pipeline_controlnet.py 67.7 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
#
# 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.


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
23
from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
24

Álvaro Somoza's avatar
Álvaro Somoza committed
25
from ...callbacks import MultiPipelineCallbacks, PipelineCallback
26
from ...image_processor import PipelineImageInput, VaeImageProcessor
27
from ...loaders import FromSingleFileMixin, IPAdapterMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin
28
from ...models import AutoencoderKL, ControlNetModel, ImageProjection, MultiControlNetModel, UNet2DConditionModel
29
from ...models.lora import adjust_lora_scale_text_encoder
30
from ...schedulers import KarrasDiffusionSchedulers
31
32
33
from ...utils import (
    USE_PEFT_BACKEND,
    deprecate,
34
    is_torch_xla_available,
35
36
37
38
39
    logging,
    replace_example_docstring,
    scale_lora_layers,
    unscale_lora_layers,
)
40
from ...utils.torch_utils import is_compiled_module, is_torch_version, randn_tensor
41
from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
42
from ..stable_diffusion.pipeline_output import StableDiffusionPipelineOutput
43
44
45
from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker


46
47
48
49
50
51
52
if is_torch_xla_available():
    import torch_xla.core.xla_model as xm

    XLA_AVAILABLE = True
else:
    XLA_AVAILABLE = False

53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
logger = logging.get_logger(__name__)  # pylint: disable=invalid-name


EXAMPLE_DOC_STRING = """
    Examples:
        ```py
        >>> # !pip install opencv-python transformers accelerate
        >>> from diffusers import StableDiffusionControlNetPipeline, ControlNetModel, UniPCMultistepScheduler
        >>> from diffusers.utils import load_image
        >>> import numpy as np
        >>> import torch

        >>> import cv2
        >>> from PIL import Image

        >>> # download an image
        >>> image = load_image(
        ...     "https://hf.co/datasets/huggingface/documentation-images/resolve/main/diffusers/input_image_vermeer.png"
        ... )
        >>> image = np.array(image)

        >>> # get canny image
        >>> image = cv2.Canny(image, 100, 200)
        >>> image = image[:, :, None]
        >>> image = np.concatenate([image, image, image], axis=2)
        >>> canny_image = Image.fromarray(image)

        >>> # load control net and stable diffusion v1-5
        >>> controlnet = ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-canny", torch_dtype=torch.float16)
        >>> pipe = StableDiffusionControlNetPipeline.from_pretrained(
83
        ...     "stable-diffusion-v1-5/stable-diffusion-v1-5", controlnet=controlnet, torch_dtype=torch.float16
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
        ... )

        >>> # speed up diffusion process with faster scheduler and memory optimization
        >>> pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
        >>> # remove following line if xformers is not installed
        >>> pipe.enable_xformers_memory_efficient_attention()

        >>> pipe.enable_model_cpu_offload()

        >>> # generate image
        >>> generator = torch.manual_seed(0)
        >>> image = pipe(
        ...     "futuristic-looking woman", num_inference_steps=20, generator=generator, image=canny_image
        ... ).images[0]
        ```
"""


102
103
104
105
106
107
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
def retrieve_timesteps(
    scheduler,
    num_inference_steps: Optional[int] = None,
    device: Optional[Union[str, torch.device]] = None,
    timesteps: Optional[List[int]] = None,
108
    sigmas: Optional[List[float]] = None,
109
110
    **kwargs,
):
111
    r"""
112
113
114
115
116
117
118
    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.

    Args:
        scheduler (`SchedulerMixin`):
            The scheduler to get timesteps from.
        num_inference_steps (`int`):
119
120
            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
            must be `None`.
121
122
123
        device (`str` or `torch.device`, *optional*):
            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
        timesteps (`List[int]`, *optional*):
124
125
126
127
128
            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
            `num_inference_steps` and `sigmas` must be `None`.
        sigmas (`List[float]`, *optional*):
            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
            `num_inference_steps` and `timesteps` must be `None`.
129
130
131
132
133

    Returns:
        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
        second element is the number of inference steps.
    """
134
135
    if timesteps is not None and sigmas is not None:
        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
136
137
138
139
140
141
142
143
144
145
    if timesteps is not None:
        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
        if not accepts_timesteps:
            raise ValueError(
                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
                f" timestep schedules. Please check whether you are using the correct scheduler."
            )
        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
        timesteps = scheduler.timesteps
        num_inference_steps = len(timesteps)
146
147
148
149
150
151
152
153
154
155
    elif sigmas is not None:
        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
        if not accept_sigmas:
            raise ValueError(
                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
                f" sigmas schedules. Please check whether you are using the correct scheduler."
            )
        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
        timesteps = scheduler.timesteps
        num_inference_steps = len(timesteps)
156
157
158
159
160
161
    else:
        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
        timesteps = scheduler.timesteps
    return timesteps, num_inference_steps


162
class StableDiffusionControlNetPipeline(
163
164
165
    DiffusionPipeline,
    StableDiffusionMixin,
    TextualInversionLoaderMixin,
166
    StableDiffusionLoraLoaderMixin,
167
168
    IPAdapterMixin,
    FromSingleFileMixin,
169
):
170
171
172
    r"""
    Pipeline for text-to-image generation using Stable Diffusion with ControlNet guidance.

Steven Liu's avatar
Steven Liu committed
173
174
    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.).
175

Steven Liu's avatar
Steven Liu committed
176
177
    The pipeline also inherits the following loading methods:
        - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
178
179
        - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights
        - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights
180
        - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
181
        - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
182
183
184

    Args:
        vae ([`AutoencoderKL`]):
Steven Liu's avatar
Steven Liu committed
185
186
187
188
189
190
191
            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.
192
        controlnet ([`ControlNetModel`] or `List[ControlNetModel]`):
Steven Liu's avatar
Steven Liu committed
193
194
195
            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.
196
197
198
199
200
        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.
Aryan's avatar
Aryan committed
201
202
            Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for
            more details about a model's potential harms.
Steven Liu's avatar
Steven Liu committed
203
204
        feature_extractor ([`~transformers.CLIPImageProcessor`]):
            A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
205
    """
206

207
    model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae"
208
    _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
209
    _exclude_from_cpu_offload = ["safety_checker"]
210
    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds", "image"]
211
212
213
214
215
216
217
218
219
220
221

    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,
222
        image_encoder: CLIPVisionModelWithProjection = None,
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
        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,
255
            image_encoder=image_encoder,
256
        )
hlky's avatar
hlky committed
257
        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8
258
259
260
261
        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True)
        self.control_image_processor = VaeImageProcessor(
            vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True, do_normalize=False
        )
262
263
264
265
266
267
268
269
270
271
        self.register_to_config(requires_safety_checker=requires_safety_checker)

    # 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,
272
273
        prompt_embeds: Optional[torch.Tensor] = None,
        negative_prompt_embeds: Optional[torch.Tensor] = None,
274
        lora_scale: Optional[float] = None,
275
        **kwargs,
276
277
278
279
280
281
282
283
284
285
286
287
288
    ):
        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,
289
            **kwargs,
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
        )

        # 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,
305
306
        prompt_embeds: Optional[torch.Tensor] = None,
        negative_prompt_embeds: Optional[torch.Tensor] = None,
307
        lora_scale: Optional[float] = None,
308
        clip_skip: Optional[int] = None,
309
310
311
312
313
    ):
        r"""
        Encodes the prompt into text encoder hidden states.

        Args:
314
            prompt (`str` or `List[str]`, *optional*):
315
316
317
318
319
320
321
322
323
324
325
                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`).
326
            prompt_embeds (`torch.Tensor`, *optional*):
327
328
                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.
329
            negative_prompt_embeds (`torch.Tensor`, *optional*):
330
331
332
                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.
333
            lora_scale (`float`, *optional*):
334
335
336
337
                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.
338
        """
339
340
        # set lora scale so that monkey patched LoRA
        # function of text encoder can correctly access it
341
        if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin):
342
343
            self._lora_scale = lora_scale

344
            # dynamically adjust the LoRA scale
345
            if not USE_PEFT_BACKEND:
346
347
348
                adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
            else:
                scale_lora_layers(self.text_encoder, lora_scale)
349

350
351
352
353
354
355
356
357
        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
358
            # textual inversion: process multi-vector tokens if necessary
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
            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

388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
            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)
404

405
406
407
408
409
410
411
412
        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)
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439

        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
440
            # textual inversion: process multi-vector tokens if necessary
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
            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]

468
            negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
469
470
471
472

            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)

473
        if self.text_encoder is not None:
474
            if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND:
475
476
                # Retrieve the original scale by scaling back the LoRA layers
                unscale_lora_layers(self.text_encoder, lora_scale)
477

478
        return prompt_embeds, negative_prompt_embeds
479

480
    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
481
    def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
482
483
484
485
486
487
        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)
488
489
490
491
492
493
494
495
496
497
498
499
500
501
        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)
502

503
            return image_embeds, uncond_image_embeds
504

505
    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
506
    def prepare_ip_adapter_image_embeds(
507
        self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
508
    ):
YiYi Xu's avatar
YiYi Xu committed
509
510
511
        image_embeds = []
        if do_classifier_free_guidance:
            negative_image_embeds = []
512
513
514
        if ip_adapter_image_embeds is None:
            if not isinstance(ip_adapter_image, list):
                ip_adapter_image = [ip_adapter_image]
515

516
517
518
519
            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."
                )
520

521
522
523
524
525
526
527
            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
                )
528

YiYi Xu's avatar
YiYi Xu committed
529
                image_embeds.append(single_image_embeds[None, :])
530
                if do_classifier_free_guidance:
YiYi Xu's avatar
YiYi Xu committed
531
                    negative_image_embeds.append(single_negative_image_embeds[None, :])
532
        else:
533
534
535
            for single_image_embeds in ip_adapter_image_embeds:
                if do_classifier_free_guidance:
                    single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
YiYi Xu's avatar
YiYi Xu committed
536
                    negative_image_embeds.append(single_negative_image_embeds)
537
538
                image_embeds.append(single_image_embeds)

YiYi Xu's avatar
YiYi Xu committed
539
540
541
542
543
544
545
546
547
548
549
        ip_adapter_image_embeds = []
        for i, single_image_embeds in enumerate(image_embeds):
            single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0)
            if do_classifier_free_guidance:
                single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0)
                single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0)

            single_image_embeds = single_image_embeds.to(device=device)
            ip_adapter_image_embeds.append(single_image_embeds)

        return ip_adapter_image_embeds
550

551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
    # 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):
568
569
570
        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)

571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
        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

    def check_inputs(
        self,
        prompt,
        image,
        callback_steps,
        negative_prompt=None,
        prompt_embeds=None,
        negative_prompt_embeds=None,
604
605
        ip_adapter_image=None,
        ip_adapter_image_embeds=None,
606
        controlnet_conditioning_scale=1.0,
607
608
        control_guidance_start=0.0,
        control_guidance_end=1.0,
609
        callback_on_step_end_tensor_inputs=None,
610
    ):
611
        if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
612
613
614
615
616
            raise ValueError(
                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
                f" {type(callback_steps)}."
            )

617
618
619
620
621
622
623
        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]}"
            )

624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
        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}."
                )

        # 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):
671
672
673
674
675
676
677
                transposed_image = [list(t) for t in zip(*image)]
                if len(transposed_image) != len(self.controlnet.nets):
                    raise ValueError(
                        f"For multiple controlnets: if you pass`image` as a list of list, each sublist must have the same length as the number of controlnets, but the sublists in `image` got {len(transposed_image)} images and {len(self.controlnet.nets)} ControlNets."
                    )
                for image_ in transposed_image:
                    self.check_image(image_, prompt, prompt_embeds)
678
679
            elif len(image) != len(self.controlnet.nets):
                raise ValueError(
680
                    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."
681
                )
682
683
684
            else:
                for image_ in image:
                    self.check_image(image_, prompt, prompt_embeds)
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
        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):
703
704
705
706
                    raise ValueError(
                        "A single batch of varying conditioning scale settings (e.g. [[1.0, 0.5], [0.2, 0.8]]) is not supported at the moment. "
                        "The conditioning scale must be fixed across the batch."
                    )
707
708
709
710
711
712
713
714
715
716
            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

Neil Wang's avatar
Neil Wang committed
717
718
        if not isinstance(control_guidance_start, (tuple, list)):
            control_guidance_start = [control_guidance_start]
Patrick von Platen's avatar
Patrick von Platen committed
719

Neil Wang's avatar
Neil Wang committed
720
721
        if not isinstance(control_guidance_end, (tuple, list)):
            control_guidance_end = [control_guidance_end]
Patrick von Platen's avatar
Patrick von Platen committed
722

723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
        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.")

744
745
746
747
748
        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."
            )

749
750
751
752
753
        if ip_adapter_image_embeds is not None:
            if not isinstance(ip_adapter_image_embeds, list):
                raise ValueError(
                    f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
                )
754
            elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
755
                raise ValueError(
756
                    f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
757
758
                )

759
760
761
    def check_image(self, image, prompt, prompt_embeds):
        image_is_pil = isinstance(image, PIL.Image.Image)
        image_is_tensor = isinstance(image, torch.Tensor)
762
        image_is_np = isinstance(image, np.ndarray)
763
764
        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)
765
        image_is_np_list = isinstance(image, list) and isinstance(image[0], np.ndarray)
766

767
768
769
770
771
772
773
774
        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
        ):
775
            raise TypeError(
776
                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)}"
777
778
779
780
            )

        if image_is_pil:
            image_batch_size = 1
781
        else:
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
            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_image(
        self,
        image,
        width,
        height,
        batch_size,
        num_images_per_prompt,
        device,
        dtype,
        do_classifier_free_guidance=False,
        guess_mode=False,
    ):
808
        image = self.control_image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32)
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
        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.StableDiffusionPipeline.prepare_latents
    def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
828
829
830
831
832
833
        shape = (
            batch_size,
            num_channels_latents,
            int(height) // self.vae_scale_factor,
            int(width) // self.vae_scale_factor,
        )
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
        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."
            )

        if latents is None:
            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
        else:
            latents = latents.to(device)

        # scale the initial noise by the standard deviation required by the scheduler
        latents = latents * self.scheduler.init_noise_sigma
        return latents

849
    # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
850
851
    def get_guidance_scale_embedding(
        self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32
852
    ) -> torch.Tensor:
853
854
855
856
        """
        See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298

        Args:
857
858
            w (`torch.Tensor`):
                Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings.
859
            embedding_dim (`int`, *optional*, defaults to 512):
860
861
862
                Dimension of the embeddings to generate.
            dtype (`torch.dtype`, *optional*, defaults to `torch.float32`):
                Data type of the generated embeddings.
863
864

        Returns:
865
            `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`.
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
        """
        assert len(w.shape) == 1
        w = w * 1000.0

        half_dim = embedding_dim // 2
        emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
        emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
        emb = w.to(dtype)[:, None] * emb[None, :]
        emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
        if embedding_dim % 2 == 1:  # zero pad
            emb = torch.nn.functional.pad(emb, (0, 1))
        assert emb.shape == (w.shape[0], embedding_dim)
        return emb

    @property
    def guidance_scale(self):
        return self._guidance_scale

884
885
886
887
    @property
    def clip_skip(self):
        return self._clip_skip

888
889
890
891
892
893
894
    # 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 and self.unet.config.time_cond_proj_dim is None

895
896
897
898
899
900
901
902
    @property
    def cross_attention_kwargs(self):
        return self._cross_attention_kwargs

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

903
904
905
906
    @property
    def interrupt(self):
        return self._interrupt

907
908
909
910
911
    @torch.no_grad()
    @replace_example_docstring(EXAMPLE_DOC_STRING)
    def __call__(
        self,
        prompt: Union[str, List[str]] = None,
912
        image: PipelineImageInput = None,
913
914
915
        height: Optional[int] = None,
        width: Optional[int] = None,
        num_inference_steps: int = 50,
916
        timesteps: List[int] = None,
917
        sigmas: List[float] = None,
918
919
920
921
922
        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,
923
924
925
        latents: Optional[torch.Tensor] = None,
        prompt_embeds: Optional[torch.Tensor] = None,
        negative_prompt_embeds: Optional[torch.Tensor] = None,
926
        ip_adapter_image: Optional[PipelineImageInput] = None,
927
        ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None,
928
929
930
931
932
        output_type: Optional[str] = "pil",
        return_dict: bool = True,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        controlnet_conditioning_scale: Union[float, List[float]] = 1.0,
        guess_mode: bool = False,
933
934
        control_guidance_start: Union[float, List[float]] = 0.0,
        control_guidance_end: Union[float, List[float]] = 1.0,
935
        clip_skip: Optional[int] = None,
Álvaro Somoza's avatar
Álvaro Somoza committed
936
937
938
        callback_on_step_end: Optional[
            Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
        ] = None,
939
940
        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
        **kwargs,
941
942
    ):
        r"""
Steven Liu's avatar
Steven Liu committed
943
        The call function to the pipeline for generation.
944
945
946

        Args:
            prompt (`str` or `List[str]`, *optional*):
Steven Liu's avatar
Steven Liu committed
947
                The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
948
949
            image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, `List[np.ndarray]`,:
                    `List[List[torch.Tensor]]`, `List[List[np.ndarray]]` or `List[List[PIL.Image.Image]]`):
Steven Liu's avatar
Steven Liu committed
950
                The ControlNet input condition to provide guidance to the `unet` for generation. If the type is
951
952
953
954
955
                specified as `torch.Tensor`, 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. When `prompt` is a list, and if a list of images is passed for a single
956
957
                ControlNet, each will be paired with each prompt in the `prompt` list. This also applies to multiple
                ControlNets, where a list of image lists can be passed to batch for each prompt and each ControlNet.
Steven Liu's avatar
Steven Liu committed
958
            height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
959
                The height in pixels of the generated image.
Steven Liu's avatar
Steven Liu committed
960
            width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
961
962
963
964
                The width in pixels of the generated image.
            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.
965
966
967
968
            timesteps (`List[int]`, *optional*):
                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
                passed will be used. Must be in descending order.
969
970
971
972
            sigmas (`List[float]`, *optional*):
                Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
                their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
                will be used.
973
            guidance_scale (`float`, *optional*, defaults to 7.5):
Steven Liu's avatar
Steven Liu committed
974
975
                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`.
976
            negative_prompt (`str` or `List[str]`, *optional*):
Steven Liu's avatar
Steven Liu committed
977
978
                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`).
979
980
981
            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
982
983
                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.
984
            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
Steven Liu's avatar
Steven Liu committed
985
986
                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
                generation deterministic.
987
            latents (`torch.Tensor`, *optional*):
Steven Liu's avatar
Steven Liu committed
988
                Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
989
                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
Steven Liu's avatar
Steven Liu committed
990
                tensor is generated by sampling using the supplied random `generator`.
991
            prompt_embeds (`torch.Tensor`, *optional*):
Steven Liu's avatar
Steven Liu committed
992
993
                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.
994
            negative_prompt_embeds (`torch.Tensor`, *optional*):
Steven Liu's avatar
Steven Liu committed
995
996
                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.
997
            ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
998
            ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*):
999
1000
1001
1002
                Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of
                IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should
                contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not
                provided, embeddings are computed from the `ip_adapter_image` input argument.
1003
            output_type (`str`, *optional*, defaults to `"pil"`):
Steven Liu's avatar
Steven Liu committed
1004
                The output format of the generated image. Choose between `PIL.Image` or `np.array`.
1005
1006
1007
1008
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
                plain tuple.
            callback (`Callable`, *optional*):
Steven Liu's avatar
Steven Liu committed
1009
                A function that calls every `callback_steps` steps during inference. The function is called with the
1010
                following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`.
1011
            callback_steps (`int`, *optional*, defaults to 1):
Steven Liu's avatar
Steven Liu committed
1012
1013
                The frequency at which the `callback` function is called. If not specified, the callback is called at
                every step.
1014
            cross_attention_kwargs (`dict`, *optional*):
Steven Liu's avatar
Steven Liu committed
1015
1016
                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).
1017
            controlnet_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 1.0):
Steven Liu's avatar
Steven Liu committed
1018
1019
1020
                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.
1021
            guess_mode (`bool`, *optional*, defaults to `False`):
Steven Liu's avatar
Steven Liu committed
1022
1023
                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.
1024
            control_guidance_start (`float` or `List[float]`, *optional*, defaults to 0.0):
Steven Liu's avatar
Steven Liu committed
1025
                The percentage of total steps at which the ControlNet starts applying.
1026
            control_guidance_end (`float` or `List[float]`, *optional*, defaults to 1.0):
Steven Liu's avatar
Steven Liu committed
1027
                The percentage of total steps at which the ControlNet stops applying.
1028
1029
1030
            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.
Álvaro Somoza's avatar
Álvaro Somoza committed
1031
1032
1033
1034
1035
            callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*):
                A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of
                each denoising step during the inference. 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`.
1036
1037
1038
            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
M. Tolga Cangöz's avatar
M. Tolga Cangöz committed
1039
                `._callback_tensor_inputs` attribute of your pipeline class.
1040
1041
1042
1043
1044

        Examples:

        Returns:
            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
Steven Liu's avatar
Steven Liu committed
1045
1046
1047
1048
                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.
1049
        """
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066

        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`",
            )

Álvaro Somoza's avatar
Álvaro Somoza committed
1067
1068
1069
        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs

1070
1071
1072
1073
1074
1075
1076
1077
1078
        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
1079
1080
1081
1082
            control_guidance_start, control_guidance_end = (
                mult * [control_guidance_start],
                mult * [control_guidance_end],
            )
1083
1084
1085
1086
1087
1088
1089
1090
1091

        # 1. Check inputs. Raise error if not correct
        self.check_inputs(
            prompt,
            image,
            callback_steps,
            negative_prompt,
            prompt_embeds,
            negative_prompt_embeds,
1092
1093
            ip_adapter_image,
            ip_adapter_image_embeds,
1094
            controlnet_conditioning_scale,
1095
1096
            control_guidance_start,
            control_guidance_end,
1097
            callback_on_step_end_tensor_inputs,
1098
1099
        )

1100
        self._guidance_scale = guidance_scale
1101
1102
        self._clip_skip = clip_skip
        self._cross_attention_kwargs = cross_attention_kwargs
1103
        self._interrupt = False
1104

1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
        # 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]

        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
1126
        text_encoder_lora_scale = (
1127
            self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
1128
        )
1129
        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
1130
1131
1132
            prompt,
            device,
            num_images_per_prompt,
1133
            self.do_classifier_free_guidance,
1134
1135
1136
            negative_prompt,
            prompt_embeds=prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds,
1137
            lora_scale=text_encoder_lora_scale,
1138
            clip_skip=self.clip_skip,
1139
        )
1140
1141
1142
        # 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
1143
        if self.do_classifier_free_guidance:
1144
            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
1145

1146
        if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
1147
            image_embeds = self.prepare_ip_adapter_image_embeds(
1148
1149
1150
1151
1152
                ip_adapter_image,
                ip_adapter_image_embeds,
                device,
                batch_size * num_images_per_prompt,
                self.do_classifier_free_guidance,
1153
            )
1154

1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
        # 4. Prepare image
        if isinstance(controlnet, ControlNetModel):
            image = self.prepare_image(
                image=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,
1165
                do_classifier_free_guidance=self.do_classifier_free_guidance,
1166
1167
                guess_mode=guess_mode,
            )
1168
            height, width = image.shape[-2:]
1169
1170
1171
        elif isinstance(controlnet, MultiControlNetModel):
            images = []

1172
1173
1174
1175
1176
            # Nested lists as ControlNet condition
            if isinstance(image[0], list):
                # Transpose the nested image list
                image = [list(t) for t in zip(*image)]

1177
1178
1179
1180
1181
1182
1183
1184
1185
            for image_ in image:
                image_ = self.prepare_image(
                    image=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,
1186
                    do_classifier_free_guidance=self.do_classifier_free_guidance,
1187
1188
1189
1190
1191
1192
                    guess_mode=guess_mode,
                )

                images.append(image_)

            image = images
1193
            height, width = image[0].shape[-2:]
1194
1195
1196
1197
        else:
            assert False

        # 5. Prepare timesteps
1198
1199
1200
        timesteps, num_inference_steps = retrieve_timesteps(
            self.scheduler, num_inference_steps, device, timesteps, sigmas
        )
1201
        self._num_timesteps = len(timesteps)
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215

        # 6. Prepare latent variables
        num_channels_latents = self.unet.config.in_channels
        latents = self.prepare_latents(
            batch_size * num_images_per_prompt,
            num_channels_latents,
            height,
            width,
            prompt_embeds.dtype,
            device,
            generator,
            latents,
        )

1216
1217
1218
1219
1220
1221
1222
1223
        # 6.5 Optionally get Guidance Scale Embedding
        timestep_cond = None
        if self.unet.config.time_cond_proj_dim is not None:
            guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
            timestep_cond = self.get_guidance_scale_embedding(
                guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
            ).to(device=device, dtype=latents.dtype)

1224
1225
1226
        # 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)

1227
        # 7.1 Add image embeds for IP-Adapter
Aryan's avatar
Aryan committed
1228
1229
1230
1231
1232
        added_cond_kwargs = (
            {"image_embeds": image_embeds}
            if ip_adapter_image is not None or ip_adapter_image_embeds is not None
            else None
        )
1233
1234

        # 7.2 Create tensor stating which controlnets to keep
1235
        controlnet_keep = []
1236
        for i in range(len(timesteps)):
1237
            keeps = [
1238
                1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e)
1239
1240
                for s, e in zip(control_guidance_start, control_guidance_end)
            ]
1241
            controlnet_keep.append(keeps[0] if isinstance(controlnet, ControlNetModel) else keeps)
1242

1243
1244
        # 8. Denoising loop
        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
1245
1246
1247
        is_unet_compiled = is_compiled_module(self.unet)
        is_controlnet_compiled = is_compiled_module(self.controlnet)
        is_torch_higher_equal_2_1 = is_torch_version(">=", "2.1")
1248
1249
        with self.progress_bar(total=num_inference_steps) as progress_bar:
            for i, t in enumerate(timesteps):
1250
1251
1252
                if self.interrupt:
                    continue

1253
1254
1255
1256
                # Relevant thread:
                # https://dev-discuss.pytorch.org/t/cudagraphs-in-pytorch-2-0/1428
                if (is_unet_compiled and is_controlnet_compiled) and is_torch_higher_equal_2_1:
                    torch._inductor.cudagraph_mark_step_begin()
1257
                # expand the latents if we are doing classifier free guidance
1258
                latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
1259
1260
1261
                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)

                # controlnet(s) inference
1262
                if guess_mode and self.do_classifier_free_guidance:
1263
                    # Infer ControlNet only for the conditional batch.
1264
1265
                    control_model_input = latents
                    control_model_input = self.scheduler.scale_model_input(control_model_input, t)
1266
1267
                    controlnet_prompt_embeds = prompt_embeds.chunk(2)[1]
                else:
1268
                    control_model_input = latent_model_input
1269
1270
                    controlnet_prompt_embeds = prompt_embeds

1271
1272
1273
                if isinstance(controlnet_keep[i], list):
                    cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])]
                else:
1274
1275
1276
1277
                    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]
1278

1279
                down_block_res_samples, mid_block_res_sample = self.controlnet(
1280
                    control_model_input,
1281
1282
1283
                    t,
                    encoder_hidden_states=controlnet_prompt_embeds,
                    controlnet_cond=image,
1284
                    conditioning_scale=cond_scale,
1285
1286
1287
1288
                    guess_mode=guess_mode,
                    return_dict=False,
                )

1289
                if guess_mode and self.do_classifier_free_guidance:
1290
                    # Inferred ControlNet only for the conditional batch.
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
                    # 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
                noise_pred = self.unet(
                    latent_model_input,
                    t,
                    encoder_hidden_states=prompt_embeds,
1301
                    timestep_cond=timestep_cond,
1302
                    cross_attention_kwargs=self.cross_attention_kwargs,
1303
1304
                    down_block_additional_residuals=down_block_res_samples,
                    mid_block_additional_residual=mid_block_res_sample,
1305
                    added_cond_kwargs=added_cond_kwargs,
1306
1307
1308
1309
                    return_dict=False,
                )[0]

                # perform guidance
1310
                if self.do_classifier_free_guidance:
1311
                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
1312
                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
1313
1314
1315
1316

                # 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]

1317
1318
1319
1320
1321
1322
1323
1324
1325
                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)
1326
                    image = callback_outputs.pop("image", image)
1327

1328
1329
1330
1331
                # 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:
1332
1333
                        step_idx = i // getattr(self.scheduler, "order", 1)
                        callback(step_idx, t, latents)
1334

1335
1336
                if XLA_AVAILABLE:
                    xm.mark_step()
1337
1338
1339
1340
1341
1342
1343
1344
        # 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
1345
1346
1347
            image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[
                0
            ]
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
            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)

1360
1361
        # Offload all models
        self.maybe_free_model_hooks()
1362
1363
1364
1365
1366

        if not return_dict:
            return (image, has_nsfw_concept)

        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)