload last model to CPU
|
| 501 |
+
if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
|
| 502 |
+
self.final_offload_hook.offload()
|
| 503 |
+
|
| 504 |
+
if not return_dict:
|
| 505 |
+
return (image, has_nsfw_concept)
|
| 506 |
+
|
| 507 |
+
return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
|