flux/pipeline_flux_controlnet_inpainting.py

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

import numpy as np
import PIL
import torch
from transformers import (
    CLIPTextModel,
    CLIPTokenizer,
    T5EncoderModel,
    T5TokenizerFast,
)

from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
from diffusers.loaders import FluxLoraLoaderMixin, FromSingleFileMixin, TextualInversionLoaderMixin
from diffusers.models.autoencoders import AutoencoderKL
from .controlnet_flux import FluxControlNetModel, FluxMultiControlNetModel
from .transformer_flux import FluxTransformer2DModel
from .scheduling_flow_match_euler_discrete import FlowMatchEulerDiscreteScheduler
from diffusers.utils import (
    USE_PEFT_BACKEND,
    is_torch_xla_available,
    logging,
    replace_example_docstring,
    scale_lora_layers,
    unscale_lora_layers,
)
from diffusers.utils.torch_utils import randn_tensor
from diffusers.pipelines.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__)

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

        >>> controlnet = FluxControlNetModel.from_pretrained(
        ...     "InstantX/FLUX.1-dev-controlnet-canny", torch_dtype=torch.float16
        ... )
        >>> pipe = FluxControlNetInpaintPipeline.from_pretrained(
        ...     "black-forest-labs/FLUX.1-schnell", controlnet=controlnet, torch_dtype=torch.float16
        ... )
        >>> pipe.to("cuda")

        >>> control_image = load_image(
        ...     "https://huggingface.co/InstantX/FLUX.1-dev-Controlnet-Canny-alpha/resolve/main/canny.jpg"
        ... )
        >>> init_image = load_image(
        ...     "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png"
        ... )
        >>> mask_image = load_image(
        ...     "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"
        ... )

        >>> prompt = "A girl holding a sign that says InstantX"
        >>> image = pipe(
        ...     prompt,
        ...     image=init_image,
        ...     mask_image=mask_image,
        ...     control_image=control_image,
        ...     control_guidance_start=0.2,
        ...     control_guidance_end=0.8,
        ...     controlnet_conditioning_scale=0.7,
        ...     strength=0.7,
        ...     num_inference_steps=28,
        ...     guidance_scale=3.5,
        ... ).images[0]
        >>> image.save("flux_controlnet_inpaint.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,
    max_shift: float = 1.16,
):
    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


# 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")


# 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,
):
    r"""
    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


class FluxControlNetInpaintPipeline(DiffusionPipeline, FluxLoraLoaderMixin, FromSingleFileMixin):
    r"""
    The Flux controlnet pipeline for inpainting.

    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).
    """

    model_cpu_offload_seq = "text_encoder->text_encoder_2->transformer->vae"
    _optional_components = []
    _callback_tensor_inputs = ["latents", "prompt_embeds"]

    def __init__(
        self,
        scheduler: FlowMatchEulerDiscreteScheduler,
        vae: AutoencoderKL,
        text_encoder: CLIPTextModel,
        tokenizer: CLIPTokenizer,
        transformer: FluxTransformer2DModel,
        controlnet: Union[
            FluxControlNetModel, List[FluxControlNetModel], Tuple[FluxControlNetModel], FluxMultiControlNetModel
        ],
        text_encoder_2: T5EncoderModel | None = None,
        tokenizer_2: T5TokenizerFast | None = None,
    ):
        super().__init__()
        if isinstance(controlnet, (list, tuple)):
            controlnet = FluxMultiControlNetModel(controlnet)

        self.register_modules(
            scheduler=scheduler,
            vae=vae,
            text_encoder=text_encoder,
            tokenizer=tokenizer,
            text_encoder_2=text_encoder_2,
            tokenizer_2=tokenizer_2,
            transformer=transformer,
            controlnet=controlnet,
        )

        self.vae_scale_factor = (
            2 ** (len(self.vae.config.block_out_channels)) if hasattr(self, "vae") and self.vae is not None else 16
        )
        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
        self.mask_processor = VaeImageProcessor(
            vae_scale_factor=self.vae_scale_factor,
            vae_latent_channels=self.vae.config.latent_channels,
            do_normalize=False,
            do_binarize=True,
            do_convert_grayscale=True,
        )
        self.tokenizer_max_length = (
            self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77
        )
        self.default_sample_size = 64

    # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._get_t5_prompt_embeds
    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)

        if isinstance(self, TextualInversionLoaderMixin):
            prompt = self.maybe_convert_prompt(prompt, self.tokenizer_2)

        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

    # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._get_clip_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)

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

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

        return prompt_embeds

    # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.encode_prompt
    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,
        t5_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.
            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
        if t5_prompt_embeds is not None:
            text_ids = torch.zeros(prompt_embeds.shape[1] + t5_prompt_embeds.shape[1], 3).to(device=device, dtype=dtype)
        else:
            text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=device, dtype=dtype)
            

        return prompt_embeds, pooled_prompt_embeds, text_ids

    # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_inpaint.StableDiffusion3InpaintPipeline._encode_vae_image
    def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator):
        if isinstance(generator, list):
            image_latents = [
                retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i])
                for i in range(image.shape[0])
            ]
            image_latents = torch.cat(image_latents, dim=0)
        else:
            image_latents = retrieve_latents(self.vae.encode(image), generator=generator)

        image_latents = (image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor

        return image_latents

    # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_img2img.StableDiffusion3Img2ImgPipeline.get_timesteps
    def get_timesteps(self, num_inference_steps, strength, device):
        # get the original timestep using init_timestep
        init_timestep = min(num_inference_steps * strength, num_inference_steps)

        t_start = int(max(num_inference_steps - init_timestep, 0))
        timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
        if hasattr(self.scheduler, "set_begin_index"):
            self.scheduler.set_begin_index(t_start * self.scheduler.order)

        return timesteps, num_inference_steps - t_start

    def check_inputs(
        self,
        prompt,
        prompt_2,
        image,
        mask_image,
        strength,
        height,
        width,
        output_type,
        prompt_embeds=None,
        pooled_prompt_embeds=None,
        callback_on_step_end_tensor_inputs=None,
        padding_mask_crop=None,
        max_sequence_length=None,
    ):
        if strength < 0 or strength > 1:
            raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")

        if height % 8 != 0 or width % 8 != 0:
            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")

        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)}")

        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`."
            )

        if padding_mask_crop is not None:
            if not isinstance(image, PIL.Image.Image):
                raise ValueError(
                    f"The image should be a PIL image when inpainting mask crop, but is of type" f" {type(image)}."
                )
            if not isinstance(mask_image, PIL.Image.Image):
                raise ValueError(
                    f"The mask image should be a PIL image when inpainting mask crop, but is of type"
                    f" {type(mask_image)}."
                )
            if output_type != "pil":
                raise ValueError(f"The output type should be PIL when inpainting mask crop, but is" f" {output_type}.")

        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):
        latent_image_ids = torch.zeros(height // 2, width // 2, 3)
        latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height // 2)[:, None]
        latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width // 2)[None, :]

        latent_image_id_height, latent_image_id_width, latent_image_id_channels = latent_image_ids.shape

        latent_image_ids = latent_image_ids.reshape(
            latent_image_id_height * latent_image_id_width, latent_image_id_channels
        )

        return latent_image_ids.to(device=device, dtype=dtype)

    @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

        height = height // vae_scale_factor
        width = width // vae_scale_factor

        latents = latents.view(batch_size, height, width, channels // 4, 2, 2)
        latents = latents.permute(0, 3, 1, 4, 2, 5)

        latents = latents.reshape(batch_size, channels // (2 * 2), height * 2, width * 2)

        return latents

    # Copied from diffusers.pipelines.flux.pipeline_flux_inpaint.FluxInpaintPipeline.prepare_latents
    def prepare_latents(
        self,
        image,
        timestep,
        batch_size,
        num_channels_latents,
        height,
        width,
        dtype,
        device,
        generator,
        latents=None,
    ):
        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."
            )

        height = 2 * (int(height) // self.vae_scale_factor)
        width = 2 * (int(width) // self.vae_scale_factor)

        shape = (batch_size, num_channels_latents, height, width)
        latent_image_ids = self._prepare_latent_image_ids(batch_size, height, width, device, dtype)

        image = image.to(device=device, dtype=dtype)
        image_latents = self._encode_vae_image(image=image, generator=generator)

        if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0:
            # expand init_latents for batch_size
            additional_image_per_prompt = batch_size // image_latents.shape[0]
            image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0)
        elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0:
            raise ValueError(
                f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts."
            )
        else:
            image_latents = torch.cat([image_latents], dim=0)

        if latents is None:
            noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
            latents = self.scheduler.scale_noise(image_latents, timestep, noise)
        else:
            noise = latents.to(device)
            latents = noise

        noise = self._pack_latents(noise, batch_size, num_channels_latents, height, width)
        image_latents = self._pack_latents(image_latents, batch_size, num_channels_latents, height, width)
        latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width)
        return latents, noise, image_latents, latent_image_ids

    # Copied from diffusers.pipelines.flux.pipeline_flux_inpaint.FluxInpaintPipeline.prepare_mask_latents
    def prepare_mask_latents(
        self,
        mask,
        masked_image,
        batch_size,
        num_channels_latents,
        num_images_per_prompt,
        height,
        width,
        dtype,
        device,
        generator,
    ):
        height = 2 * (int(height) // self.vae_scale_factor)
        width = 2 * (int(width) // self.vae_scale_factor)
        # resize the mask to latents shape as we concatenate the mask to the latents
        # we do that before converting to dtype to avoid breaking in case we're using cpu_offload
        # and half precision
        mask = torch.nn.functional.interpolate(mask, size=(height, width))
        mask = mask.to(device=device, dtype=dtype)

        batch_size = batch_size * num_images_per_prompt

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

        if masked_image.shape[1] == 16:
            masked_image_latents = masked_image
        else:
            masked_image_latents = retrieve_latents(self.vae.encode(masked_image), generator=generator)

        masked_image_latents = (masked_image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor

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

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

        masked_image_latents = self._pack_latents(
            masked_image_latents,
            batch_size,
            num_channels_latents,
            height,
            width,
        )
        mask = self._pack_latents(
            mask.repeat(1, num_channels_latents, 1, 1),
            batch_size,
            num_channels_latents,
            height,
            width,
        )

        return mask, masked_image_latents

    # 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,
        image: PipelineImageInput = None,
        mask_image: PipelineImageInput = None,
        masked_image_latents: PipelineImageInput = None,
        control_image: PipelineImageInput = None,
        height: Optional[int] = None,
        width: Optional[int] = None,
        strength: float = 0.6,
        padding_mask_crop: Optional[int] = None,
        timesteps: List[int] = None,
        num_inference_steps: int = 28,
        guidance_scale: float = 7.0,
        control_guidance_start: Union[float, List[float]] = 0.0,
        control_guidance_end: Union[float, List[float]] = 1.0,
        control_mode: Optional[Union[int, List[int]]] = None,
        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,
        t5_prompt_embeds: Optional[torch.FloatTensor] = None,
        prompt_embeds_control: Optional[torch.FloatTensor] = None,
        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
        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,
    ):
        """
        Function invoked when calling the pipeline for generation.

        Args:
            prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts to guide the image generation.
            prompt_2 (`str` or `List[str]`, *optional*):
                The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`.
            image (`PIL.Image.Image` or `List[PIL.Image.Image]` or `torch.FloatTensor`):
                The image(s) to inpaint.
            mask_image (`PIL.Image.Image` or `List[PIL.Image.Image]` or `torch.FloatTensor`):
                The mask image(s) to use for inpainting. White pixels in the mask will be repainted, while black pixels
                will be preserved.
            masked_image_latents (`torch.FloatTensor`, *optional*):
                Pre-generated masked image latents.
            control_image (`PIL.Image.Image` or `List[PIL.Image.Image]` or `torch.FloatTensor`):
                The ControlNet input condition. Image to control the generation.
            height (`int`, *optional*, defaults to self.default_sample_size * self.vae_scale_factor):
                The height in pixels of the generated image.
            width (`int`, *optional*, defaults to self.default_sample_size * self.vae_scale_factor):
                The width in pixels of the generated image.
            strength (`float`, *optional*, defaults to 0.6):
                Conceptually, indicates how much to inpaint the masked area. Must be between 0 and 1.
            padding_mask_crop (`int`, *optional*):
                The size of the padding to use when cropping the mask.
            num_inference_steps (`int`, *optional*, defaults to 28):
                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
                expense of slower inference.
            timesteps (`List[int]`, *optional*):
                Custom timesteps to use for the denoising process.
            guidance_scale (`float`, *optional*, defaults to 7.0):
                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
            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.
            control_mode (`int` or `List[int]`, *optional*):
                The mode for the ControlNet. If multiple ControlNets are used, this should be a list.
            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 transformer.
            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 more [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.
            prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated pooled text embeddings.
            output_type (`str`, *optional*, defaults to `"pil"`):
                The output format of the generate image. Choose between `PIL.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*):
                Additional keyword arguments to be passed to the joint attention mechanism.
            callback_on_step_end (`Callable`, *optional*):
                A function that calls at the end of each denoising step during the inference.
            callback_on_step_end_tensor_inputs (`List[str]`, *optional*):
                The list of tensor inputs for the `callback_on_step_end` function.
            max_sequence_length (`int`, *optional*, defaults to 512):
                The maximum length of the sequence to be generated.

        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

        global_height = height
        global_width = width

        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],
            )

        # 1. Check inputs
        self.check_inputs(
            prompt,
            prompt_2,
            image,
            mask_image,
            strength,
            height,
            width,
            output_type=output_type,
            prompt_embeds=prompt_embeds,
            pooled_prompt_embeds=pooled_prompt_embeds,
            callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
            padding_mask_crop=padding_mask_crop,
            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

        # 3. Encode input prompt
        lora_scale = (
            self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None
        )
        prompt_embeds, pooled_prompt_embeds, text_ids = self.encode_prompt(
            prompt=prompt,
            prompt_2=prompt_2,
            prompt_embeds=prompt_embeds,
            t5_prompt_embeds=t5_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,
        )

        # 4. Preprocess mask and image
        if padding_mask_crop is not None:
            crops_coords = self.mask_processor.get_crop_region(
                mask_image, global_width, global_height, pad=padding_mask_crop
            )
            resize_mode = "fill"
        else:
            crops_coords = None
            resize_mode = "default"

        original_image = image
        init_image = self.image_processor.preprocess(
            image, height=global_height, width=global_width, crops_coords=crops_coords, resize_mode=resize_mode
        )
        init_image = init_image.to(dtype=torch.float32)

        # 5. 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=height,
                height=width,
                batch_size=batch_size * num_images_per_prompt,
                num_images_per_prompt=num_images_per_prompt,
                device=device,
                dtype=self.vae.dtype,
            )
            height, width = control_image.shape[-2:]

            # vae encode
            control_image = self.vae.encode(control_image).latent_dist.sample()
            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,
            )

            # set control mode
            if control_mode is not None:
                control_mode = torch.tensor(control_mode).to(device, dtype=torch.long)
                control_mode = control_mode.reshape([-1, 1])

        elif isinstance(self.controlnet, FluxMultiControlNetModel):
            control_images = []

            for control_image_ in control_image:
                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,
                    dtype=self.vae.dtype,
                )
                height, width = control_image_.shape[-2:]

                # vae encode
                control_image_ = self.vae.encode(control_image_).latent_dist.sample()
                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,
                )

                control_images.append(control_image_)

            control_image = control_images

            ## set control mode
            #control_mode_ = []
            #if isinstance(control_mode, list):
            #    for cmode in control_mode:
            #        if cmode is None:
            #            control_mode_.append(-1)
            #        else:
            #            control_mode_.append(cmode)
            #control_mode = torch.tensor(control_mode_).to(device, dtype=torch.long)
            #control_mode = control_mode.reshape([-1, 1])
            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

        # 6. Prepare timesteps

        sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps)
        image_seq_len = (int(global_height) // self.vae_scale_factor) * (int(global_width) // self.vae_scale_factor)
        mu = calculate_shift(
            image_seq_len,
            self.scheduler.config.base_image_seq_len,
            self.scheduler.config.max_image_seq_len,
            self.scheduler.config.base_shift,
            self.scheduler.config.max_shift,
        )
        timesteps, num_inference_steps = retrieve_timesteps(
            self.scheduler,
            num_inference_steps,
            device,
            timesteps,
            sigmas,
            mu=mu,
        )
        timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device)

        if num_inference_steps < 1:
            raise ValueError(
                f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline"
                f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline."
            )
        latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)

        # 7. Prepare latent variables

        latents, noise, image_latents, latent_image_ids = self.prepare_latents(
            init_image,
            latent_timestep,
            batch_size * num_images_per_prompt,
            num_channels_latents,
            global_height,
            global_width,
            prompt_embeds.dtype,
            device,
            generator,
            latents,
        )

        # 8. Prepare mask latents
        mask_condition = self.mask_processor.preprocess(
            mask_image, height=global_height, width=global_width, resize_mode=resize_mode, crops_coords=crops_coords
        )
        if masked_image_latents is None:
            masked_image = init_image * (mask_condition < 0.5)
        else:
            masked_image = masked_image_latents

        mask, masked_image_latents = self.prepare_mask_latents(
            mask_condition,
            masked_image,
            batch_size,
            num_channels_latents,
            num_images_per_prompt,
            global_height,
            global_width,
            prompt_embeds.dtype,
            device,
            generator,
        )

        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)

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

        with self.progress_bar(total=num_inference_steps) as progress_bar:
            for i, t in enumerate(timesteps):
                if self.interrupt:
                    continue

                timestep = t.expand(latents.shape[0]).to(latents.dtype)

                # predict the noise residual
                #if self.controlnet.config.guidance_embeds:
                guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32)
                guidance = guidance.expand(latents.shape[0])
                #else:
                #    guidance = None

                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]

                controlnet_block_samples, controlnet_single_block_samples = self.controlnet(
                    hidden_states=latents,
                    controlnet_cond=control_image,
                    controlnet_mode=control_mode,
                    conditioning_scale=cond_scale,
                    timestep=timestep / 1000,
                    guidance=guidance,
                    pooled_projections=pooled_prompt_embeds,
                    encoder_hidden_states=prompt_embeds_control,
                    t5_encoder_hidden_states=t5_prompt_embeds,
                    txt_ids=text_ids,
                    img_ids=latent_image_ids,
                    joint_attention_kwargs=self.joint_attention_kwargs,
                    return_dict=False,
                )

                if self.transformer.config.guidance_embeds:
                    guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32)
                    guidance = guidance.expand(latents.shape[0])
                else:
                    guidance = None

                noise_pred = self.transformer(
                    hidden_states=latents,
                    timestep=timestep / 1000,
                    guidance=guidance,
                    pooled_projections=pooled_prompt_embeds,
                    encoder_hidden_states=prompt_embeds,
                    t5_encoder_hidden_states=t5_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,
                )[0]

                # 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]

                # For inpainting, we need to apply the mask and add the masked image latents
                init_latents_proper = image_latents
                init_mask = mask

                if i < len(timesteps) - 1:
                    noise_timestep = timesteps[i + 1]
                    init_latents_proper = self.scheduler.scale_noise(
                        init_latents_proper, torch.tensor([noise_timestep]), noise
                    )

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

                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)

                # call the callback, if provided
                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)

                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()

        # Post-processing
        if output_type == "latent":
            image = latents
        else:
            latents = self._unpack_latents(latents, global_height, global_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)