pipeline.py

# Copyright 2024 Black Forest Labs and The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

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

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

from diffusers.image_processor import VaeImageProcessor
from diffusers.loaders import SD3LoraLoaderMixin
from diffusers.models.autoencoders import AutoencoderKL
# from diffusers.models.transformers import FluxTransformer2DModel
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 recaption import enhance_prompt
from transformer import FluxTransformer2DConditionalModel


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

    XLA_AVAILABLE = True
else:
    XLA_AVAILABLE = False


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

EXAMPLE_DOC_STRING = """
    Examples:
        ```py
        >>> import torch
        >>> from diffusers import FluxPipeline

        >>> pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-schnell", torch_dtype=torch.bfloat16)
        >>> pipe.to("cuda")
        >>> prompt = "A cat holding a sign that says hello world"
        >>> # Depending on the variant being used, the pipeline call will slightly vary.
        >>> # Refer to the pipeline documentation for more details.
        >>> image = pipe(prompt, num_inference_steps=4, guidance_scale=0.0).images[0]
        >>> image.save("flux.png")
        ```
"""


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.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,
):
    """
    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 FluxConditionalPipeline(DiffusionPipeline, SD3LoraLoaderMixin):
    r"""
    The Flux pipeline for text-to-image generation.

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

    Args:
        transformer ([`FluxTransformer2DModel`]):
            Conditional Transformer (MMDiT) architecture to denoise the encoded image latents.
        scheduler ([`FlowMatchEulerDiscreteScheduler`]):
            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
        vae ([`AutoencoderKL`]):
            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
        text_encoder ([`CLIPTextModelWithProjection`]):
            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection),
            specifically the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant,
            with an additional added projection layer that is initialized with a diagonal matrix with the `hidden_size`
            as its dimension.
        text_encoder_2 ([`CLIPTextModelWithProjection`]):
            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection),
            specifically the
            [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)
            variant.
        tokenizer (`CLIPTokenizer`):
            Tokenizer of class
            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
        tokenizer_2 (`CLIPTokenizer`):
            Second Tokenizer of class
            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
    """

    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: FluxTransformer2DConditionalModel,
    ):
        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,
        )
        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.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

    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)

        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",
        )
        prompt_attention_mask = text_inputs.attention_mask
        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, prompt_attention_mask

    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)

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

        return prompt_embeds

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

        Args:
            prompt (`str` or `List[str]`, *optional*):
                prompt to be encoded
            prompt_2 (`str` or `List[str]`, *optional*):
                The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
                used in all text-encoders
            device: (`torch.device`):
                torch device
            num_images_per_prompt (`int`):
                number of images that should be generated per prompt
            prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
                provided, text embeddings will be generated from `prompt` input argument.
            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
                If not provided, pooled text embeddings will be generated from `prompt` input argument.
            clip_skip (`int`, *optional*):
                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
                the output of the pre-final layer will be used for computing the prompt embeddings.
            lora_scale (`float`, *optional*):
                A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
        """
        device = device or self._execution_device

        # set lora scale so that monkey patched LoRA
        # function of text encoder can correctly access it
        if lora_scale is not None and isinstance(self, SD3LoraLoaderMixin):
            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 is not None:
            batch_size = len(prompt)
        else:
            batch_size = prompt_embeds.shape[0]

        prompt_attention_mask = None
        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, prompt_attention_mask = 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, SD3LoraLoaderMixin) 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, SD3LoraLoaderMixin) and USE_PEFT_BACKEND:
                # Retrieve the original scale by scaling back the LoRA layers
                unscale_lora_layers(self.text_encoder_2, lora_scale)

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

        return prompt_embeds, pooled_prompt_embeds, text_ids, prompt_attention_mask

    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,
        image=None
    ):
        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 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
    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[None, :].repeat(batch_size, 1, 1, 1)
        latent_image_ids = latent_image_ids.reshape(
            batch_size,
            latent_image_id_height * latent_image_id_width,
            latent_image_id_channels,
        )

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

    @staticmethod
    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
    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

    def prepare_latents(
        self,
        batch_size,
        num_channels_latents,
        height,
        width,
        dtype,
        device,
        generator,
        latents=None,
    ):
        height = 2 * (int(height) // self.vae_scale_factor)
        width = 2 * (int(width) // self.vae_scale_factor)

        shape = (batch_size, num_channels_latents, height, width)

        if latents is not None:
            latent_image_ids = self._prepare_latent_image_ids(
                batch_size, height, width, device, dtype
            )
            return latents.to(device=device, dtype=dtype), latent_image_ids

        if isinstance(generator, list) and len(generator) != batch_size:
            raise ValueError(
                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
            )

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

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

        return 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 interrupt(self):
        return self._interrupt

    @torch.no_grad()
    @replace_example_docstring(EXAMPLE_DOC_STRING)
    def __call__(
        self,
        prompt: Union[str, List[str]] = None,
        prompt_mask: Optional[Union[torch.FloatTensor, List[torch.FloatTensor]]] = None,
        negative_mask: Optional[
            Union[torch.FloatTensor, List[torch.FloatTensor]]
        ] = None,
        prompt_2: Optional[Union[str, List[str]]] = None,
        height: Optional[int] = None,
        width: Optional[int] = None,
        num_inference_steps: int = 28,
        timesteps: List[int] = 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,
        guidance_scale_real_i: float = 1.0,
        guidance_scale_real_t: float = 1.0,
        negative_prompt: Union[str, List[str]] = "",
        negative_prompt_2: Union[str, List[str]] = "",
        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
        no_cfg_until_timestep: int = 2,
        image: Optional[torch.FloatTensor] = None,
        image_path = None,
        cut_output = True,
        gemini_prompt = True
    ):
        r"""
        Function invoked when calling the pipeline for generation.

        Args:
            prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
                instead.
            prompt_mask (`str` or `List[str]`, *optional*):
                The prompt or prompts to be used as a mask for the image generation. If not defined, `prompt` is used
                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.
            timesteps (`List[int]`, *optional*):
                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
                passed will be used. Must be in descending order.
            guidance_scale (`float`, *optional*, defaults to 7.0):
                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
                `guidance_scale` is defined as `w` of equation 2. of [Imagen
                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
                usually at the expense of lower image quality.
            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`.

        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._guidance_scale_real_i = guidance_scale_real_i
        self._guidance_scale_real_t = guidance_scale_real_t
        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

        prompt = enhance_prompt(image, prompt)
        # if gemini_prompt:
        #     while True:
        #         try:
        #             prompt = enhance_prompt(image, prompt)
        #             break  # Exit the loop if the function succeeds
        #         except Exception as e:
        #             print(f"An error occurred: {e}")

        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,
        )

        if negative_prompt_2 == "" and negative_prompt != "":
            negative_prompt_2 = negative_prompt

        negative_text_ids = text_ids
        if guidance_scale_real_i > 1.0 and (
            negative_prompt_embeds is None or negative_pooled_prompt_embeds is None
        ):
            (
                negative_prompt_embeds,
                negative_pooled_prompt_embeds,
                negative_text_ids,
                _,
            ) = self.encode_prompt(
                prompt=negative_prompt,
                prompt_2=negative_prompt_2,
                prompt_embeds=None,
                pooled_prompt_embeds=None,
                device=device,
                num_images_per_prompt=num_images_per_prompt,
                max_sequence_length=max_sequence_length,
                lora_scale=lora_scale,
            )

        # 3. Preprocess image
        image = self.image_processor.preprocess(image)
        # image = image[..., :512]
        image = torch.nn.functional.interpolate(image, size=512)
        black_image = torch.full((1, 3, 512, 512), -1.0)
        image = torch.cat([image, black_image], dim=3)
        latents_cond = self.vae.encode(image.to(self.vae.dtype).to(self.vae.device)).latent_dist.sample()
        latents_cond = (
            latents_cond - self.vae.config.shift_factor
        ) * self.vae.config.scaling_factor
        # from customization.utils import mask_random_quadrants
        # latent_cond = mask_random_quadrants(latent_cond)

        # 4. Prepare latent variables
        num_channels_latents = self.transformer.config.in_channels // 4
        latents, latent_image_ids = self.prepare_latents(
            batch_size * num_images_per_prompt,
            num_channels_latents,
            height,
            width,
            prompt_embeds.dtype,
            device,
            generator,
            latents,
        )
        # _pack_latents(latents, batch_size, num_channels_latents, height, width)
        latents_cond = self._pack_latents(
            latents_cond, batch_size, num_channels_latents, 2 * (int(height) // self.vae_scale_factor), 2 * (int(width) // self.vae_scale_factor)
        )

        # 5. Prepare timesteps
        sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps)
        image_seq_len = latents.shape[1]
        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,
        )
        num_warmup_steps = max(
            len(timesteps) - num_inference_steps * self.scheduler.order, 0
        )
        self._num_timesteps = len(timesteps)

        latents = latents.to(self.transformer.device)
        latent_image_ids = latent_image_ids.to(self.transformer.device)
        timesteps = timesteps.to(self.transformer.device)
        text_ids = text_ids.to(self.transformer.device)

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

                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
                timestep = t.expand(latents.shape[0]).to(latents.dtype)

                # handle guidance
                if self.transformer.config.guidance_embeds:
                    guidance = torch.tensor(
                        [guidance_scale], device=self.transformer.device
                    )
                    guidance = guidance.expand(latents.shape[0])
                else:
                    guidance = None

                extra_transformer_args = {}
                if prompt_mask is not None:
                    extra_transformer_args["attention_mask"] = prompt_mask.to(
                        device=self.transformer.device
                    )

                noise_pred = self.transformer(
                    hidden_states=latents.to(
                        device=self.transformer.device, dtype=self.transformer.dtype
                    ),
                    # YiYi notes: divide it by 1000 for now because we scale it by 1000 in the transforme rmodel (we should not keep it but I want to keep the inputs same for the model for testing)
                    timestep=timestep / 1000,
                    guidance=guidance,
                    pooled_projections=pooled_prompt_embeds.to(
                        device=self.transformer.device, dtype=self.transformer.dtype
                    ),
                    encoder_hidden_states=prompt_embeds.to(
                        device=self.transformer.device, dtype=self.transformer.dtype
                    ),
                    txt_ids=text_ids,
                    img_ids=latent_image_ids,
                    joint_attention_kwargs=self.joint_attention_kwargs,
                    return_dict=False,
                    condition_hidden_states=latents_cond.to(
                        device=self.transformer.device, dtype=self.transformer.dtype
                    ),
                    **extra_transformer_args,
                )[0]

                # TODO optionally use batch prediction to speed this up.
                if guidance_scale_real_i > 1.0 and i >= no_cfg_until_timestep:
                    noise_pred_uncond = self.transformer(
                        hidden_states=latents.to(
                            device=self.transformer.device, dtype=self.transformer.dtype
                        ),
                        # YiYi notes: divide it by 1000 for now because we scale it by 1000 in the transforme rmodel (we should not keep it but I want to keep the inputs same for the model for testing)
                        timestep=timestep / 1000,
                        guidance=guidance,
                        pooled_projections=negative_pooled_prompt_embeds.to(
                            device=self.transformer.device, dtype=self.transformer.dtype
                        ),
                        encoder_hidden_states=negative_prompt_embeds.to(
                            device=self.transformer.device, dtype=self.transformer.dtype
                        ),
                        txt_ids=negative_text_ids.to(device=self.transformer.device),
                        img_ids=latent_image_ids.to(device=self.transformer.device),
                        joint_attention_kwargs=self.joint_attention_kwargs,
                        return_dict=False,
                        condition_hidden_states=torch.zeros_like(latents_cond).to(
                            device=self.transformer.device, dtype=self.transformer.dtype
                        ),
                    )[0]
                    noise_pred_uncond_t = self.transformer(
                        hidden_states=latents.to(
                            device=self.transformer.device, dtype=self.transformer.dtype
                        ),
                        # YiYi notes: divide it by 1000 for now because we scale it by 1000 in the transforme rmodel (we should not keep it but I want to keep the inputs same for the model for testing)
                        timestep=timestep / 1000,
                        guidance=guidance,
                        pooled_projections=negative_pooled_prompt_embeds.to(
                            device=self.transformer.device, dtype=self.transformer.dtype
                        ),
                        encoder_hidden_states=negative_prompt_embeds.to(
                            device=self.transformer.device, dtype=self.transformer.dtype
                        ),
                        txt_ids=negative_text_ids.to(device=self.transformer.device),
                        img_ids=latent_image_ids.to(device=self.transformer.device),
                        joint_attention_kwargs=self.joint_attention_kwargs,
                        return_dict=False,
                        condition_hidden_states=latents_cond.to(
                            device=self.transformer.device, dtype=self.transformer.dtype
                        ),
                    )[0]

                    # noise_pred = noise_pred_uncond + guidance_scale_real * (
                    #     noise_pred - noise_pred_uncond
                    # )
                    noise_pred = noise_pred_uncond + \
                        guidance_scale_real_i * (noise_pred_uncond_t - noise_pred_uncond) + \
                        guidance_scale_real_t * (noise_pred - noise_pred_uncond_t)

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

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

                if callback_on_step_end is not None:
                    callback_kwargs = {}
                    for k in callback_on_step_end_tensor_inputs:
                        callback_kwargs[k] = locals()[k]
                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)

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

                # call the callback, if provided
                if i == len(timesteps) - 1 or (
                    (i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0
                ):
                    progress_bar.update()

                if XLA_AVAILABLE:
                    xm.mark_step()

        if output_type == "latent":
            image = latents

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

            image = self.vae.decode(
                latents.to(device=self.vae.device, dtype=self.vae.dtype),
                return_dict=False,
            )[0]
            if cut_output:
                image = image[..., 512:]
            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)


from dataclasses import dataclass
from typing import List, Union
import PIL.Image
from diffusers.utils import BaseOutput


@dataclass
class FluxPipelineOutput(BaseOutput):
    """
    Output class for Stable Diffusion pipelines.

    Args:
        images (`List[PIL.Image.Image]` or `np.ndarray`)
            List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width,
            num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline.
    """

    images: Union[List[PIL.Image.Image], np.ndarray]