pipeline_controlnet_inpaint.py 80.8 KB
Newer Older
1
# Copyright 2024 The HuggingFace Team. All rights reserved.
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
#
# 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.

# This model implementation is heavily inspired by https://github.com/haofanwang/ControlNet-for-Diffusers/

import inspect
from typing import Any, Callable, Dict, List, Optional, Tuple, Union

import numpy as np
import PIL.Image
import torch
import torch.nn.functional as F
24
from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
25

26
from ...image_processor import PipelineImageInput, VaeImageProcessor
27
from ...loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
28
from ...models import AutoencoderKL, ControlNetModel, ImageProjection, UNet2DConditionModel
29
from ...models.lora import adjust_lora_scale_text_encoder
30
from ...schedulers import KarrasDiffusionSchedulers
31
32
33
34
35
36
37
38
from ...utils import (
    USE_PEFT_BACKEND,
    deprecate,
    logging,
    replace_example_docstring,
    scale_lora_layers,
    unscale_lora_layers,
)
Dhruv Nair's avatar
Dhruv Nair committed
39
from ...utils.torch_utils import is_compiled_module, randn_tensor
40
41
42
43
44
45
46
47
48
49
50
51
from ..pipeline_utils import DiffusionPipeline
from ..stable_diffusion import StableDiffusionPipelineOutput
from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
from .multicontrolnet import MultiControlNetModel


logger = logging.get_logger(__name__)  # pylint: disable=invalid-name


EXAMPLE_DOC_STRING = """
    Examples:
        ```py
52
53
        >>> # !pip install transformers accelerate
        >>> from diffusers import StableDiffusionControlNetInpaintPipeline, ControlNetModel, DDIMScheduler
54
55
56
57
        >>> from diffusers.utils import load_image
        >>> import numpy as np
        >>> import torch

58
59
60
61
62
63
64
65
66
67
68
69
        >>> init_image = load_image(
        ...     "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main/stable_diffusion_inpaint/boy.png"
        ... )
        >>> init_image = init_image.resize((512, 512))

        >>> generator = torch.Generator(device="cpu").manual_seed(1)

        >>> mask_image = load_image(
        ...     "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main/stable_diffusion_inpaint/boy_mask.png"
        ... )
        >>> mask_image = mask_image.resize((512, 512))

70

71
72
73
74
75
76
        >>> def make_canny_condition(image):
        ...     image = np.array(image)
        ...     image = cv2.Canny(image, 100, 200)
        ...     image = image[:, :, None]
        ...     image = np.concatenate([image, image, image], axis=2)
        ...     image = Image.fromarray(image)
77
        ...     return image
78
79


80
        >>> control_image = make_canny_condition(init_image)
81

82
83
84
        >>> controlnet = ControlNetModel.from_pretrained(
        ...     "lllyasviel/control_v11p_sd15_inpaint", torch_dtype=torch.float16
        ... )
85
        >>> pipe = StableDiffusionControlNetInpaintPipeline.from_pretrained(
86
        ...     "runwayml/stable-diffusion-v1-5", controlnet=controlnet, torch_dtype=torch.float16
87
88
        ... )

89
        >>> pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
90
91
92
93
        >>> pipe.enable_model_cpu_offload()

        >>> # generate image
        >>> image = pipe(
94
        ...     "a handsome man with ray-ban sunglasses",
95
        ...     num_inference_steps=20,
96
        ...     generator=generator,
97
        ...     eta=1.0,
98
99
        ...     image=init_image,
        ...     mask_image=mask_image,
100
        ...     control_image=control_image,
101
102
103
104
105
        ... ).images[0]
        ```
"""


106
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
107
108
109
110
def retrieve_latents(
    encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
):
    if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
111
        return encoder_output.latent_dist.sample(generator)
112
113
    elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
        return encoder_output.latent_dist.mode()
114
115
116
117
118
119
    elif hasattr(encoder_output, "latents"):
        return encoder_output.latents
    else:
        raise AttributeError("Could not access latents of provided encoder_output")


120
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint.prepare_mask_and_masked_image
121
def prepare_mask_and_masked_image(image, mask, height, width, return_image=False):
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
    """
    Prepares a pair (image, mask) to be consumed by the Stable Diffusion pipeline. This means that those inputs will be
    converted to ``torch.Tensor`` with shapes ``batch x channels x height x width`` where ``channels`` is ``3`` for the
    ``image`` and ``1`` for the ``mask``.

    The ``image`` will be converted to ``torch.float32`` and normalized to be in ``[-1, 1]``. The ``mask`` will be
    binarized (``mask > 0.5``) and cast to ``torch.float32`` too.

    Args:
        image (Union[np.array, PIL.Image, torch.Tensor]): The image to inpaint.
            It can be a ``PIL.Image``, or a ``height x width x 3`` ``np.array`` or a ``channels x height x width``
            ``torch.Tensor`` or a ``batch x channels x height x width`` ``torch.Tensor``.
        mask (_type_): The mask to apply to the image, i.e. regions to inpaint.
            It can be a ``PIL.Image``, or a ``height x width`` ``np.array`` or a ``1 x height x width``
            ``torch.Tensor`` or a ``batch x 1 x height x width`` ``torch.Tensor``.


    Raises:
        ValueError: ``torch.Tensor`` images should be in the ``[-1, 1]`` range. ValueError: ``torch.Tensor`` mask
        should be in the ``[0, 1]`` range. ValueError: ``mask`` and ``image`` should have the same spatial dimensions.
        TypeError: ``mask`` is a ``torch.Tensor`` but ``image`` is not
            (ot the other way around).

    Returns:
        tuple[torch.Tensor]: The pair (mask, masked_image) as ``torch.Tensor`` with 4
            dimensions: ``batch x channels x height x width``.
    """
149
150
151
152
153
154
    deprecation_message = "The prepare_mask_and_masked_image method is deprecated and will be removed in a future version. Please use VaeImageProcessor.preprocess instead"
    deprecate(
        "prepare_mask_and_masked_image",
        "0.30.0",
        deprecation_message,
    )
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
    if image is None:
        raise ValueError("`image` input cannot be undefined.")

    if mask is None:
        raise ValueError("`mask_image` input cannot be undefined.")

    if isinstance(image, torch.Tensor):
        if not isinstance(mask, torch.Tensor):
            raise TypeError(f"`image` is a torch.Tensor but `mask` (type: {type(mask)} is not")

        # Batch single image
        if image.ndim == 3:
            assert image.shape[0] == 3, "Image outside a batch should be of shape (3, H, W)"
            image = image.unsqueeze(0)

        # Batch and add channel dim for single mask
        if mask.ndim == 2:
            mask = mask.unsqueeze(0).unsqueeze(0)

        # Batch single mask or add channel dim
        if mask.ndim == 3:
            # Single batched mask, no channel dim or single mask not batched but channel dim
            if mask.shape[0] == 1:
                mask = mask.unsqueeze(0)

            # Batched masks no channel dim
            else:
                mask = mask.unsqueeze(1)

        assert image.ndim == 4 and mask.ndim == 4, "Image and Mask must have 4 dimensions"
        assert image.shape[-2:] == mask.shape[-2:], "Image and Mask must have the same spatial dimensions"
        assert image.shape[0] == mask.shape[0], "Image and Mask must have the same batch size"

        # Check image is in [-1, 1]
        if image.min() < -1 or image.max() > 1:
            raise ValueError("Image should be in [-1, 1] range")

        # Check mask is in [0, 1]
        if mask.min() < 0 or mask.max() > 1:
            raise ValueError("Mask should be in [0, 1] range")

        # Binarize mask
        mask[mask < 0.5] = 0
        mask[mask >= 0.5] = 1

        # Image as float32
        image = image.to(dtype=torch.float32)
    elif isinstance(mask, torch.Tensor):
        raise TypeError(f"`mask` is a torch.Tensor but `image` (type: {type(image)} is not")
    else:
        # preprocess image
        if isinstance(image, (PIL.Image.Image, np.ndarray)):
            image = [image]
        if isinstance(image, list) and isinstance(image[0], PIL.Image.Image):
            # resize all images w.r.t passed height an width
            image = [i.resize((width, height), resample=PIL.Image.LANCZOS) for i in image]
            image = [np.array(i.convert("RGB"))[None, :] for i in image]
            image = np.concatenate(image, axis=0)
        elif isinstance(image, list) and isinstance(image[0], np.ndarray):
            image = np.concatenate([i[None, :] for i in image], axis=0)

        image = image.transpose(0, 3, 1, 2)
        image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0

        # preprocess mask
        if isinstance(mask, (PIL.Image.Image, np.ndarray)):
            mask = [mask]

        if isinstance(mask, list) and isinstance(mask[0], PIL.Image.Image):
            mask = [i.resize((width, height), resample=PIL.Image.LANCZOS) for i in mask]
            mask = np.concatenate([np.array(m.convert("L"))[None, None, :] for m in mask], axis=0)
            mask = mask.astype(np.float32) / 255.0
        elif isinstance(mask, list) and isinstance(mask[0], np.ndarray):
            mask = np.concatenate([m[None, None, :] for m in mask], axis=0)

        mask[mask < 0.5] = 0
        mask[mask >= 0.5] = 1
        mask = torch.from_numpy(mask)

    masked_image = image * (mask < 0.5)

236
237
238
239
    # n.b. ensure backwards compatibility as old function does not return image
    if return_image:
        return mask, masked_image, image

240
241
242
    return mask, masked_image


243
class StableDiffusionControlNetInpaintPipeline(
244
    DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, IPAdapterMixin, FromSingleFileMixin
245
):
246
    r"""
Steven Liu's avatar
Steven Liu committed
247
    Pipeline for image inpainting using Stable Diffusion with ControlNet guidance.
248

Steven Liu's avatar
Steven Liu committed
249
250
    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
    implemented for all pipelines (downloading, saving, running on a particular device, etc.).
251

Steven Liu's avatar
Steven Liu committed
252
253
    The pipeline also inherits the following loading methods:
        - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
254
255
256
        - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
        - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
        - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
257
        - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
258

259
260
    <Tip>

Steven Liu's avatar
Steven Liu committed
261
262
263
264
265
266
    This pipeline can be used with checkpoints that have been specifically fine-tuned for inpainting
    ([runwayml/stable-diffusion-inpainting](https://huggingface.co/runwayml/stable-diffusion-inpainting)) as well as
    default text-to-image Stable Diffusion checkpoints
    ([runwayml/stable-diffusion-v1-5](https://huggingface.co/runwayml/stable-diffusion-v1-5)). Default text-to-image
    Stable Diffusion checkpoints might be preferable for ControlNets that have been fine-tuned on those, such as
    [lllyasviel/control_v11p_sd15_inpaint](https://huggingface.co/lllyasviel/control_v11p_sd15_inpaint).
267
268
269

    </Tip>

270
271
    Args:
        vae ([`AutoencoderKL`]):
Steven Liu's avatar
Steven Liu committed
272
273
274
275
276
277
278
            Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
        text_encoder ([`~transformers.CLIPTextModel`]):
            Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
        tokenizer ([`~transformers.CLIPTokenizer`]):
            A `CLIPTokenizer` to tokenize text.
        unet ([`UNet2DConditionModel`]):
            A `UNet2DConditionModel` to denoise the encoded image latents.
279
        controlnet ([`ControlNetModel`] or `List[ControlNetModel]`):
Steven Liu's avatar
Steven Liu committed
280
281
282
            Provides additional conditioning to the `unet` during the denoising process. If you set multiple
            ControlNets as a list, the outputs from each ControlNet are added together to create one combined
            additional conditioning.
283
284
285
286
287
        scheduler ([`SchedulerMixin`]):
            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
        safety_checker ([`StableDiffusionSafetyChecker`]):
            Classification module that estimates whether generated images could be considered offensive or harmful.
Steven Liu's avatar
Steven Liu committed
288
289
290
291
            Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
            about a model's potential harms.
        feature_extractor ([`~transformers.CLIPImageProcessor`]):
            A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
292
    """
293

294
    model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae"
295
    _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
296
    _exclude_from_cpu_offload = ["safety_checker"]
297
    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
298
299
300
301
302
303
304
305
306
307
308

    def __init__(
        self,
        vae: AutoencoderKL,
        text_encoder: CLIPTextModel,
        tokenizer: CLIPTokenizer,
        unet: UNet2DConditionModel,
        controlnet: Union[ControlNetModel, List[ControlNetModel], Tuple[ControlNetModel], MultiControlNetModel],
        scheduler: KarrasDiffusionSchedulers,
        safety_checker: StableDiffusionSafetyChecker,
        feature_extractor: CLIPImageProcessor,
309
        image_encoder: CLIPVisionModelWithProjection = None,
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
        requires_safety_checker: bool = True,
    ):
        super().__init__()

        if safety_checker is None and requires_safety_checker:
            logger.warning(
                f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
                " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
                " results in services or applications open to the public. Both the diffusers team and Hugging Face"
                " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
                " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
                " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
            )

        if safety_checker is not None and feature_extractor is None:
            raise ValueError(
                "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
                " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
            )

        if isinstance(controlnet, (list, tuple)):
            controlnet = MultiControlNetModel(controlnet)

        self.register_modules(
            vae=vae,
            text_encoder=text_encoder,
            tokenizer=tokenizer,
            unet=unet,
            controlnet=controlnet,
            scheduler=scheduler,
            safety_checker=safety_checker,
            feature_extractor=feature_extractor,
342
            image_encoder=image_encoder,
343
344
345
        )
        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
346
347
348
        self.mask_processor = VaeImageProcessor(
            vae_scale_factor=self.vae_scale_factor, do_normalize=False, do_binarize=True, do_convert_grayscale=True
        )
349
350
351
        self.control_image_processor = VaeImageProcessor(
            vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True, do_normalize=False
        )
352
353
354
355
356
        self.register_to_config(requires_safety_checker=requires_safety_checker)

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
    def enable_vae_slicing(self):
        r"""
357
358
        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
359
360
361
362
363
364
        """
        self.vae.enable_slicing()

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
    def disable_vae_slicing(self):
        r"""
365
        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
366
367
368
369
370
371
372
        computing decoding in one step.
        """
        self.vae.disable_slicing()

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
    def enable_vae_tiling(self):
        r"""
373
374
375
        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
        processing larger images.
376
377
378
379
380
381
        """
        self.vae.enable_tiling()

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
    def disable_vae_tiling(self):
        r"""
382
        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
383
384
385
386
387
388
389
390
391
392
393
394
395
396
        computing decoding in one step.
        """
        self.vae.disable_tiling()

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
    def _encode_prompt(
        self,
        prompt,
        device,
        num_images_per_prompt,
        do_classifier_free_guidance,
        negative_prompt=None,
        prompt_embeds: Optional[torch.FloatTensor] = None,
        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
397
        lora_scale: Optional[float] = None,
398
        **kwargs,
399
400
401
402
403
404
405
406
407
408
409
410
411
    ):
        deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
        deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)

        prompt_embeds_tuple = self.encode_prompt(
            prompt=prompt,
            device=device,
            num_images_per_prompt=num_images_per_prompt,
            do_classifier_free_guidance=do_classifier_free_guidance,
            negative_prompt=negative_prompt,
            prompt_embeds=prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds,
            lora_scale=lora_scale,
412
            **kwargs,
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
        )

        # concatenate for backwards comp
        prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])

        return prompt_embeds

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
    def encode_prompt(
        self,
        prompt,
        device,
        num_images_per_prompt,
        do_classifier_free_guidance,
        negative_prompt=None,
        prompt_embeds: Optional[torch.FloatTensor] = None,
        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
        lora_scale: Optional[float] = None,
431
        clip_skip: Optional[int] = None,
432
433
434
435
436
    ):
        r"""
        Encodes the prompt into text encoder hidden states.

        Args:
437
            prompt (`str` or `List[str]`, *optional*):
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
                prompt to be encoded
            device: (`torch.device`):
                torch device
            num_images_per_prompt (`int`):
                number of images that should be generated per prompt
            do_classifier_free_guidance (`bool`):
                whether to use classifier free guidance or not
            negative_prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation. If not defined, one has to pass
                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
                less than `1`).
            prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
                provided, text embeddings will be generated from `prompt` input argument.
            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
                argument.
456
            lora_scale (`float`, *optional*):
457
458
459
460
                A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
            clip_skip (`int`, *optional*):
                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
                the output of the pre-final layer will be used for computing the prompt embeddings.
461
        """
462
463
464
465
466
        # set lora scale so that monkey patched LoRA
        # function of text encoder can correctly access it
        if lora_scale is not None and isinstance(self, LoraLoaderMixin):
            self._lora_scale = lora_scale

467
            # dynamically adjust the LoRA scale
468
            if not USE_PEFT_BACKEND:
469
470
471
                adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
            else:
                scale_lora_layers(self.text_encoder, lora_scale)
472

473
474
475
476
477
478
479
480
        if prompt is not None and isinstance(prompt, str):
            batch_size = 1
        elif prompt is not None and isinstance(prompt, list):
            batch_size = len(prompt)
        else:
            batch_size = prompt_embeds.shape[0]

        if prompt_embeds is None:
co63oc's avatar
co63oc committed
481
            # textual inversion: process multi-vector tokens if necessary
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
            if isinstance(self, TextualInversionLoaderMixin):
                prompt = self.maybe_convert_prompt(prompt, self.tokenizer)

            text_inputs = self.tokenizer(
                prompt,
                padding="max_length",
                max_length=self.tokenizer.model_max_length,
                truncation=True,
                return_tensors="pt",
            )
            text_input_ids = text_inputs.input_ids
            untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids

            if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
                text_input_ids, untruncated_ids
            ):
                removed_text = self.tokenizer.batch_decode(
                    untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
                )
                logger.warning(
                    "The following part of your input was truncated because CLIP can only handle sequences up to"
                    f" {self.tokenizer.model_max_length} tokens: {removed_text}"
                )

            if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
                attention_mask = text_inputs.attention_mask.to(device)
            else:
                attention_mask = None

511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
            if clip_skip is None:
                prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
                prompt_embeds = prompt_embeds[0]
            else:
                prompt_embeds = self.text_encoder(
                    text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
                )
                # Access the `hidden_states` first, that contains a tuple of
                # all the hidden states from the encoder layers. Then index into
                # the tuple to access the hidden states from the desired layer.
                prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
                # We also need to apply the final LayerNorm here to not mess with the
                # representations. The `last_hidden_states` that we typically use for
                # obtaining the final prompt representations passes through the LayerNorm
                # layer.
                prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
527

528
529
530
531
532
533
534
535
        if self.text_encoder is not None:
            prompt_embeds_dtype = self.text_encoder.dtype
        elif self.unet is not None:
            prompt_embeds_dtype = self.unet.dtype
        else:
            prompt_embeds_dtype = prompt_embeds.dtype

        prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562

        bs_embed, seq_len, _ = prompt_embeds.shape
        # duplicate text embeddings for each generation per prompt, using mps friendly method
        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
        prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)

        # get unconditional embeddings for classifier free guidance
        if do_classifier_free_guidance and negative_prompt_embeds is None:
            uncond_tokens: List[str]
            if negative_prompt is None:
                uncond_tokens = [""] * batch_size
            elif prompt is not None and type(prompt) is not type(negative_prompt):
                raise TypeError(
                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
                    f" {type(prompt)}."
                )
            elif isinstance(negative_prompt, str):
                uncond_tokens = [negative_prompt]
            elif batch_size != len(negative_prompt):
                raise ValueError(
                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
                    " the batch size of `prompt`."
                )
            else:
                uncond_tokens = negative_prompt

co63oc's avatar
co63oc committed
563
            # textual inversion: process multi-vector tokens if necessary
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
            if isinstance(self, TextualInversionLoaderMixin):
                uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)

            max_length = prompt_embeds.shape[1]
            uncond_input = self.tokenizer(
                uncond_tokens,
                padding="max_length",
                max_length=max_length,
                truncation=True,
                return_tensors="pt",
            )

            if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
                attention_mask = uncond_input.attention_mask.to(device)
            else:
                attention_mask = None

            negative_prompt_embeds = self.text_encoder(
                uncond_input.input_ids.to(device),
                attention_mask=attention_mask,
            )
            negative_prompt_embeds = negative_prompt_embeds[0]

        if do_classifier_free_guidance:
            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
            seq_len = negative_prompt_embeds.shape[1]

591
            negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
592
593
594
595

            negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
            negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)

596
        if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
597
            # Retrieve the original scale by scaling back the LoRA layers
598
            unscale_lora_layers(self.text_encoder, lora_scale)
599

600
        return prompt_embeds, negative_prompt_embeds
601

602
    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
603
    def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
604
605
606
607
608
609
        dtype = next(self.image_encoder.parameters()).dtype

        if not isinstance(image, torch.Tensor):
            image = self.feature_extractor(image, return_tensors="pt").pixel_values

        image = image.to(device=device, dtype=dtype)
610
611
612
613
614
615
616
617
618
619
620
621
622
623
        if output_hidden_states:
            image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
            image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
            uncond_image_enc_hidden_states = self.image_encoder(
                torch.zeros_like(image), output_hidden_states=True
            ).hidden_states[-2]
            uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
                num_images_per_prompt, dim=0
            )
            return image_enc_hidden_states, uncond_image_enc_hidden_states
        else:
            image_embeds = self.image_encoder(image).image_embeds
            image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
            uncond_image_embeds = torch.zeros_like(image_embeds)
624

625
            return image_embeds, uncond_image_embeds
626

627
    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
628
629
630
631
632
633
    def prepare_ip_adapter_image_embeds(
        self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt
    ):
        if ip_adapter_image_embeds is None:
            if not isinstance(ip_adapter_image, list):
                ip_adapter_image = [ip_adapter_image]
634

635
636
637
638
            if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
                raise ValueError(
                    f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
                )
639

640
641
642
643
644
645
646
647
648
649
650
651
            image_embeds = []
            for single_ip_adapter_image, image_proj_layer in zip(
                ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
            ):
                output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
                single_image_embeds, single_negative_image_embeds = self.encode_image(
                    single_ip_adapter_image, device, 1, output_hidden_state
                )
                single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
                single_negative_image_embeds = torch.stack(
                    [single_negative_image_embeds] * num_images_per_prompt, dim=0
                )
652

653
654
655
                if self.do_classifier_free_guidance:
                    single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
                    single_image_embeds = single_image_embeds.to(device)
656

657
658
659
                image_embeds.append(single_image_embeds)
        else:
            image_embeds = ip_adapter_image_embeds
660
661
        return image_embeds

662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
    def run_safety_checker(self, image, device, dtype):
        if self.safety_checker is None:
            has_nsfw_concept = None
        else:
            if torch.is_tensor(image):
                feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
            else:
                feature_extractor_input = self.image_processor.numpy_to_pil(image)
            safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
            image, has_nsfw_concept = self.safety_checker(
                images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
            )
        return image, has_nsfw_concept

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
    def decode_latents(self, latents):
679
680
681
        deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
        deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)

682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
        latents = 1 / self.vae.config.scaling_factor * latents
        image = self.vae.decode(latents, return_dict=False)[0]
        image = (image / 2 + 0.5).clamp(0, 1)
        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
        image = image.cpu().permute(0, 2, 3, 1).float().numpy()
        return image

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
    def prepare_extra_step_kwargs(self, generator, eta):
        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
        # and should be between [0, 1]

        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
        extra_step_kwargs = {}
        if accepts_eta:
            extra_step_kwargs["eta"] = eta

        # check if the scheduler accepts generator
        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
        if accepts_generator:
            extra_step_kwargs["generator"] = generator
        return extra_step_kwargs

707
708
709
710
711
712
713
    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps
    def get_timesteps(self, num_inference_steps, strength, device):
        # get the original timestep using init_timestep
        init_timestep = min(int(num_inference_steps * strength), num_inference_steps)

        t_start = max(num_inference_steps - init_timestep, 0)
        timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
714
715
        if hasattr(self.scheduler, "set_begin_index"):
            self.scheduler.set_begin_index(t_start * self.scheduler.order)
716
717
718

        return timesteps, num_inference_steps - t_start

719
720
721
722
    def check_inputs(
        self,
        prompt,
        image,
723
        mask_image,
724
725
726
        height,
        width,
        callback_steps,
727
        output_type,
728
729
730
        negative_prompt=None,
        prompt_embeds=None,
        negative_prompt_embeds=None,
731
732
        ip_adapter_image=None,
        ip_adapter_image_embeds=None,
733
        controlnet_conditioning_scale=1.0,
734
735
        control_guidance_start=0.0,
        control_guidance_end=1.0,
736
        callback_on_step_end_tensor_inputs=None,
737
        padding_mask_crop=None,
738
    ):
739
        if height is not None and height % 8 != 0 or width is not None and width % 8 != 0:
740
741
            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")

742
        if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
743
744
745
746
747
            raise ValueError(
                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
                f" {type(callback_steps)}."
            )

748
749
750
751
752
753
754
        if callback_on_step_end_tensor_inputs is not None and not all(
            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
        ):
            raise ValueError(
                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
            )

755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
        if prompt is not None and prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
                " only forward one of the two."
            )
        elif prompt is None and prompt_embeds is None:
            raise ValueError(
                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
            )
        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")

        if negative_prompt is not None and negative_prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
            )

        if prompt_embeds is not None and negative_prompt_embeds is not None:
            if prompt_embeds.shape != negative_prompt_embeds.shape:
                raise ValueError(
                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
                    f" {negative_prompt_embeds.shape}."
                )

781
782
783
784
785
786
787
788
789
790
791
792
793
        if padding_mask_crop is not None:
            if not isinstance(image, PIL.Image.Image):
                raise ValueError(
                    f"The image should be a PIL image when inpainting mask crop, but is of type" f" {type(image)}."
                )
            if not isinstance(mask_image, PIL.Image.Image):
                raise ValueError(
                    f"The mask image should be a PIL image when inpainting mask crop, but is of type"
                    f" {type(mask_image)}."
                )
            if output_type != "pil":
                raise ValueError(f"The output type should be PIL when inpainting mask crop, but is" f" {output_type}.")

794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
        # `prompt` needs more sophisticated handling when there are multiple
        # conditionings.
        if isinstance(self.controlnet, MultiControlNetModel):
            if isinstance(prompt, list):
                logger.warning(
                    f"You have {len(self.controlnet.nets)} ControlNets and you have passed {len(prompt)}"
                    " prompts. The conditionings will be fixed across the prompts."
                )

        # Check `image`
        is_compiled = hasattr(F, "scaled_dot_product_attention") and isinstance(
            self.controlnet, torch._dynamo.eval_frame.OptimizedModule
        )
        if (
            isinstance(self.controlnet, ControlNetModel)
            or is_compiled
            and isinstance(self.controlnet._orig_mod, ControlNetModel)
        ):
            self.check_image(image, prompt, prompt_embeds)
        elif (
            isinstance(self.controlnet, MultiControlNetModel)
            or is_compiled
            and isinstance(self.controlnet._orig_mod, MultiControlNetModel)
        ):
            if not isinstance(image, list):
                raise TypeError("For multiple controlnets: `image` must be type `list`")

            # When `image` is a nested list:
            # (e.g. [[canny_image_1, pose_image_1], [canny_image_2, pose_image_2]])
            elif any(isinstance(i, list) for i in image):
                raise ValueError("A single batch of multiple conditionings are supported at the moment.")
            elif len(image) != len(self.controlnet.nets):
                raise ValueError(
827
                    f"For multiple controlnets: `image` must have the same length as the number of controlnets, but got {len(image)} images and {len(self.controlnet.nets)} ControlNets."
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
                )

            for image_ in image:
                self.check_image(image_, prompt, prompt_embeds)
        else:
            assert False

        # Check `controlnet_conditioning_scale`
        if (
            isinstance(self.controlnet, ControlNetModel)
            or is_compiled
            and isinstance(self.controlnet._orig_mod, ControlNetModel)
        ):
            if not isinstance(controlnet_conditioning_scale, float):
                raise TypeError("For single controlnet: `controlnet_conditioning_scale` must be type `float`.")
        elif (
            isinstance(self.controlnet, MultiControlNetModel)
            or is_compiled
            and isinstance(self.controlnet._orig_mod, MultiControlNetModel)
        ):
            if isinstance(controlnet_conditioning_scale, list):
                if any(isinstance(i, list) for i in controlnet_conditioning_scale):
                    raise ValueError("A single batch of multiple conditionings are supported at the moment.")
            elif isinstance(controlnet_conditioning_scale, list) and len(controlnet_conditioning_scale) != len(
                self.controlnet.nets
            ):
                raise ValueError(
                    "For multiple controlnets: When `controlnet_conditioning_scale` is specified as `list`, it must have"
                    " the same length as the number of controlnets"
                )
        else:
            assert False

861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
        if len(control_guidance_start) != len(control_guidance_end):
            raise ValueError(
                f"`control_guidance_start` has {len(control_guidance_start)} elements, but `control_guidance_end` has {len(control_guidance_end)} elements. Make sure to provide the same number of elements to each list."
            )

        if isinstance(self.controlnet, MultiControlNetModel):
            if len(control_guidance_start) != len(self.controlnet.nets):
                raise ValueError(
                    f"`control_guidance_start`: {control_guidance_start} has {len(control_guidance_start)} elements but there are {len(self.controlnet.nets)} controlnets available. Make sure to provide {len(self.controlnet.nets)}."
                )

        for start, end in zip(control_guidance_start, control_guidance_end):
            if start >= end:
                raise ValueError(
                    f"control guidance start: {start} cannot be larger or equal to control guidance end: {end}."
                )
            if start < 0.0:
                raise ValueError(f"control guidance start: {start} can't be smaller than 0.")
            if end > 1.0:
                raise ValueError(f"control guidance end: {end} can't be larger than 1.0.")

882
883
884
885
886
        if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
            raise ValueError(
                "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
            )

887
    # Copied from diffusers.pipelines.controlnet.pipeline_controlnet.StableDiffusionControlNetPipeline.check_image
888
889
890
    def check_image(self, image, prompt, prompt_embeds):
        image_is_pil = isinstance(image, PIL.Image.Image)
        image_is_tensor = isinstance(image, torch.Tensor)
891
        image_is_np = isinstance(image, np.ndarray)
892
893
        image_is_pil_list = isinstance(image, list) and isinstance(image[0], PIL.Image.Image)
        image_is_tensor_list = isinstance(image, list) and isinstance(image[0], torch.Tensor)
894
        image_is_np_list = isinstance(image, list) and isinstance(image[0], np.ndarray)
895

896
897
898
899
900
901
902
903
        if (
            not image_is_pil
            and not image_is_tensor
            and not image_is_np
            and not image_is_pil_list
            and not image_is_tensor_list
            and not image_is_np_list
        ):
904
            raise TypeError(
905
                f"image must be passed and be one of PIL image, numpy array, torch tensor, list of PIL images, list of numpy arrays or list of torch tensors, but is {type(image)}"
906
907
908
909
            )

        if image_is_pil:
            image_batch_size = 1
910
        else:
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
            image_batch_size = len(image)

        if prompt is not None and isinstance(prompt, str):
            prompt_batch_size = 1
        elif prompt is not None and isinstance(prompt, list):
            prompt_batch_size = len(prompt)
        elif prompt_embeds is not None:
            prompt_batch_size = prompt_embeds.shape[0]

        if image_batch_size != 1 and image_batch_size != prompt_batch_size:
            raise ValueError(
                f"If image batch size is not 1, image batch size must be same as prompt batch size. image batch size: {image_batch_size}, prompt batch size: {prompt_batch_size}"
            )

    def prepare_control_image(
        self,
        image,
        width,
        height,
        batch_size,
        num_images_per_prompt,
        device,
        dtype,
934
935
        crops_coords,
        resize_mode,
936
937
938
        do_classifier_free_guidance=False,
        guess_mode=False,
    ):
939
940
941
        image = self.control_image_processor.preprocess(
            image, height=height, width=width, crops_coords=crops_coords, resize_mode=resize_mode
        ).to(dtype=torch.float32)
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
        image_batch_size = image.shape[0]

        if image_batch_size == 1:
            repeat_by = batch_size
        else:
            # image batch size is the same as prompt batch size
            repeat_by = num_images_per_prompt

        image = image.repeat_interleave(repeat_by, dim=0)

        image = image.to(device=device, dtype=dtype)

        if do_classifier_free_guidance and not guess_mode:
            image = torch.cat([image] * 2)

        return image

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint.StableDiffusionInpaintPipeline.prepare_latents
960
961
962
963
964
965
966
967
968
969
970
971
972
    def prepare_latents(
        self,
        batch_size,
        num_channels_latents,
        height,
        width,
        dtype,
        device,
        generator,
        latents=None,
        image=None,
        timestep=None,
        is_strength_max=True,
973
974
        return_noise=False,
        return_image_latents=False,
975
    ):
976
977
978
979
980
981
982
        shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
        if isinstance(generator, list) and len(generator) != batch_size:
            raise ValueError(
                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
            )

983
984
985
986
987
988
        if (image is None or timestep is None) and not is_strength_max:
            raise ValueError(
                "Since strength < 1. initial latents are to be initialised as a combination of Image + Noise."
                "However, either the image or the noise timestep has not been provided."
            )

989
990
        if return_image_latents or (latents is None and not is_strength_max):
            image = image.to(device=device, dtype=dtype)
991
992
993
994
995

            if image.shape[1] == 4:
                image_latents = image
            else:
                image_latents = self._encode_vae_image(image=image, generator=generator)
996
            image_latents = image_latents.repeat(batch_size // image_latents.shape[0], 1, 1, 1)
997

998
        if latents is None:
999
            noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
1000
            # if strength is 1. then initialise the latents to noise, else initial to image + noise
1001
            latents = noise if is_strength_max else self.scheduler.add_noise(image_latents, noise, timestep)
1002
1003
            # if pure noise then scale the initial latents by the  Scheduler's init sigma
            latents = latents * self.scheduler.init_noise_sigma if is_strength_max else latents
1004
        else:
1005
1006
            noise = latents.to(device)
            latents = noise * self.scheduler.init_noise_sigma
1007

1008
1009
1010
1011
1012
1013
1014
1015
1016
        outputs = (latents,)

        if return_noise:
            outputs += (noise,)

        if return_image_latents:
            outputs += (image_latents,)

        return outputs
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint.StableDiffusionInpaintPipeline.prepare_mask_latents
    def prepare_mask_latents(
        self, mask, masked_image, batch_size, height, width, dtype, device, generator, do_classifier_free_guidance
    ):
        # resize the mask to latents shape as we concatenate the mask to the latents
        # we do that before converting to dtype to avoid breaking in case we're using cpu_offload
        # and half precision
        mask = torch.nn.functional.interpolate(
            mask, size=(height // self.vae_scale_factor, width // self.vae_scale_factor)
        )
        mask = mask.to(device=device, dtype=dtype)

        masked_image = masked_image.to(device=device, dtype=dtype)
1031
1032
1033
1034
1035

        if masked_image.shape[1] == 4:
            masked_image_latents = masked_image
        else:
            masked_image_latents = self._encode_vae_image(masked_image, generator=generator)
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063

        # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method
        if mask.shape[0] < batch_size:
            if not batch_size % mask.shape[0] == 0:
                raise ValueError(
                    "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to"
                    f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number"
                    " of masks that you pass is divisible by the total requested batch size."
                )
            mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1)
        if masked_image_latents.shape[0] < batch_size:
            if not batch_size % masked_image_latents.shape[0] == 0:
                raise ValueError(
                    "The passed images and the required batch size don't match. Images are supposed to be duplicated"
                    f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed."
                    " Make sure the number of images that you pass is divisible by the total requested batch size."
                )
            masked_image_latents = masked_image_latents.repeat(batch_size // masked_image_latents.shape[0], 1, 1, 1)

        mask = torch.cat([mask] * 2) if do_classifier_free_guidance else mask
        masked_image_latents = (
            torch.cat([masked_image_latents] * 2) if do_classifier_free_guidance else masked_image_latents
        )

        # aligning device to prevent device errors when concating it with the latent model input
        masked_image_latents = masked_image_latents.to(device=device, dtype=dtype)
        return mask, masked_image_latents

1064
1065
1066
1067
    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint.StableDiffusionInpaintPipeline._encode_vae_image
    def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator):
        if isinstance(generator, list):
            image_latents = [
1068
                retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i])
1069
1070
1071
1072
                for i in range(image.shape[0])
            ]
            image_latents = torch.cat(image_latents, dim=0)
        else:
1073
            image_latents = retrieve_latents(self.vae.encode(image), generator=generator)
1074
1075
1076
1077
1078

        image_latents = self.vae.config.scaling_factor * image_latents

        return image_latents

1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_freeu
    def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
        r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497.

        The suffixes after the scaling factors represent the stages where they are being applied.

        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values
        that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.

        Args:
            s1 (`float`):
                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
                mitigate "oversmoothing effect" in the enhanced denoising process.
            s2 (`float`):
                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
                mitigate "oversmoothing effect" in the enhanced denoising process.
            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
        """
        if not hasattr(self, "unet"):
            raise ValueError("The pipeline must have `unet` for using FreeU.")
        self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2)

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_freeu
    def disable_freeu(self):
        """Disables the FreeU mechanism if enabled."""
        self.unet.disable_freeu()

1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
    @property
    def guidance_scale(self):
        return self._guidance_scale

    @property
    def clip_skip(self):
        return self._clip_skip

    # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
    # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
    # corresponds to doing no classifier free guidance.
    @property
    def do_classifier_free_guidance(self):
        return self._guidance_scale > 1

    @property
    def cross_attention_kwargs(self):
        return self._cross_attention_kwargs

    @property
    def num_timesteps(self):
        return self._num_timesteps

1130
1131
1132
1133
1134
    @torch.no_grad()
    @replace_example_docstring(EXAMPLE_DOC_STRING)
    def __call__(
        self,
        prompt: Union[str, List[str]] = None,
1135
1136
1137
        image: PipelineImageInput = None,
        mask_image: PipelineImageInput = None,
        control_image: PipelineImageInput = None,
1138
1139
        height: Optional[int] = None,
        width: Optional[int] = None,
1140
        padding_mask_crop: Optional[int] = None,
1141
        strength: float = 1.0,
1142
1143
1144
1145
1146
1147
1148
1149
1150
        num_inference_steps: int = 50,
        guidance_scale: float = 7.5,
        negative_prompt: Optional[Union[str, List[str]]] = None,
        num_images_per_prompt: Optional[int] = 1,
        eta: float = 0.0,
        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
        latents: Optional[torch.FloatTensor] = None,
        prompt_embeds: Optional[torch.FloatTensor] = None,
        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
1151
        ip_adapter_image: Optional[PipelineImageInput] = None,
1152
        ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
1153
1154
1155
1156
1157
        output_type: Optional[str] = "pil",
        return_dict: bool = True,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        controlnet_conditioning_scale: Union[float, List[float]] = 0.5,
        guess_mode: bool = False,
1158
1159
        control_guidance_start: Union[float, List[float]] = 0.0,
        control_guidance_end: Union[float, List[float]] = 1.0,
1160
        clip_skip: Optional[int] = None,
1161
1162
1163
        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
        **kwargs,
1164
1165
    ):
        r"""
Steven Liu's avatar
Steven Liu committed
1166
        The call function to the pipeline for generation.
1167
1168
1169

        Args:
            prompt (`str` or `List[str]`, *optional*):
Steven Liu's avatar
Steven Liu committed
1170
                The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
1171
1172
            image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`,
                    `List[PIL.Image.Image]`, or `List[np.ndarray]`):
Steven Liu's avatar
Steven Liu committed
1173
1174
1175
1176
1177
                `Image`, NumPy array or tensor representing an image batch to be used as the starting point. For both
                NumPy array and PyTorch tensor, the expected value range is between `[0, 1]`. If it's a tensor or a
                list or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a NumPy array or
                a list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)`. It can also accept image
                latents as `image`, but if passing latents directly it is not encoded again.
1178
1179
            mask_image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`,
                    `List[PIL.Image.Image]`, or `List[np.ndarray]`):
Steven Liu's avatar
Steven Liu committed
1180
                `Image`, NumPy array or tensor representing an image batch to mask `image`. White pixels in the mask
1181
                are repainted while black pixels are preserved. If `mask_image` is a PIL image, it is converted to a
Steven Liu's avatar
Steven Liu committed
1182
1183
1184
1185
                single channel (luminance) before use. If it's a NumPy array or PyTorch tensor, it should contain one
                color channel (L) instead of 3, so the expected shape for PyTorch tensor would be `(B, 1, H, W)`, `(B,
                H, W)`, `(1, H, W)`, `(H, W)`. And for NumPy array, it would be for `(B, H, W, 1)`, `(B, H, W)`, `(H,
                W, 1)`, or `(H, W)`.
1186
            control_image (`torch.FloatTensor`, `PIL.Image.Image`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`,
1187
                    `List[List[torch.FloatTensor]]`, or `List[List[PIL.Image.Image]]`):
Steven Liu's avatar
Steven Liu committed
1188
1189
1190
1191
1192
1193
1194
                The ControlNet input condition to provide guidance to the `unet` for generation. If the type is
                specified as `torch.FloatTensor`, it is passed to ControlNet as is. `PIL.Image.Image` can also be
                accepted as an image. The dimensions of the output image defaults to `image`'s dimensions. If height
                and/or width are passed, `image` is resized accordingly. If multiple ControlNets are specified in
                `init`, images must be passed as a list such that each element of the list can be correctly batched for
                input to a single ControlNet.
            height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
1195
                The height in pixels of the generated image.
Steven Liu's avatar
Steven Liu committed
1196
            width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
1197
                The width in pixels of the generated image.
1198
1199
1200
1201
1202
1203
            padding_mask_crop (`int`, *optional*, defaults to `None`):
                The size of margin in the crop to be applied to the image and masking. If `None`, no crop is applied to image and mask_image. If
                `padding_mask_crop` is not `None`, it will first find a rectangular region with the same aspect ration of the image and
                contains all masked area, and then expand that area based on `padding_mask_crop`. The image and mask_image will then be cropped based on
                the expanded area before resizing to the original image size for inpainting. This is useful when the masked area is small while the image is large
                and contain information inreleant for inpainging, such as background.
Steven Liu's avatar
Steven Liu committed
1204
1205
1206
1207
1208
1209
            strength (`float`, *optional*, defaults to 1.0):
                Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a
                starting point and more noise is added the higher the `strength`. The number of denoising steps depends
                on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising
                process runs for the full number of iterations specified in `num_inference_steps`. A value of 1
                essentially ignores `image`.
1210
1211
1212
1213
            num_inference_steps (`int`, *optional*, defaults to 50):
                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
                expense of slower inference.
            guidance_scale (`float`, *optional*, defaults to 7.5):
Steven Liu's avatar
Steven Liu committed
1214
1215
                A higher guidance scale value encourages the model to generate images closely linked to the text
                `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
1216
            negative_prompt (`str` or `List[str]`, *optional*):
Steven Liu's avatar
Steven Liu committed
1217
1218
                The prompt or prompts to guide what to not include in image generation. If not defined, you need to
                pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
1219
1220
1221
            num_images_per_prompt (`int`, *optional*, defaults to 1):
                The number of images to generate per prompt.
            eta (`float`, *optional*, defaults to 0.0):
Steven Liu's avatar
Steven Liu committed
1222
1223
                Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
                to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
1224
            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
Steven Liu's avatar
Steven Liu committed
1225
1226
                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
                generation deterministic.
1227
            latents (`torch.FloatTensor`, *optional*):
Steven Liu's avatar
Steven Liu committed
1228
                Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
1229
                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
Steven Liu's avatar
Steven Liu committed
1230
                tensor is generated by sampling using the supplied random `generator`.
1231
            prompt_embeds (`torch.FloatTensor`, *optional*):
Steven Liu's avatar
Steven Liu committed
1232
1233
                Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
                provided, text embeddings are generated from the `prompt` input argument.
1234
            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
Steven Liu's avatar
Steven Liu committed
1235
1236
                Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
                not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
1237
            ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
1238
1239
1240
            ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
                Pre-generated image embeddings for IP-Adapter. If not
                provided, embeddings are computed from the `ip_adapter_image` input argument.
1241
            output_type (`str`, *optional*, defaults to `"pil"`):
Steven Liu's avatar
Steven Liu committed
1242
                The output format of the generated image. Choose between `PIL.Image` or `np.array`.
1243
1244
1245
1246
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
                plain tuple.
            cross_attention_kwargs (`dict`, *optional*):
Steven Liu's avatar
Steven Liu committed
1247
1248
                A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
                [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
1249
            controlnet_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 0.5):
Steven Liu's avatar
Steven Liu committed
1250
1251
1252
                The outputs of the ControlNet are multiplied by `controlnet_conditioning_scale` before they are added
                to the residual in the original `unet`. If multiple ControlNets are specified in `init`, you can set
                the corresponding scale as a list.
1253
            guess_mode (`bool`, *optional*, defaults to `False`):
Steven Liu's avatar
Steven Liu committed
1254
1255
                The ControlNet encoder tries to recognize the content of the input image even if you remove all
                prompts. A `guidance_scale` value between 3.0 and 5.0 is recommended.
1256
            control_guidance_start (`float` or `List[float]`, *optional*, defaults to 0.0):
Steven Liu's avatar
Steven Liu committed
1257
                The percentage of total steps at which the ControlNet starts applying.
1258
            control_guidance_end (`float` or `List[float]`, *optional*, defaults to 1.0):
Steven Liu's avatar
Steven Liu committed
1259
                The percentage of total steps at which the ControlNet stops applying.
1260
1261
1262
            clip_skip (`int`, *optional*):
                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
                the output of the pre-final layer will be used for computing the prompt embeddings.
1263
1264
1265
1266
1267
1268
1269
1270
1271
            callback_on_step_end (`Callable`, *optional*):
                A function that calls at the end of each denoising steps during the inference. The function is called
                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
                `callback_on_step_end_tensor_inputs`.
            callback_on_step_end_tensor_inputs (`List`, *optional*):
                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
                `._callback_tensor_inputs` attribute of your pipeine class.
1272
1273
1274
1275
1276

        Examples:

        Returns:
            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
Steven Liu's avatar
Steven Liu committed
1277
1278
1279
1280
                If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
                otherwise a `tuple` is returned where the first element is a list with the generated images and the
                second element is a list of `bool`s indicating whether the corresponding generated image contains
                "not-safe-for-work" (nsfw) content.
1281
        """
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298

        callback = kwargs.pop("callback", None)
        callback_steps = kwargs.pop("callback_steps", None)

        if callback is not None:
            deprecate(
                "callback",
                "1.0.0",
                "Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
            )
        if callback_steps is not None:
            deprecate(
                "callback_steps",
                "1.0.0",
                "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
            )

1299
1300
1301
1302
1303
1304
1305
1306
1307
        controlnet = self.controlnet._orig_mod if is_compiled_module(self.controlnet) else self.controlnet

        # align format for control guidance
        if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list):
            control_guidance_start = len(control_guidance_end) * [control_guidance_start]
        elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list):
            control_guidance_end = len(control_guidance_start) * [control_guidance_end]
        elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list):
            mult = len(controlnet.nets) if isinstance(controlnet, MultiControlNetModel) else 1
1308
1309
1310
1311
            control_guidance_start, control_guidance_end = (
                mult * [control_guidance_start],
                mult * [control_guidance_end],
            )
1312

1313
1314
1315
1316
        # 1. Check inputs. Raise error if not correct
        self.check_inputs(
            prompt,
            control_image,
1317
            mask_image,
1318
1319
1320
            height,
            width,
            callback_steps,
1321
            output_type,
1322
1323
1324
            negative_prompt,
            prompt_embeds,
            negative_prompt_embeds,
1325
1326
            ip_adapter_image,
            ip_adapter_image_embeds,
1327
            controlnet_conditioning_scale,
1328
1329
            control_guidance_start,
            control_guidance_end,
1330
            callback_on_step_end_tensor_inputs,
1331
            padding_mask_crop,
1332
1333
        )

1334
1335
1336
1337
        self._guidance_scale = guidance_scale
        self._clip_skip = clip_skip
        self._cross_attention_kwargs = cross_attention_kwargs

1338
1339
1340
1341
1342
1343
1344
1345
        # 2. Define call parameters
        if prompt is not None and isinstance(prompt, str):
            batch_size = 1
        elif prompt is not None and isinstance(prompt, list):
            batch_size = len(prompt)
        else:
            batch_size = prompt_embeds.shape[0]

1346
1347
1348
1349
1350
1351
1352
1353
        if padding_mask_crop is not None:
            height, width = self.image_processor.get_default_height_width(image, height, width)
            crops_coords = self.mask_processor.get_crop_region(mask_image, width, height, pad=padding_mask_crop)
            resize_mode = "fill"
        else:
            crops_coords = None
            resize_mode = "default"

1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
        device = self._execution_device

        if isinstance(controlnet, MultiControlNetModel) and isinstance(controlnet_conditioning_scale, float):
            controlnet_conditioning_scale = [controlnet_conditioning_scale] * len(controlnet.nets)

        global_pool_conditions = (
            controlnet.config.global_pool_conditions
            if isinstance(controlnet, ControlNetModel)
            else controlnet.nets[0].config.global_pool_conditions
        )
        guess_mode = guess_mode or global_pool_conditions

        # 3. Encode input prompt
1367
        text_encoder_lora_scale = (
1368
            self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
1369
        )
1370
        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
1371
1372
1373
            prompt,
            device,
            num_images_per_prompt,
1374
            self.do_classifier_free_guidance,
1375
1376
1377
            negative_prompt,
            prompt_embeds=prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds,
1378
            lora_scale=text_encoder_lora_scale,
1379
            clip_skip=self.clip_skip,
1380
        )
1381
1382
1383
        # For classifier free guidance, we need to do two forward passes.
        # Here we concatenate the unconditional and text embeddings into a single batch
        # to avoid doing two forward passes
1384
        if self.do_classifier_free_guidance:
1385
            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
1386

1387
        if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
1388
            image_embeds = self.prepare_ip_adapter_image_embeds(
1389
                ip_adapter_image, ip_adapter_image_embeds, device, batch_size * num_images_per_prompt
1390
            )
1391

1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
        # 4. Prepare image
        if isinstance(controlnet, ControlNetModel):
            control_image = self.prepare_control_image(
                image=control_image,
                width=width,
                height=height,
                batch_size=batch_size * num_images_per_prompt,
                num_images_per_prompt=num_images_per_prompt,
                device=device,
                dtype=controlnet.dtype,
1402
1403
                crops_coords=crops_coords,
                resize_mode=resize_mode,
1404
                do_classifier_free_guidance=self.do_classifier_free_guidance,
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
                guess_mode=guess_mode,
            )
        elif isinstance(controlnet, MultiControlNetModel):
            control_images = []

            for control_image_ in control_image:
                control_image_ = self.prepare_control_image(
                    image=control_image_,
                    width=width,
                    height=height,
                    batch_size=batch_size * num_images_per_prompt,
                    num_images_per_prompt=num_images_per_prompt,
                    device=device,
                    dtype=controlnet.dtype,
1419
1420
                    crops_coords=crops_coords,
                    resize_mode=resize_mode,
1421
                    do_classifier_free_guidance=self.do_classifier_free_guidance,
1422
1423
1424
1425
1426
1427
1428
1429
1430
                    guess_mode=guess_mode,
                )

                control_images.append(control_image_)

            control_image = control_images
        else:
            assert False

1431
        # 4.1 Preprocess mask and image - resizes image and mask w.r.t height and width
1432
1433
1434
1435
        original_image = image
        init_image = self.image_processor.preprocess(
            image, height=height, width=width, crops_coords=crops_coords, resize_mode=resize_mode
        )
1436
1437
        init_image = init_image.to(dtype=torch.float32)

1438
1439
1440
        mask = self.mask_processor.preprocess(
            mask_image, height=height, width=width, resize_mode=resize_mode, crops_coords=crops_coords
        )
1441
1442
1443

        masked_image = init_image * (mask < 0.5)
        _, _, height, width = init_image.shape
1444

1445
1446
        # 5. Prepare timesteps
        self.scheduler.set_timesteps(num_inference_steps, device=device)
1447
1448
1449
1450
1451
1452
1453
        timesteps, num_inference_steps = self.get_timesteps(
            num_inference_steps=num_inference_steps, strength=strength, device=device
        )
        # at which timestep to set the initial noise (n.b. 50% if strength is 0.5)
        latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
        # create a boolean to check if the strength is set to 1. if so then initialise the latents with pure noise
        is_strength_max = strength == 1.0
1454
        self._num_timesteps = len(timesteps)
1455
1456
1457

        # 6. Prepare latent variables
        num_channels_latents = self.vae.config.latent_channels
1458
1459
1460
        num_channels_unet = self.unet.config.in_channels
        return_image_latents = num_channels_unet == 4
        latents_outputs = self.prepare_latents(
1461
1462
1463
1464
1465
1466
1467
1468
            batch_size * num_images_per_prompt,
            num_channels_latents,
            height,
            width,
            prompt_embeds.dtype,
            device,
            generator,
            latents,
1469
1470
1471
1472
1473
            image=init_image,
            timestep=latent_timestep,
            is_strength_max=is_strength_max,
            return_noise=True,
            return_image_latents=return_image_latents,
1474
1475
        )

1476
1477
1478
1479
1480
        if return_image_latents:
            latents, noise, image_latents = latents_outputs
        else:
            latents, noise = latents_outputs

1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
        # 7. Prepare mask latent variables
        mask, masked_image_latents = self.prepare_mask_latents(
            mask,
            masked_image,
            batch_size * num_images_per_prompt,
            height,
            width,
            prompt_embeds.dtype,
            device,
            generator,
1491
            self.do_classifier_free_guidance,
1492
1493
1494
1495
1496
        )

        # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)

1497
        # 7.1 Add image embeds for IP-Adapter
Aryan's avatar
Aryan committed
1498
1499
1500
1501
1502
        added_cond_kwargs = (
            {"image_embeds": image_embeds}
            if ip_adapter_image is not None or ip_adapter_image_embeds is not None
            else None
        )
1503
1504

        # 7.2 Create tensor stating which controlnets to keep
1505
        controlnet_keep = []
1506
        for i in range(len(timesteps)):
1507
            keeps = [
1508
                1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e)
1509
1510
                for s, e in zip(control_guidance_start, control_guidance_end)
            ]
1511
            controlnet_keep.append(keeps[0] if isinstance(controlnet, ControlNetModel) else keeps)
1512

1513
1514
1515
1516
1517
        # 8. Denoising loop
        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
        with self.progress_bar(total=num_inference_steps) as progress_bar:
            for i, t in enumerate(timesteps):
                # expand the latents if we are doing classifier free guidance
1518
                latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
1519
1520
                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)

1521
                # controlnet(s) inference
1522
                if guess_mode and self.do_classifier_free_guidance:
1523
                    # Infer ControlNet only for the conditional batch.
1524
1525
                    control_model_input = latents
                    control_model_input = self.scheduler.scale_model_input(control_model_input, t)
1526
1527
                    controlnet_prompt_embeds = prompt_embeds.chunk(2)[1]
                else:
1528
                    control_model_input = latent_model_input
1529
1530
                    controlnet_prompt_embeds = prompt_embeds

1531
1532
1533
                if isinstance(controlnet_keep[i], list):
                    cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])]
                else:
1534
1535
1536
1537
                    controlnet_cond_scale = controlnet_conditioning_scale
                    if isinstance(controlnet_cond_scale, list):
                        controlnet_cond_scale = controlnet_cond_scale[0]
                    cond_scale = controlnet_cond_scale * controlnet_keep[i]
1538

1539
                down_block_res_samples, mid_block_res_sample = self.controlnet(
1540
                    control_model_input,
1541
1542
1543
                    t,
                    encoder_hidden_states=controlnet_prompt_embeds,
                    controlnet_cond=control_image,
1544
                    conditioning_scale=cond_scale,
1545
1546
1547
1548
                    guess_mode=guess_mode,
                    return_dict=False,
                )

1549
                if guess_mode and self.do_classifier_free_guidance:
1550
1551
1552
1553
1554
1555
1556
                    # Infered ControlNet only for the conditional batch.
                    # To apply the output of ControlNet to both the unconditional and conditional batches,
                    # add 0 to the unconditional batch to keep it unchanged.
                    down_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples]
                    mid_block_res_sample = torch.cat([torch.zeros_like(mid_block_res_sample), mid_block_res_sample])

                # predict the noise residual
1557
1558
1559
                if num_channels_unet == 9:
                    latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents], dim=1)

1560
1561
1562
1563
                noise_pred = self.unet(
                    latent_model_input,
                    t,
                    encoder_hidden_states=prompt_embeds,
1564
                    cross_attention_kwargs=self.cross_attention_kwargs,
1565
1566
                    down_block_additional_residuals=down_block_res_samples,
                    mid_block_additional_residual=mid_block_res_sample,
1567
                    added_cond_kwargs=added_cond_kwargs,
1568
1569
1570
1571
                    return_dict=False,
                )[0]

                # perform guidance
1572
                if self.do_classifier_free_guidance:
1573
1574
1575
1576
1577
1578
                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
                    noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)

                # compute the previous noisy sample x_t -> x_t-1
                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]

1579
                if num_channels_unet == 4:
1580
                    init_latents_proper = image_latents
1581
                    if self.do_classifier_free_guidance:
1582
1583
1584
                        init_mask, _ = mask.chunk(2)
                    else:
                        init_mask = mask
1585
1586

                    if i < len(timesteps) - 1:
1587
1588
1589
1590
                        noise_timestep = timesteps[i + 1]
                        init_latents_proper = self.scheduler.add_noise(
                            init_latents_proper, noise, torch.tensor([noise_timestep])
                        )
1591
1592
1593

                    latents = (1 - init_mask) * init_latents_proper + init_mask * latents

1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
                if callback_on_step_end is not None:
                    callback_kwargs = {}
                    for k in callback_on_step_end_tensor_inputs:
                        callback_kwargs[k] = locals()[k]
                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)

                    latents = callback_outputs.pop("latents", latents)
                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)

1604
1605
1606
1607
                # call the callback, if provided
                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                    progress_bar.update()
                    if callback is not None and i % callback_steps == 0:
1608
1609
                        step_idx = i // getattr(self.scheduler, "order", 1)
                        callback(step_idx, t, latents)
1610
1611
1612
1613
1614
1615
1616
1617
1618

        # If we do sequential model offloading, let's offload unet and controlnet
        # manually for max memory savings
        if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
            self.unet.to("cpu")
            self.controlnet.to("cpu")
            torch.cuda.empty_cache()

        if not output_type == "latent":
Will Berman's avatar
Will Berman committed
1619
1620
1621
            image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[
                0
            ]
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
            image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
        else:
            image = latents
            has_nsfw_concept = None

        if has_nsfw_concept is None:
            do_denormalize = [True] * image.shape[0]
        else:
            do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]

        image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)

1634
1635
1636
        if padding_mask_crop is not None:
            image = [self.image_processor.apply_overlay(mask_image, original_image, i, crops_coords) for i in image]

1637
1638
        # Offload all models
        self.maybe_free_model_hooks()
1639
1640
1641
1642
1643

        if not return_dict:
            return (image, has_nsfw_concept)

        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)