train/src/pipeline_flux_kontext_control.py

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

import numpy as np
import torch
from .transformer_flux import FluxTransformer2DModel
from transformers import (
    CLIPImageProcessor,
    CLIPTextModel,
    CLIPTokenizer,
    CLIPVisionModelWithProjection,
    T5EncoderModel,
    T5TokenizerFast,
)

from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
from diffusers.loaders import FluxIPAdapterMixin, FluxLoraLoaderMixin, FromSingleFileMixin, TextualInversionLoaderMixin
from diffusers.models import AutoencoderKL
from diffusers.schedulers 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 diffusers.pipelines.flux.pipeline_output import FluxPipelineOutput
from torchvision.transforms.functional import pad


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

PREFERRED_KONTEXT_RESOLUTIONS = [
    (672, 1568),
    (688, 1504),
    (720, 1456),
    (752, 1392),
    (800, 1328),
    (832, 1248),
    (880, 1184),
    (944, 1104),
    (1024, 1024),
    (1104, 944),
    (1184, 880),
    (1248, 832),
    (1328, 800),
    (1392, 752),
    (1456, 720),
    (1504, 688),
    (1568, 672),
]


def calculate_shift(
    image_seq_len,
    base_seq_len: int = 256,
    max_seq_len: int = 4096,
    base_shift: float = 0.5,
    max_shift: float = 1.15,
):
    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


def prepare_latent_image_ids_(height, width, device, dtype):
    latent_image_ids = torch.zeros(height, width, 3, device=device, dtype=dtype)
    latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height, device=device)[:, None]  # y
    latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width, device=device)[None, :]  # x
    return latent_image_ids


def prepare_latent_subject_ids(height, width, device, dtype):
    latent_image_ids = torch.zeros(height, width, 3, device=device, dtype=dtype)
    latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height, device=device)[:, None]
    latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width, device=device)[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)


def resize_position_encoding(
    batch_size, original_height, original_width, target_height, target_width, device, dtype
):
    latent_image_ids = prepare_latent_image_ids_(original_height // 2, original_width // 2, device, dtype)
    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
    )

    scale_h = original_height / target_height
    scale_w = original_width / target_width
    latent_image_ids_resized = torch.zeros(target_height // 2, target_width // 2, 3, device=device, dtype=dtype)
    latent_image_ids_resized[..., 1] = (
        latent_image_ids_resized[..., 1] + torch.arange(target_height // 2, device=device)[:, None] * scale_h
    )
    latent_image_ids_resized[..., 2] = (
        latent_image_ids_resized[..., 2] + torch.arange(target_width // 2, device=device)[None, :] * scale_w
    )

    cond_latent_image_id_height, cond_latent_image_id_width, cond_latent_image_id_channels = (
        latent_image_ids_resized.shape
    )
    cond_latent_image_ids = latent_image_ids_resized.reshape(
        cond_latent_image_id_height * cond_latent_image_id_width, cond_latent_image_id_channels
    )
    return latent_image_ids, cond_latent_image_ids


# 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


# 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")


class FluxKontextControlPipeline(
    DiffusionPipeline,
    FluxLoraLoaderMixin,
    FromSingleFileMixin,
    TextualInversionLoaderMixin,
):
    r"""
    The Flux Kontext pipeline for image-to-image and text-to-image generation with EasyControl.

    Reference: https://bfl.ai/announcements/flux-1-kontext-dev

    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,
        text_encoder_2: T5EncoderModel,
        tokenizer_2: T5TokenizerFast,
        transformer: FluxTransformer2DModel,
        image_encoder: CLIPVisionModelWithProjection = None,
        feature_extractor: CLIPImageProcessor = None,
    ):
        super().__init__()

        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,
            image_encoder=None,
            feature_extractor=None,
        )
        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8
        # Flux latents are packed into 2x2 patches, so use VAE factor multiplied by patch size for image processing
        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2)
        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 = 128
        self.latent_channels = self.vae.config.latent_channels if getattr(self, "vae", None) else 16
        # EasyControl: cache multiple control LoRA processors
        self.control_lora_processors: Dict[str, Dict[str, Any]] = {}
        self.control_lora_cond_sizes: Dict[str, Any] = {}
        self.current_control_type: Optional[str] = None

    # 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,
        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
        text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=device, dtype=dtype)

        return prompt_embeds, pooled_prompt_embeds, text_ids

    # Adapted from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.check_inputs
    def check_inputs(
        self,
        prompt,
        prompt_2,
        height,
        width,
        prompt_embeds=None,
        pooled_prompt_embeds=None,
        callback_on_step_end_tensor_inputs=None,
        max_sequence_length=None,
    ):
        if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0:
            raise ValueError(
                f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} 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 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, width, 3)
        latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height)[:, None]
        latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width)[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

        # 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))

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

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

        return latents

    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.flux.pipeline_flux.FluxPipeline.enable_vae_slicing
    def enable_vae_slicing(self):
        r"""
        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
        """
        self.vae.enable_slicing()

    # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.disable_vae_slicing
    def disable_vae_slicing(self):
        r"""
        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
        computing decoding in one step.
        """
        self.vae.disable_slicing()

    # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.enable_vae_tiling
    def enable_vae_tiling(self):
        r"""
        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
        processing larger images.
        """
        self.vae.enable_tiling()

    # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.disable_vae_tiling
    def disable_vae_tiling(self):
        r"""
        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
        computing decoding in one step.
        """
        self.vae.disable_tiling()

    def prepare_latents(
        self,
        batch_size,
        num_channels_latents,
        height,
        width,
        dtype,
        device,
        generator,
        image,
        subject_images,
        spatial_images,
        latents=None,
        cond_size=512,
    ):
        height = 2 * (int(height) // (self.vae_scale_factor * 2))
        width = 2 * (int(width) // (self.vae_scale_factor * 2))
        height_cond = 2 * (cond_size // (self.vae_scale_factor * 2))
        width_cond = 2 * (cond_size // (self.vae_scale_factor * 2))

        image_latents = image_ids = None
        # Prepare noise latents
        shape = (batch_size, num_channels_latents, height, width)
        if latents is None:
            noise_latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
        else:
            noise_latents = latents.to(device=device, dtype=dtype)

        noise_latents = self._pack_latents(noise_latents, batch_size, num_channels_latents, height, width)
        # print(noise_latents.shape)
        noise_latent_image_ids, cond_latent_image_ids_resized = resize_position_encoding(
            batch_size, height, width, height_cond, width_cond, device, dtype
        )
        # noise IDs are marked with 0 in the first channel
        noise_latent_image_ids[..., 0] = 0

        cond_latents_to_concat = []
        latents_ids_to_concat = [noise_latent_image_ids]

        # 1. Prepare `image` (Kontext) latents
        if image is not None:
            image = image.to(device=device, dtype=dtype)
            if image.shape[1] != self.latent_channels:
                image_latents = self._encode_vae_image(image=image, generator=generator)
            else:
                image_latents = image

            image_latent_h, image_latent_w = image_latents.shape[2:]
            image_latents = self._pack_latents(
                image_latents, batch_size, num_channels_latents, image_latent_h, image_latent_w
            )
            image_ids = self._prepare_latent_image_ids(
                batch_size, image_latent_h // 2, image_latent_w // 2, device, dtype
            )
            image_ids[..., 0] = 1  # Mark as condition
            latents_ids_to_concat.append(image_ids)

        # 2. Prepare `subject_images` latents
        if subject_images is not None and len(subject_images) > 0:
            subject_images = subject_images.to(device=device, dtype=dtype)
            subject_image_latents = self._encode_vae_image(image=subject_images, generator=generator)
            subject_latents = self._pack_latents(
                subject_image_latents, batch_size, num_channels_latents, height_cond * len(subject_images), width_cond
            )

            latent_subject_ids = prepare_latent_subject_ids(height_cond // 2, width_cond // 2, device, dtype)
            latent_subject_ids[..., 0] = 1
            latent_subject_ids[:, 1] += image_latent_h // 2
            subject_latent_image_ids = torch.cat([latent_subject_ids for _ in range(len(subject_images))], dim=0)

            cond_latents_to_concat.append(subject_latents)
            latents_ids_to_concat.append(subject_latent_image_ids)

        # 3. Prepare `spatial_images` latents
        if spatial_images is not None and len(spatial_images) > 0:
            spatial_images = spatial_images.to(device=device, dtype=dtype)
            spatial_image_latents = self._encode_vae_image(image=spatial_images, generator=generator)
            cond_latents = self._pack_latents(
                spatial_image_latents, batch_size, num_channels_latents, height_cond * len(spatial_images), width_cond
            )
            cond_latent_image_ids_resized[..., 0] = 2
            cond_latent_image_ids = torch.cat(
                [cond_latent_image_ids_resized for _ in range(len(spatial_images))], dim=0
            )

            cond_latents_to_concat.append(cond_latents)
            latents_ids_to_concat.append(cond_latent_image_ids)

        cond_latents = torch.cat(cond_latents_to_concat, dim=1) if cond_latents_to_concat else None
        latent_image_ids = torch.cat(latents_ids_to_concat, dim=0)

        return noise_latents, image_latents, cond_latents, latent_image_ids

    @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 current_timestep(self):
        return self._current_timestep

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

    @torch.no_grad()
    def __call__(
        self,
        image: Optional[PipelineImageInput] = None,
        prompt: Union[str, List[str]] = None,
        prompt_2: Optional[Union[str, List[str]]] = None,
        height: Optional[int] = None,
        width: Optional[int] = None,
        num_inference_steps: int = 28,
        sigmas: Optional[List[float]] = None,
        guidance_scale: float = 3.5,
        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,
        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,
        cond_size: int = 512,
        control_dict: Optional[Dict[str, Any]] = None,
    ):
        r"""
        Function invoked when calling the pipeline for generation.

        Args:
            image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`):
                `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both
                numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list
                or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a
                list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image
                latents as `image`, but if passing latents directly it is not encoded again.
            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.
            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.
            guidance_scale (`float`, *optional*, defaults to 3.5):
                Guidance scale as defined in [Classifier-Free Diffusion
                Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2.
                of [Imagen Paper](https://huggingface.co/papers/2205.11487). 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.
            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.
            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`.
            cond_size (`int`, *optional*, defaults to 512):
                The size for conditioning images.

        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

        # 1. Check inputs. Raise error if not correct
        self.check_inputs(
            prompt,
            prompt_2,
            height,
            width,
            prompt_embeds=prompt_embeds,
            pooled_prompt_embeds=pooled_prompt_embeds,
            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._current_timestep = None
        self._interrupt = False

        spatial_images = control_dict.get("spatial_images", [])
        subject_images = control_dict.get("subject_images", [])

        # 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

        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,
            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,
        )

        # 3. Preprocess images
        if image is not None and not (isinstance(image, torch.Tensor) and image.size(1) == self.latent_channels):
            img = image[0] if isinstance(image, list) else image
            image_height, image_width = self.image_processor.get_default_height_width(img)
            aspect_ratio = image_width / image_height
            # Kontext is trained on specific resolutions, using one of them is recommended
            _, image_width, image_height = min(
                (abs(aspect_ratio - w / h), w, h) for w, h in PREFERRED_KONTEXT_RESOLUTIONS
            )
            multiple_of = self.vae_scale_factor * 2
            image_width = image_width // multiple_of * multiple_of
            image_height = image_height // multiple_of * multiple_of
            image = self.image_processor.resize(image, image_height, image_width)
            image = self.image_processor.preprocess(image, image_height, image_width)
            image = image.to(dtype=self.vae.dtype)

        if len(subject_images) > 0:
            subject_image_ls = []
            for subject_image in subject_images:
                w, h = subject_image.size[:2]
                scale = cond_size / max(h, w)
                new_h, new_w = int(h * scale), int(w * scale)
                subject_image = self.image_processor.preprocess(subject_image, height=new_h, width=new_w)
                subject_image = subject_image.to(dtype=self.vae.dtype)
                pad_h = cond_size - subject_image.shape[-2]
                pad_w = cond_size - subject_image.shape[-1]
                subject_image = pad(
                    subject_image, padding=(int(pad_w / 2), int(pad_h / 2), int(pad_w / 2), int(pad_h / 2)), fill=0
                )
                subject_image_ls.append(subject_image)
            subject_images = torch.cat(subject_image_ls, dim=-2)
        else:
            subject_images = None

        if len(spatial_images) > 0:
            condition_image_ls = []
            for img in spatial_images:
                condition_image = self.image_processor.preprocess(img, height=cond_size, width=cond_size)
                condition_image = condition_image.to(dtype=self.vae.dtype)
                condition_image_ls.append(condition_image)
            spatial_images = torch.cat(condition_image_ls, dim=-2)
        else:
            spatial_images = None

        # 4. Prepare latent variables
        num_channels_latents = self.transformer.config.in_channels // 4
        latents, image_latents, cond_latents, latent_image_ids = self.prepare_latents(
            batch_size * num_images_per_prompt,
            num_channels_latents,
            height,
            width,
            prompt_embeds.dtype,
            device,
            generator,
            image,
            subject_images,
            spatial_images,
            latents,
            cond_size,
        )

        # 5. Prepare timesteps
        sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas
        image_seq_len = latents.shape[1]
        mu = calculate_shift(
            image_seq_len,
            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),
            self.scheduler.config.get("max_shift", 1.15),
        )
        timesteps, num_inference_steps = retrieve_timesteps(
            self.scheduler,
            num_inference_steps,
            device,
            sigmas=sigmas,
            mu=mu,
        )
        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
        self._num_timesteps = len(timesteps)

        # handle guidance
        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

        # 6. Denoising loop
        self.scheduler.set_begin_index(0)
        with self.progress_bar(total=num_inference_steps) as progress_bar:
            for i, t in enumerate(timesteps):
                if self.interrupt:
                    continue

                latent_model_input = latents
                if image_latents is not None:
                    latent_model_input = torch.cat([latent_model_input, image_latents], dim=1)
                
                self._current_timestep = t
                timestep = t.expand(latents.shape[0]).to(latents.dtype)
                noise_pred = self.transformer(
                    hidden_states=latent_model_input,
                    cond_hidden_states=cond_latents,
                    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,
                )[0]

                noise_pred = noise_pred[:, : latents.size(1)]

                # 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:
                    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()

        self._current_timestep = None

        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)