pipeline_flux_controlnet.py 56 KB
Newer Older
王奇勋's avatar
王奇勋 committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
# Copyright 2024 Black Forest Labs, The HuggingFace Team and The InstantX Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import inspect
16
from typing import Any, Callable, Dict, List, Optional, Tuple, Union
王奇勋's avatar
王奇勋 committed
17
18
19
20

import numpy as np
import torch
from transformers import (
21
    CLIPImageProcessor,
王奇勋's avatar
王奇勋 committed
22
23
    CLIPTextModel,
    CLIPTokenizer,
24
    CLIPVisionModelWithProjection,
王奇勋's avatar
王奇勋 committed
25
26
27
28
29
    T5EncoderModel,
    T5TokenizerFast,
)

from ...image_processor import PipelineImageInput, VaeImageProcessor
30
from ...loaders import FluxIPAdapterMixin, FluxLoraLoaderMixin, FromSingleFileMixin, TextualInversionLoaderMixin
王奇勋's avatar
王奇勋 committed
31
from ...models.autoencoders import AutoencoderKL
32
from ...models.controlnets.controlnet_flux import FluxControlNetModel, FluxMultiControlNetModel
王奇勋's avatar
王奇勋 committed
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
from ...models.transformers import FluxTransformer2DModel
from ...schedulers import FlowMatchEulerDiscreteScheduler
from ...utils import (
    USE_PEFT_BACKEND,
    is_torch_xla_available,
    logging,
    replace_example_docstring,
    scale_lora_layers,
    unscale_lora_layers,
)
from ...utils.torch_utils import randn_tensor
from ..pipeline_utils import DiffusionPipeline
from .pipeline_output import FluxPipelineOutput


if is_torch_xla_available():
    import torch_xla.core.xla_model as xm

    XLA_AVAILABLE = True
else:
    XLA_AVAILABLE = False


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

EXAMPLE_DOC_STRING = """
    Examples:
        ```py
        >>> import torch
        >>> from diffusers.utils import load_image
        >>> from diffusers import FluxControlNetPipeline
        >>> from diffusers import FluxControlNetModel

66
        >>> controlnet_model = "InstantX/FLUX.1-dev-controlnet-canny"
王奇勋's avatar
王奇勋 committed
67
68
69
70
71
72
73
74
75
76
        >>> controlnet = FluxControlNetModel.from_pretrained(controlnet_model, torch_dtype=torch.bfloat16)
        >>> pipe = FluxControlNetPipeline.from_pretrained(
        ...     base_model, controlnet=controlnet, torch_dtype=torch.bfloat16
        ... )
        >>> pipe.to("cuda")
        >>> control_image = load_image("https://huggingface.co/InstantX/SD3-Controlnet-Canny/resolve/main/canny.jpg")
        >>> prompt = "A girl in city, 25 years old, cool, futuristic"
        >>> image = pipe(
        ...     prompt,
        ...     control_image=control_image,
77
78
79
        ...     control_guidance_start=0.2,
        ...     control_guidance_end=0.8,
        ...     controlnet_conditioning_scale=1.0,
王奇勋's avatar
王奇勋 committed
80
81
82
83
84
85
86
87
88
89
90
91
92
93
        ...     num_inference_steps=28,
        ...     guidance_scale=3.5,
        ... ).images[0]
        >>> image.save("flux.png")
        ```
"""


# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift
def calculate_shift(
    image_seq_len,
    base_seq_len: int = 256,
    max_seq_len: int = 4096,
    base_shift: float = 0.5,
94
    max_shift: float = 1.15,
王奇勋's avatar
王奇勋 committed
95
96
97
98
99
100
101
):
    m = (max_shift - base_shift) / (max_seq_len - base_seq_len)
    b = base_shift - m * base_seq_len
    mu = image_seq_len * m + b
    return mu


102
103
104
105
106
107
108
109
110
111
112
113
114
115
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
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":
        return encoder_output.latent_dist.sample(generator)
    elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
        return encoder_output.latent_dist.mode()
    elif hasattr(encoder_output, "latents"):
        return encoder_output.latents
    else:
        raise AttributeError("Could not access latents of provided encoder_output")


王奇勋's avatar
王奇勋 committed
116
117
118
119
120
121
122
123
124
# 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,
    sigmas: Optional[List[float]] = None,
    **kwargs,
):
125
    r"""
王奇勋's avatar
王奇勋 committed
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
    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`):
            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
            must be `None`.
        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*):
            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`.

    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.
    """
    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")
    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)
    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)
    else:
        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
        timesteps = scheduler.timesteps
    return timesteps, num_inference_steps


176
class FluxControlNetPipeline(DiffusionPipeline, FluxLoraLoaderMixin, FromSingleFileMixin, FluxIPAdapterMixin):
王奇勋's avatar
王奇勋 committed
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
    r"""
    The Flux pipeline for text-to-image generation.

    Reference: https://blackforestlabs.ai/announcing-black-forest-labs/

    Args:
        transformer ([`FluxTransformer2DModel`]):
            Conditional Transformer (MMDiT) architecture to denoise the encoded image latents.
        scheduler ([`FlowMatchEulerDiscreteScheduler`]):
            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
        vae ([`AutoencoderKL`]):
            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
        text_encoder ([`CLIPTextModel`]):
            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
        text_encoder_2 ([`T5EncoderModel`]):
            [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically
            the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant.
        tokenizer (`CLIPTokenizer`):
            Tokenizer of class
            [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer).
        tokenizer_2 (`T5TokenizerFast`):
            Second Tokenizer of class
            [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast).
    """

203
204
    model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->transformer->vae"
    _optional_components = ["image_encoder", "feature_extractor"]
王奇勋's avatar
王奇勋 committed
205
206
207
208
209
210
211
212
213
214
215
    _callback_tensor_inputs = ["latents", "prompt_embeds"]

    def __init__(
        self,
        scheduler: FlowMatchEulerDiscreteScheduler,
        vae: AutoencoderKL,
        text_encoder: CLIPTextModel,
        tokenizer: CLIPTokenizer,
        text_encoder_2: T5EncoderModel,
        tokenizer_2: T5TokenizerFast,
        transformer: FluxTransformer2DModel,
216
217
218
        controlnet: Union[
            FluxControlNetModel, List[FluxControlNetModel], Tuple[FluxControlNetModel], FluxMultiControlNetModel
        ],
219
220
        image_encoder: CLIPVisionModelWithProjection = None,
        feature_extractor: CLIPImageProcessor = None,
王奇勋's avatar
王奇勋 committed
221
222
    ):
        super().__init__()
hlky's avatar
hlky committed
223
224
        if isinstance(controlnet, (list, tuple)):
            controlnet = FluxMultiControlNetModel(controlnet)
王奇勋's avatar
王奇勋 committed
225
226
227
228
229
230
231
232
233
234

        self.register_modules(
            vae=vae,
            text_encoder=text_encoder,
            text_encoder_2=text_encoder_2,
            tokenizer=tokenizer,
            tokenizer_2=tokenizer_2,
            transformer=transformer,
            scheduler=scheduler,
            controlnet=controlnet,
235
236
            image_encoder=image_encoder,
            feature_extractor=feature_extractor,
王奇勋's avatar
王奇勋 committed
237
        )
hlky's avatar
hlky committed
238
        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8
Dhruv Nair's avatar
Dhruv Nair committed
239
240
241
        # Flux latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible
        # by the patch size. So the vae scale factor is multiplied by the patch size to account for this
        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2)
王奇勋's avatar
王奇勋 committed
242
243
244
        self.tokenizer_max_length = (
            self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77
        )
245
        self.default_sample_size = 128
王奇勋's avatar
王奇勋 committed
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260

    def _get_t5_prompt_embeds(
        self,
        prompt: Union[str, List[str]] = None,
        num_images_per_prompt: int = 1,
        max_sequence_length: int = 512,
        device: Optional[torch.device] = None,
        dtype: Optional[torch.dtype] = None,
    ):
        device = device or self._execution_device
        dtype = dtype or self.text_encoder.dtype

        prompt = [prompt] if isinstance(prompt, str) else prompt
        batch_size = len(prompt)

261
262
263
        if isinstance(self, TextualInversionLoaderMixin):
            prompt = self.maybe_convert_prompt(prompt, self.tokenizer)

王奇勋's avatar
王奇勋 committed
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
        text_inputs = self.tokenizer_2(
            prompt,
            padding="max_length",
            max_length=max_sequence_length,
            truncation=True,
            return_length=False,
            return_overflowing_tokens=False,
            return_tensors="pt",
        )
        text_input_ids = text_inputs.input_ids
        untruncated_ids = self.tokenizer_2(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_2.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1])
            logger.warning(
                "The following part of your input was truncated because `max_sequence_length` is set to "
                f" {max_sequence_length} tokens: {removed_text}"
            )

        prompt_embeds = self.text_encoder_2(text_input_ids.to(device), output_hidden_states=False)[0]

        dtype = self.text_encoder_2.dtype
        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)

        _, seq_len, _ = prompt_embeds.shape

        # duplicate text embeddings and attention mask 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(batch_size * num_images_per_prompt, seq_len, -1)

        return prompt_embeds

    def _get_clip_prompt_embeds(
        self,
        prompt: Union[str, List[str]],
        num_images_per_prompt: int = 1,
        device: Optional[torch.device] = None,
    ):
        device = device or self._execution_device

        prompt = [prompt] if isinstance(prompt, str) else prompt
        batch_size = len(prompt)

307
308
309
        if isinstance(self, TextualInversionLoaderMixin):
            prompt = self.maybe_convert_prompt(prompt, self.tokenizer)

王奇勋's avatar
王奇勋 committed
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
        text_inputs = self.tokenizer(
            prompt,
            padding="max_length",
            max_length=self.tokenizer_max_length,
            truncation=True,
            return_overflowing_tokens=False,
            return_length=False,
            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_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_max_length} tokens: {removed_text}"
            )
        prompt_embeds = self.text_encoder(text_input_ids.to(device), output_hidden_states=False)

        # Use pooled output of CLIPTextModel
        prompt_embeds = prompt_embeds.pooler_output
        prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)

        # duplicate text embeddings for each generation per prompt, using mps friendly method
335
        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt)
王奇勋's avatar
王奇勋 committed
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
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
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
        prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, -1)

        return prompt_embeds

    def encode_prompt(
        self,
        prompt: Union[str, List[str]],
        prompt_2: Union[str, List[str]],
        device: Optional[torch.device] = None,
        num_images_per_prompt: int = 1,
        prompt_embeds: Optional[torch.FloatTensor] = None,
        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
        max_sequence_length: int = 512,
        lora_scale: Optional[float] = None,
    ):
        r"""

        Args:
            prompt (`str` or `List[str]`, *optional*):
                prompt to be encoded
            prompt_2 (`str` or `List[str]`, *optional*):
                The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
                used in all text-encoders
            device: (`torch.device`):
                torch device
            num_images_per_prompt (`int`):
                number of images that should be generated per prompt
            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.
            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
                If not provided, pooled text embeddings will be generated from `prompt` input argument.
            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.
            lora_scale (`float`, *optional*):
                A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
        """
        device = device or self._execution_device

        # 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, FluxLoraLoaderMixin):
            self._lora_scale = lora_scale

            # dynamically adjust the LoRA scale
            if self.text_encoder is not None and USE_PEFT_BACKEND:
                scale_lora_layers(self.text_encoder, lora_scale)
            if self.text_encoder_2 is not None and USE_PEFT_BACKEND:
                scale_lora_layers(self.text_encoder_2, lora_scale)

        prompt = [prompt] if isinstance(prompt, str) else prompt

        if prompt_embeds is None:
            prompt_2 = prompt_2 or prompt
            prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2

            # We only use the pooled prompt output from the CLIPTextModel
            pooled_prompt_embeds = self._get_clip_prompt_embeds(
                prompt=prompt,
                device=device,
                num_images_per_prompt=num_images_per_prompt,
            )
            prompt_embeds = self._get_t5_prompt_embeds(
                prompt=prompt_2,
                num_images_per_prompt=num_images_per_prompt,
                max_sequence_length=max_sequence_length,
                device=device,
            )

        if self.text_encoder is not None:
            if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND:
                # Retrieve the original scale by scaling back the LoRA layers
                unscale_lora_layers(self.text_encoder, lora_scale)

        if self.text_encoder_2 is not None:
            if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND:
                # Retrieve the original scale by scaling back the LoRA layers
                unscale_lora_layers(self.text_encoder_2, lora_scale)

        dtype = self.text_encoder.dtype if self.text_encoder is not None else self.transformer.dtype
YiYi Xu's avatar
YiYi Xu committed
418
        text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=device, dtype=dtype)
王奇勋's avatar
王奇勋 committed
419
420
421

        return prompt_embeds, pooled_prompt_embeds, text_ids

422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
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
    # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.encode_image
    def encode_image(self, image, device, num_images_per_prompt):
        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)
        image_embeds = self.image_encoder(image).image_embeds
        image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
        return image_embeds

    # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.prepare_ip_adapter_image_embeds
    def prepare_ip_adapter_image_embeds(
        self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt
    ):
        image_embeds = []
        if ip_adapter_image_embeds is None:
            if not isinstance(ip_adapter_image, list):
                ip_adapter_image = [ip_adapter_image]

            if len(ip_adapter_image) != len(self.transformer.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.transformer.encoder_hid_proj.image_projection_layers)} IP Adapters."
                )

            for single_ip_adapter_image, image_proj_layer in zip(
                ip_adapter_image, self.transformer.encoder_hid_proj.image_projection_layers
            ):
                single_image_embeds = self.encode_image(single_ip_adapter_image, device, 1)

                image_embeds.append(single_image_embeds[None, :])
        else:
            for single_image_embeds in ip_adapter_image_embeds:
                image_embeds.append(single_image_embeds)

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

        return ip_adapter_image_embeds

王奇勋's avatar
王奇勋 committed
466
467
468
469
470
471
    def check_inputs(
        self,
        prompt,
        prompt_2,
        height,
        width,
472
473
        negative_prompt=None,
        negative_prompt_2=None,
王奇勋's avatar
王奇勋 committed
474
        prompt_embeds=None,
475
        negative_prompt_embeds=None,
王奇勋's avatar
王奇勋 committed
476
        pooled_prompt_embeds=None,
477
        negative_pooled_prompt_embeds=None,
王奇勋's avatar
王奇勋 committed
478
479
480
        callback_on_step_end_tensor_inputs=None,
        max_sequence_length=None,
    ):
Dhruv Nair's avatar
Dhruv Nair committed
481
482
483
        if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0:
            logger.warning(
                f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly"
484
            )
王奇勋's avatar
王奇勋 committed
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
511

        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]}"
            )

        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_2 is not None and prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `prompt_2`: {prompt_2} 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)}")
        elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)):
            raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}")

512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
        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."
            )
        elif negative_prompt_2 is not None and negative_prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} 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}."
                )

王奇勋's avatar
王奇勋 committed
531
532
533
534
        if prompt_embeds is not None and pooled_prompt_embeds is None:
            raise ValueError(
                "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`."
            )
535
536
537
538
        if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None:
            raise ValueError(
                "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`."
            )
王奇勋's avatar
王奇勋 committed
539
540
541
542
543
544
545

        if max_sequence_length is not None and max_sequence_length > 512:
            raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}")

    @staticmethod
    # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._prepare_latent_image_ids
    def _prepare_latent_image_ids(batch_size, height, width, device, dtype):
546
        latent_image_ids = torch.zeros(height, width, 3)
547
548
        latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height)[:, None]
        latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width)[None, :]
王奇勋's avatar
王奇勋 committed
549
550
551
552

        latent_image_id_height, latent_image_id_width, latent_image_id_channels = latent_image_ids.shape

        latent_image_ids = latent_image_ids.reshape(
YiYi Xu's avatar
YiYi Xu committed
553
            latent_image_id_height * latent_image_id_width, latent_image_id_channels
王奇勋's avatar
王奇勋 committed
554
555
        )

556
        return latent_image_ids.to(device=device, dtype=dtype)
王奇勋's avatar
王奇勋 committed
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571

    @staticmethod
    # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._pack_latents
    def _pack_latents(latents, batch_size, num_channels_latents, height, width):
        latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2)
        latents = latents.permute(0, 2, 4, 1, 3, 5)
        latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4)

        return latents

    @staticmethod
    # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._unpack_latents
    def _unpack_latents(latents, height, width, vae_scale_factor):
        batch_size, num_patches, channels = latents.shape

Dhruv Nair's avatar
Dhruv Nair committed
572
573
574
575
        # VAE applies 8x compression on images but we must also account for packing which requires
        # latent height and width to be divisible by 2.
        height = 2 * (int(height) // (vae_scale_factor * 2))
        width = 2 * (int(width) // (vae_scale_factor * 2))
王奇勋's avatar
王奇勋 committed
576

577
        latents = latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2)
王奇勋's avatar
王奇勋 committed
578
579
        latents = latents.permute(0, 3, 1, 4, 2, 5)

580
        latents = latents.reshape(batch_size, channels // (2 * 2), height, width)
王奇勋's avatar
王奇勋 committed
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595

        return latents

    # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.prepare_latents
    def prepare_latents(
        self,
        batch_size,
        num_channels_latents,
        height,
        width,
        dtype,
        device,
        generator,
        latents=None,
    ):
Dhruv Nair's avatar
Dhruv Nair committed
596
597
598
599
        # VAE applies 8x compression on images but we must also account for packing which requires
        # latent height and width to be divisible by 2.
        height = 2 * (int(height) // (self.vae_scale_factor * 2))
        width = 2 * (int(width) // (self.vae_scale_factor * 2))
王奇勋's avatar
王奇勋 committed
600
601
602
603

        shape = (batch_size, num_channels_latents, height, width)

        if latents is not None:
604
            latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype)
王奇勋's avatar
王奇勋 committed
605
606
607
608
609
610
611
612
613
614
615
            return latents.to(device=device, dtype=dtype), latent_image_ids

        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."
            )

        latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
        latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width)

616
        latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype)
王奇勋's avatar
王奇勋 committed
617
618
619
620
621
622
623
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
671
672
673
674
675
676

        return latents, latent_image_ids

    # Copied from diffusers.pipelines.controlnet_sd3.pipeline_stable_diffusion_3_controlnet.StableDiffusion3ControlNetPipeline.prepare_image
    def prepare_image(
        self,
        image,
        width,
        height,
        batch_size,
        num_images_per_prompt,
        device,
        dtype,
        do_classifier_free_guidance=False,
        guess_mode=False,
    ):
        if isinstance(image, torch.Tensor):
            pass
        else:
            image = self.image_processor.preprocess(image, height=height, width=width)

        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

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

    @property
    def joint_attention_kwargs(self):
        return self._joint_attention_kwargs

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

    @property
    def interrupt(self):
        return self._interrupt

    @torch.no_grad()
    @replace_example_docstring(EXAMPLE_DOC_STRING)
    def __call__(
        self,
        prompt: Union[str, List[str]] = None,
        prompt_2: Optional[Union[str, List[str]]] = None,
677
678
679
        negative_prompt: Union[str, List[str]] = None,
        negative_prompt_2: Optional[Union[str, List[str]]] = None,
        true_cfg_scale: float = 1.0,
王奇勋's avatar
王奇勋 committed
680
681
682
        height: Optional[int] = None,
        width: Optional[int] = None,
        num_inference_steps: int = 28,
hlky's avatar
hlky committed
683
        sigmas: Optional[List[float]] = None,
王奇勋's avatar
王奇勋 committed
684
        guidance_scale: float = 7.0,
685
686
        control_guidance_start: Union[float, List[float]] = 0.0,
        control_guidance_end: Union[float, List[float]] = 1.0,
王奇勋's avatar
王奇勋 committed
687
        control_image: PipelineImageInput = None,
688
        control_mode: Optional[Union[int, List[int]]] = None,
王奇勋's avatar
王奇勋 committed
689
690
691
692
693
694
        controlnet_conditioning_scale: Union[float, List[float]] = 1.0,
        num_images_per_prompt: Optional[int] = 1,
        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
        latents: Optional[torch.FloatTensor] = None,
        prompt_embeds: Optional[torch.FloatTensor] = None,
        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
695
696
697
698
699
700
        ip_adapter_image: Optional[PipelineImageInput] = None,
        ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None,
        negative_ip_adapter_image: Optional[PipelineImageInput] = None,
        negative_ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None,
        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
王奇勋's avatar
王奇勋 committed
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
        output_type: Optional[str] = "pil",
        return_dict: bool = True,
        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
        max_sequence_length: int = 512,
    ):
        r"""
        Function invoked when calling the pipeline for generation.

        Args:
            prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
                instead.
            prompt_2 (`str` or `List[str]`, *optional*):
                The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
                will be used instead
            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                The height in pixels of the generated image. This is set to 1024 by default for the best results.
            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                The width in pixels of the generated image. This is set to 1024 by default for the best results.
            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.
hlky's avatar
hlky committed
725
726
727
728
            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.
王奇勋's avatar
王奇勋 committed
729
730
731
732
733
734
            guidance_scale (`float`, *optional*, defaults to 7.0):
                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
                `guidance_scale` is defined as `w` of equation 2. of [Imagen
                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
                usually at the expense of lower image quality.
735
736
737
738
            control_guidance_start (`float` or `List[float]`, *optional*, defaults to 0.0):
                The percentage of total steps at which the ControlNet starts applying.
            control_guidance_end (`float` or `List[float]`, *optional*, defaults to 1.0):
                The percentage of total steps at which the ControlNet stops applying.
739
740
741
742
743
744
745
746
747
748
749
750
751
752
            control_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]]`):
                The ControlNet input condition to provide guidance to the `unet` for generation. If the type is
                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.
            controlnet_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 1.0):
                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.
            control_mode (`int` or `List[int]`,, *optional*, defaults to None):
                The control mode when applying ControlNet-Union.
王奇勋's avatar
王奇勋 committed
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
            num_images_per_prompt (`int`, *optional*, defaults to 1):
                The number of images to generate per prompt.
            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
                to make generation deterministic.
            latents (`torch.FloatTensor`, *optional*):
                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
                tensor will ge generated by sampling using the supplied random `generator`.
            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.
            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
                If not provided, pooled text embeddings will be generated from `prompt` input argument.
768
769
770
771
772
773
774
775
776
777
778
            ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
            ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*):
                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)`. If not
                provided, embeddings are computed from the `ip_adapter_image` input argument.
            negative_ip_adapter_image:
                (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
            negative_ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*):
                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)`. If not
                provided, embeddings are computed from the `ip_adapter_image` input argument.
王奇勋's avatar
王奇勋 committed
779
780
781
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
808
809
            output_type (`str`, *optional*, defaults to `"pil"`):
                The output format of the generate image. Choose between
                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`~pipelines.flux.FluxPipelineOutput`] instead of a plain tuple.
            joint_attention_kwargs (`dict`, *optional*):
                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
                `self.processor` in
                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
            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 pipeline class.
            max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`.

        Examples:

        Returns:
            [`~pipelines.flux.FluxPipelineOutput`] or `tuple`: [`~pipelines.flux.FluxPipelineOutput`] if `return_dict`
            is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated
            images.
        """

        height = height or self.default_sample_size * self.vae_scale_factor
        width = width or self.default_sample_size * self.vae_scale_factor

810
811
812
813
814
815
816
817
818
819
820
        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(self.controlnet.nets) if isinstance(self.controlnet, FluxMultiControlNetModel) else 1
            control_guidance_start, control_guidance_end = (
                mult * [control_guidance_start],
                mult * [control_guidance_end],
            )

王奇勋's avatar
王奇勋 committed
821
822
823
824
825
826
        # 1. Check inputs. Raise error if not correct
        self.check_inputs(
            prompt,
            prompt_2,
            height,
            width,
827
828
            negative_prompt=negative_prompt,
            negative_prompt_2=negative_prompt_2,
王奇勋's avatar
王奇勋 committed
829
            prompt_embeds=prompt_embeds,
830
            negative_prompt_embeds=negative_prompt_embeds,
王奇勋's avatar
王奇勋 committed
831
            pooled_prompt_embeds=pooled_prompt_embeds,
832
            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
王奇勋's avatar
王奇勋 committed
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
            callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
            max_sequence_length=max_sequence_length,
        )

        self._guidance_scale = guidance_scale
        self._joint_attention_kwargs = joint_attention_kwargs
        self._interrupt = False

        # 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
        dtype = self.transformer.dtype

Aryan's avatar
Aryan committed
852
        # 3. Prepare text embeddings
王奇勋's avatar
王奇勋 committed
853
854
855
        lora_scale = (
            self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None
        )
856
        do_true_cfg = true_cfg_scale > 1 and negative_prompt is not None
王奇勋's avatar
王奇勋 committed
857
858
859
860
861
862
863
864
865
866
867
868
869
870
        (
            prompt_embeds,
            pooled_prompt_embeds,
            text_ids,
        ) = self.encode_prompt(
            prompt=prompt,
            prompt_2=prompt_2,
            prompt_embeds=prompt_embeds,
            pooled_prompt_embeds=pooled_prompt_embeds,
            device=device,
            num_images_per_prompt=num_images_per_prompt,
            max_sequence_length=max_sequence_length,
            lora_scale=lora_scale,
        )
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
        if do_true_cfg:
            (
                negative_prompt_embeds,
                negative_pooled_prompt_embeds,
                _,
            ) = self.encode_prompt(
                prompt=negative_prompt,
                prompt_2=negative_prompt_2,
                prompt_embeds=negative_prompt_embeds,
                pooled_prompt_embeds=negative_pooled_prompt_embeds,
                device=device,
                num_images_per_prompt=num_images_per_prompt,
                max_sequence_length=max_sequence_length,
                lora_scale=lora_scale,
            )
王奇勋's avatar
王奇勋 committed
886
887
888
889
890
891
892
893
894
895
896

        # 3. Prepare control image
        num_channels_latents = self.transformer.config.in_channels // 4
        if isinstance(self.controlnet, FluxControlNetModel):
            control_image = self.prepare_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,
897
                dtype=self.vae.dtype,
王奇勋's avatar
王奇勋 committed
898
899
900
            )
            height, width = control_image.shape[-2:]

901
902
903
904
            # xlab controlnet has a input_hint_block and instantx controlnet does not
            controlnet_blocks_repeat = False if self.controlnet.input_hint_block is None else True
            if self.controlnet.input_hint_block is None:
                # vae encode
905
                control_image = retrieve_latents(self.vae.encode(control_image), generator=generator)
906
907
908
909
910
911
912
913
914
915
916
                control_image = (control_image - self.vae.config.shift_factor) * self.vae.config.scaling_factor

                # pack
                height_control_image, width_control_image = control_image.shape[2:]
                control_image = self._pack_latents(
                    control_image,
                    batch_size * num_images_per_prompt,
                    num_channels_latents,
                    height_control_image,
                    width_control_image,
                )
王奇勋's avatar
王奇勋 committed
917

918
            # Here we ensure that `control_mode` has the same length as the control_image.
919
            if control_mode is not None:
920
921
                if not isinstance(control_mode, int):
                    raise ValueError(" For `FluxControlNet`, `control_mode` should be an `int` or `None`")
922
                control_mode = torch.tensor(control_mode).to(device, dtype=torch.long)
923
                control_mode = control_mode.view(-1, 1).expand(control_image.shape[0], 1)
924
925
926

        elif isinstance(self.controlnet, FluxMultiControlNetModel):
            control_images = []
927
928
929
            # xlab controlnet has a input_hint_block and instantx controlnet does not
            controlnet_blocks_repeat = False if self.controlnet.nets[0].input_hint_block is None else True
            for i, control_image_ in enumerate(control_image):
930
931
932
933
934
935
936
                control_image_ = self.prepare_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,
937
                    dtype=self.vae.dtype,
938
939
940
                )
                height, width = control_image_.shape[-2:]

941
942
                if self.controlnet.nets[0].input_hint_block is None:
                    # vae encode
943
                    control_image_ = retrieve_latents(self.vae.encode(control_image_), generator=generator)
944
945
946
947
948
949
950
951
952
953
954
                    control_image_ = (control_image_ - self.vae.config.shift_factor) * self.vae.config.scaling_factor

                    # pack
                    height_control_image, width_control_image = control_image_.shape[2:]
                    control_image_ = self._pack_latents(
                        control_image_,
                        batch_size * num_images_per_prompt,
                        num_channels_latents,
                        height_control_image,
                        width_control_image,
                    )
955
956
957
958
                control_images.append(control_image_)

            control_image = control_images

959
960
961
962
963
964
965
966
            # Here we ensure that `control_mode` has the same length as the control_image.
            if isinstance(control_mode, list) and len(control_mode) != len(control_image):
                raise ValueError(
                    "For Multi-ControlNet, `control_mode` must be a list of the same "
                    + " length as the number of controlnets (control images) specified"
                )
            if not isinstance(control_mode, list):
                control_mode = [control_mode] * len(control_image)
967
            # set control mode
968
969
970
971
972
973
974
            control_modes = []
            for cmode in control_mode:
                if cmode is None:
                    cmode = -1
                control_mode = torch.tensor(cmode).expand(control_images[0].shape[0]).to(device, dtype=torch.long)
                control_modes.append(control_mode)
            control_mode = control_modes
975

王奇勋's avatar
王奇勋 committed
976
977
978
979
980
981
982
983
984
985
986
987
988
989
        # 4. Prepare latent variables
        num_channels_latents = self.transformer.config.in_channels // 4
        latents, latent_image_ids = self.prepare_latents(
            batch_size * num_images_per_prompt,
            num_channels_latents,
            height,
            width,
            prompt_embeds.dtype,
            device,
            generator,
            latents,
        )

        # 5. Prepare timesteps
hlky's avatar
hlky committed
990
        sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas
王奇勋's avatar
王奇勋 committed
991
992
993
        image_seq_len = latents.shape[1]
        mu = calculate_shift(
            image_seq_len,
994
995
996
            self.scheduler.config.get("base_image_seq_len", 256),
            self.scheduler.config.get("max_image_seq_len", 4096),
            self.scheduler.config.get("base_shift", 0.5),
997
            self.scheduler.config.get("max_shift", 1.15),
王奇勋's avatar
王奇勋 committed
998
999
1000
1001
1002
        )
        timesteps, num_inference_steps = retrieve_timesteps(
            self.scheduler,
            num_inference_steps,
            device,
hlky's avatar
hlky committed
1003
            sigmas=sigmas,
王奇勋's avatar
王奇勋 committed
1004
1005
1006
1007
1008
1009
            mu=mu,
        )

        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
        self._num_timesteps = len(timesteps)

1010
1011
1012
1013
1014
1015
1016
1017
1018
        # 6. Create tensor stating which controlnets to keep
        controlnet_keep = []
        for i in range(len(timesteps)):
            keeps = [
                1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e)
                for s, e in zip(control_guidance_start, control_guidance_end)
            ]
            controlnet_keep.append(keeps[0] if isinstance(self.controlnet, FluxControlNetModel) else keeps)

1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
        if (ip_adapter_image is not None or ip_adapter_image_embeds is not None) and (
            negative_ip_adapter_image is None and negative_ip_adapter_image_embeds is None
        ):
            negative_ip_adapter_image = np.zeros((width, height, 3), dtype=np.uint8)
        elif (ip_adapter_image is None and ip_adapter_image_embeds is None) and (
            negative_ip_adapter_image is not None or negative_ip_adapter_image_embeds is not None
        ):
            ip_adapter_image = np.zeros((width, height, 3), dtype=np.uint8)

        if self.joint_attention_kwargs is None:
            self._joint_attention_kwargs = {}

        image_embeds = None
        negative_image_embeds = None
        if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
            image_embeds = self.prepare_ip_adapter_image_embeds(
                ip_adapter_image,
                ip_adapter_image_embeds,
                device,
                batch_size * num_images_per_prompt,
            )
        if negative_ip_adapter_image is not None or negative_ip_adapter_image_embeds is not None:
            negative_image_embeds = self.prepare_ip_adapter_image_embeds(
                negative_ip_adapter_image,
                negative_ip_adapter_image_embeds,
                device,
                batch_size * num_images_per_prompt,
            )

1048
        # 7. Denoising loop
王奇勋's avatar
王奇勋 committed
1049
1050
1051
1052
1053
        with self.progress_bar(total=num_inference_steps) as progress_bar:
            for i, t in enumerate(timesteps):
                if self.interrupt:
                    continue

1054
1055
                if image_embeds is not None:
                    self._joint_attention_kwargs["ip_adapter_image_embeds"] = image_embeds
王奇勋's avatar
王奇勋 committed
1056
1057
1058
                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
                timestep = t.expand(latents.shape[0]).to(latents.dtype)

1059
1060
1061
1062
1063
1064
                if isinstance(self.controlnet, FluxMultiControlNetModel):
                    use_guidance = self.controlnet.nets[0].config.guidance_embeds
                else:
                    use_guidance = self.controlnet.config.guidance_embeds

                guidance = torch.tensor([guidance_scale], device=device) if use_guidance else None
1065
                guidance = guidance.expand(latents.shape[0]) if guidance is not None else None
王奇勋's avatar
王奇勋 committed
1066

1067
1068
1069
1070
1071
1072
1073
1074
                if isinstance(controlnet_keep[i], list):
                    cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])]
                else:
                    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]

王奇勋's avatar
王奇勋 committed
1075
1076
1077
1078
                # controlnet
                controlnet_block_samples, controlnet_single_block_samples = self.controlnet(
                    hidden_states=latents,
                    controlnet_cond=control_image,
1079
                    controlnet_mode=control_mode,
1080
                    conditioning_scale=cond_scale,
王奇勋's avatar
王奇勋 committed
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
                    timestep=timestep / 1000,
                    guidance=guidance,
                    pooled_projections=pooled_prompt_embeds,
                    encoder_hidden_states=prompt_embeds,
                    txt_ids=text_ids,
                    img_ids=latent_image_ids,
                    joint_attention_kwargs=self.joint_attention_kwargs,
                    return_dict=False,
                )

1091
1092
1093
1094
1095
                guidance = (
                    torch.tensor([guidance_scale], device=device) if self.transformer.config.guidance_embeds else None
                )
                guidance = guidance.expand(latents.shape[0]) if guidance is not None else None

王奇勋's avatar
王奇勋 committed
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
                noise_pred = self.transformer(
                    hidden_states=latents,
                    timestep=timestep / 1000,
                    guidance=guidance,
                    pooled_projections=pooled_prompt_embeds,
                    encoder_hidden_states=prompt_embeds,
                    controlnet_block_samples=controlnet_block_samples,
                    controlnet_single_block_samples=controlnet_single_block_samples,
                    txt_ids=text_ids,
                    img_ids=latent_image_ids,
                    joint_attention_kwargs=self.joint_attention_kwargs,
                    return_dict=False,
1108
                    controlnet_blocks_repeat=controlnet_blocks_repeat,
王奇勋's avatar
王奇勋 committed
1109
1110
                )[0]

1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
                if do_true_cfg:
                    if negative_image_embeds is not None:
                        self._joint_attention_kwargs["ip_adapter_image_embeds"] = negative_image_embeds
                    neg_noise_pred = self.transformer(
                        hidden_states=latents,
                        timestep=timestep / 1000,
                        guidance=guidance,
                        pooled_projections=negative_pooled_prompt_embeds,
                        encoder_hidden_states=negative_prompt_embeds,
                        controlnet_block_samples=controlnet_block_samples,
                        controlnet_single_block_samples=controlnet_single_block_samples,
                        txt_ids=text_ids,
                        img_ids=latent_image_ids,
                        joint_attention_kwargs=self.joint_attention_kwargs,
                        return_dict=False,
                        controlnet_blocks_repeat=controlnet_blocks_repeat,
                    )[0]
                    noise_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred)

王奇勋's avatar
王奇勋 committed
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
                # compute the previous noisy sample x_t -> x_t-1
                latents_dtype = latents.dtype
                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]

                if latents.dtype != latents_dtype:
                    if torch.backends.mps.is_available():
                        # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
                        latents = latents.to(latents_dtype)

                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)

                # 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 XLA_AVAILABLE:
                    xm.mark_step()

        if output_type == "latent":
            image = latents

        else:
            latents = self._unpack_latents(latents, height, width, self.vae_scale_factor)
            latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor

            image = self.vae.decode(latents, return_dict=False)[0]
            image = self.image_processor.postprocess(image, output_type=output_type)

        # Offload all models
        self.maybe_free_model_hooks()

        if not return_dict:
            return (image,)

        return FluxPipelineOutput(images=image)