pipeline_gligen_sdxl.py

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

import PIL.Image
import torch

from transformers import (
    CLIPImageProcessor,
    CLIPTextModel,
    CLIPTextModelWithProjection,
    CLIPTokenizer,
    CLIPVisionModelWithProjection,
)
from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
from diffusers.loaders import (
    FromSingleFileMixin,
    IPAdapterMixin,
    StableDiffusionXLLoraLoaderMixin,
    TextualInversionLoaderMixin,
)
from diffusers.models import (
    AutoencoderKL, 
    ImageProjection,
    UNet2DConditionModel
)
from diffusers.models.attention import GatedSelfAttentionDense
from diffusers.models.attention_processor import (
    AttnProcessor2_0,
    FusedAttnProcessor2_0,
    XFormersAttnProcessor,
)
from diffusers.models.lora import adjust_lora_scale_text_encoder
from diffusers.pipelines.pipeline_utils import (
    DiffusionPipeline,
    StableDiffusionMixin
)
from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl import rescale_noise_cfg, retrieve_timesteps
from diffusers.pipelines.stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
from diffusers.schedulers import (
    KarrasDiffusionSchedulers
)
from diffusers.utils import (
    USE_PEFT_BACKEND,
    deprecate,
    is_invisible_watermark_available,
    is_torch_xla_available,
    logging,
    replace_example_docstring,
    scale_lora_layers,
    unscale_lora_layers,
)
from diffusers.utils.torch_utils import randn_tensor


if is_invisible_watermark_available():
    from diffusers.pipelines.stable_diffusion_xl.watermark import StableDiffusionXLWatermarker
    
if is_torch_xla_available():
    import torch_xla.core.xla_model as xm # type: ignore

    XLA_AVAILABLE = True
else:
    XLA_AVAILABLE = False

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

EXAMPLE_DOC_STRING = """
    Examples:
        ```py
        >>> import torch
        >>> from pipeline_gligen_sdxl import StableDiffusionXLGLIGENPipeline

        >>> pipe = StableDiffusionXLGLIGENPipeline.from_pretrained(
        ...     "xxx", torch_dtype=torch.float16
        ... )
        >>> pipe = pipe.to("cuda")
        
        >>> prompt = "a waterfall and a modern high speed train running through the tunnel in a beautiful forest with fall foliage"
        >>> boxes = [[0.1387, 0.2051, 0.4277, 0.7090], [0.4980, 0.4355, 0.8516, 0.7266]]
        >>> phrases = ["a waterfall", "a modern high speed train running through the tunnel"]

        >>> images = pipe(
        ...     prompt=prompt,
        ...     gligen_phrases=phrases,
        ...     gligen_boxes=boxes,
        ...     gligen_scheduled_sampling_beta=1,
        ...     output_type="pil",
        ...     num_inference_steps=50,
        ... ).images

        >>> images[0].save("./gligen-xl-generation-text-box.jpg")
        ```
"""  

class StableDiffusionXLGLIGENPipeline(
    DiffusionPipeline,
    StableDiffusionMixin,
    FromSingleFileMixin,
    StableDiffusionXLLoraLoaderMixin,
    TextualInversionLoaderMixin,
    IPAdapterMixin,
):
    r"""
    Pipeline for GLIGEN layout text-to-image generation using Stable Diffusion XL.
    """
    
    model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae"
    _optional_components = [
        "tokenizer",
        "tokenizer_2",
        "text_encoder",
        "text_encoder_2",
        "image_encoder",
        "feature_extractor",
    ]
    _callback_tensor_inputs = [
        "latents",
        "prompt_embeds",
        "negative_prompt_embeds",
        "add_text_embeds",
        "add_time_ids",
        "negative_pooled_prompt_embeds",
        "negative_add_time_ids",
    ]
    
    
    def __init__(
        self,
        vae: AutoencoderKL,
        text_encoder: CLIPTextModel,
        text_encoder_2: CLIPTextModelWithProjection,
        tokenizer: CLIPTokenizer,
        tokenizer_2: CLIPTokenizer,
        unet: UNet2DConditionModel,
        scheduler: KarrasDiffusionSchedulers,
        image_encoder: CLIPVisionModelWithProjection = None,
        feature_extractor: CLIPImageProcessor = None,
        force_zeros_for_empty_prompt: bool = True,
        add_watermarker: Optional[bool] = None,
    ):
        super().__init__()

        self.register_modules(
            vae=vae,
            text_encoder=text_encoder,
            text_encoder_2=text_encoder_2,
            tokenizer=tokenizer,
            tokenizer_2=tokenizer_2,
            unet=unet,
            scheduler=scheduler,
            image_encoder=image_encoder,
            feature_extractor=feature_extractor,
        )
        self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)

        self.default_sample_size = self.unet.config.sample_size

        add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available()

        if add_watermarker:
            self.watermark = StableDiffusionXLWatermarker()
        else:
            self.watermark = None
    # copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionPipelineXL.encode_prompt       
    def encode_prompt(
        self,
        prompt: str,
        prompt_2: Optional[str] = None,
        device: Optional[torch.device] = None,
        num_images_per_prompt: int = 1,
        do_classifier_free_guidance: bool = True,
        negative_prompt: Optional[str] = None,
        negative_prompt_2: Optional[str] = None,
        prompt_embeds: Optional[torch.Tensor] = None,
        negative_prompt_embeds: Optional[torch.Tensor] = None,
        pooled_prompt_embeds: Optional[torch.Tensor] = None,
        negative_pooled_prompt_embeds: Optional[torch.Tensor] = None,
        lora_scale: Optional[float] = None,
        clip_skip: Optional[int] = None,
    ):
        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, StableDiffusionXLLoraLoaderMixin):
            self._lora_scale = lora_scale

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

            if self.text_encoder_2 is not None:
                if not USE_PEFT_BACKEND:
                    adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale)
                else:
                    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]

        # Define tokenizers and text encoders
        tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2]
        text_encoders = (
            [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
        )

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

            # textual inversion: process multi-vector tokens if necessary
            prompt_embeds_list = []
            prompts = [prompt, prompt_2]
            for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders):
                if isinstance(self, TextualInversionLoaderMixin):
                    prompt = self.maybe_convert_prompt(prompt, tokenizer)

                text_inputs = tokenizer(
                    prompt,
                    padding="max_length",
                    max_length=tokenizer.model_max_length,
                    truncation=True,
                    return_tensors="pt",
                )

                text_input_ids = text_inputs.input_ids
                untruncated_ids = 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 = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1])
                    logger.warning(
                        "The following part of your input was truncated because CLIP can only handle sequences up to"
                        f" {tokenizer.model_max_length} tokens: {removed_text}"
                    )

                prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
                
                # We are only ALWAYS interested in the pooled output of the final text encoder
                pooled_prompt_embeds = prompt_embeds[0]
                if clip_skip is None:
                    prompt_embeds = prompt_embeds.hidden_states[-2]
                else:
                    # "2" because SDXL always indexes from the penultimate layer.
                    prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)]

                prompt_embeds_list.append(prompt_embeds)
                
            prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)

        # get unconditional embeddings for classifier free guidance
        zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt
        if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt:
            negative_prompt_embeds = torch.zeros_like(prompt_embeds)
            negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
        elif do_classifier_free_guidance and negative_prompt_embeds is None:
            negative_prompt = negative_prompt or ""
            negative_prompt_2 = negative_prompt_2 or negative_prompt

            # normalize str to list
            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
            negative_prompt_2 = (
                batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2
            )

            uncond_tokens: List[str]
            if prompt is not None and type(prompt) is not type(negative_prompt):
                raise TypeError(
                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
                    f" {type(prompt)}."
                )
            elif batch_size != len(negative_prompt):
                raise ValueError(
                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
                    " the batch size of `prompt`."
                )
            else:
                uncond_tokens = [negative_prompt, negative_prompt_2]

            negative_prompt_embeds_list = []
            for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders):
                if isinstance(self, TextualInversionLoaderMixin):
                    negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer)

                max_length = prompt_embeds.shape[1]
                uncond_input = tokenizer(
                    negative_prompt,
                    padding="max_length",
                    max_length=max_length,
                    truncation=True,
                    return_tensors="pt",
                )

                negative_prompt_embeds = text_encoder(
                    uncond_input.input_ids.to(device),
                    output_hidden_states=True,
                )
                # We are only ALWAYS interested in the pooled output of the final text encoder
                negative_pooled_prompt_embeds = negative_prompt_embeds[0]
                negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]

                negative_prompt_embeds_list.append(negative_prompt_embeds)

            negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)

        if self.text_encoder_2 is not None:
            prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
        else:
            prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device)

        bs_embed, seq_len, _ = prompt_embeds.shape
        # duplicate text embeddings for each generation per prompt, using mps friendly method
        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
        prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
        
        if do_classifier_free_guidance:
            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
            seq_len = negative_prompt_embeds.shape[1]

            if self.text_encoder_2 is not None:
                negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
            else:
                negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device)

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

        pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
            bs_embed * num_images_per_prompt, -1
        )
        if do_classifier_free_guidance:
            negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
                bs_embed * num_images_per_prompt, -1
            )

        if self.text_encoder is not None:
            if isinstance(self, StableDiffusionXLLoraLoaderMixin) 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, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
                # Retrieve the original scale by scaling back the LoRA layers
                unscale_lora_layers(self.text_encoder_2, lora_scale)
                
        return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
    def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
        dtype = next(self.image_encoder.parameters()).dtype

        if not isinstance(image, torch.Tensor):
            image = self.feature_extractor(image, return_tensors="pt").pixel_values

        image = image.to(device=device, dtype=dtype)
        if output_hidden_states:
            image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
            image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
            uncond_image_enc_hidden_states = self.image_encoder(
                torch.zeros_like(image), output_hidden_states=True
            ).hidden_states[-2]
            uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
                num_images_per_prompt, dim=0
            )
            return image_enc_hidden_states, uncond_image_enc_hidden_states
        else:
            image_embeds = self.image_encoder(image).image_embeds
            image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
            uncond_image_embeds = torch.zeros_like(image_embeds)

            return image_embeds, uncond_image_embeds
        
    def encode_prompt_gligen(
        self,
        prompt: str,
        prompt_2: Optional[str] = None,
        device: Optional[torch.device] = None,
        num_images_per_prompt: int = 1,
        gligen_embeds: Optional[torch.Tensor] = None,
        lora_scale: Optional[float] = None,
        clip_skip: Optional[int] = None,
    ):
        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, StableDiffusionXLLoraLoaderMixin):
            self._lora_scale = lora_scale

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

            if self.text_encoder_2 is not None:
                if not USE_PEFT_BACKEND:
                    adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale)
                else:
                    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]

        # Define tokenizers and text encoders
        tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2]
        text_encoders = (
            [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
        )

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

            # textual inversion: process multi-vector tokens if necessary
            gligen_embeds_list = []
            prompts = [prompt, prompt_2]
            for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders):
                if isinstance(self, TextualInversionLoaderMixin):
                    prompt = self.maybe_convert_prompt(prompt, tokenizer)

                text_inputs = tokenizer(
                    prompt,
                    padding="max_length",
                    max_length=tokenizer.model_max_length,
                    truncation=True,
                    return_tensors="pt",
                )

                text_input_ids = text_inputs.input_ids
                untruncated_ids = 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 = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1])
                    logger.warning(
                        "The following part of your input was truncated because CLIP can only handle sequences up to"
                        f" {tokenizer.model_max_length} tokens: {removed_text}"
                    )

                prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=False)
                
                if isinstance(text_encoder, CLIPTextModel):
                    gligen_embeds_list.append(prompt_embeds.pooler_output)
                elif isinstance(text_encoder, CLIPTextModelWithProjection):
                    gligen_embeds_list.append(prompt_embeds.text_embeds)
        
        gligen_embeds = torch.concat(gligen_embeds_list, dim=-1)

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

        return gligen_embeds
        
    # Copied from SDXL
    def prepare_ip_adapter_image_embeds(
        self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
    ):
        image_embeds = []
        if do_classifier_free_guidance:
            negative_image_embeds = []
        if ip_adapter_image_embeds is None:
            if not isinstance(ip_adapter_image, list):
                ip_adapter_image = [ip_adapter_image]

            if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
                raise ValueError(
                    f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
                )

            for single_ip_adapter_image, image_proj_layer in zip(
                ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
            ):
                output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
                single_image_embeds, single_negative_image_embeds = self.encode_image(
                    single_ip_adapter_image, device, 1, output_hidden_state
                )

                image_embeds.append(single_image_embeds[None, :])
                if do_classifier_free_guidance:
                    negative_image_embeds.append(single_negative_image_embeds[None, :])
        else:
            for single_image_embeds in ip_adapter_image_embeds:
                if do_classifier_free_guidance:
                    single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
                    negative_image_embeds.append(single_negative_image_embeds)
                image_embeds.append(single_image_embeds)

        ip_adapter_image_embeds = []
        for i, single_image_embeds in enumerate(image_embeds):
            single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0)
            if do_classifier_free_guidance:
                single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0)
                single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0)

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

        return ip_adapter_image_embeds
    
    # Copied form SDXL
    def prepare_extra_step_kwargs(self, generator, eta):
        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
        # and should be between [0, 1]

        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
        extra_step_kwargs = {}
        if accepts_eta:
            extra_step_kwargs["eta"] = eta

        # check if the scheduler accepts generator
        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
        if accepts_generator:
            extra_step_kwargs["generator"] = generator
        return extra_step_kwargs
    
    # Copied from SDXL and StableDiffusionGLIGENPipeline
    def check_inputs(
        self,
        prompt,
        prompt_2,
        height,
        width,
        callback_steps,
        gligen_phrases,
        gligen_boxes,
        negative_prompt=None,
        negative_prompt_2=None,
        prompt_embeds=None,
        negative_prompt_embeds=None,
        pooled_prompt_embeds=None,
        negative_pooled_prompt_embeds=None,
        ip_adapter_image=None,
        ip_adapter_image_embeds=None,
        callback_on_step_end_tensor_inputs=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_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
            raise ValueError(
                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
                f" {type(callback_steps)}."
            )

        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 negative_prompt is not None and negative_prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
            )
        elif negative_prompt_2 is not None and negative_prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:"
                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
            )

        if prompt_embeds is not None and negative_prompt_embeds is not None:
            if prompt_embeds.shape != negative_prompt_embeds.shape:
                raise ValueError(
                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
                    f" {negative_prompt_embeds.shape}."
                )

        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 negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None:
            raise ValueError(
                "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`."
            )

        if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
            raise ValueError(
                "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
            )

        if ip_adapter_image_embeds is not None:
            if not isinstance(ip_adapter_image_embeds, list):
                raise ValueError(
                    f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
                )
            elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
                raise ValueError(
                    f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
                )
                
        if len(gligen_phrases) != len(gligen_boxes):
            raise ValueError(
                "length of `gligen_phrases` and `gligen_boxes` has to be same, but"
                f" got: `gligen_phrases` {len(gligen_phrases)} != `gligen_boxes` {len(gligen_boxes)}"
            )
            
    # Copied from SDXL
    def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
        shape = (
            batch_size,
            num_channels_latents,
            int(height) // self.vae_scale_factor,
            int(width) // self.vae_scale_factor,
        )
        if isinstance(generator, list) and len(generator) != batch_size:
            raise ValueError(
                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
            )

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

        # scale the initial noise by the standard deviation required by the scheduler
        latents = latents * self.scheduler.init_noise_sigma
        return latents
    
    # Copied from SDGligenPipeline
    def enable_fuser(self, enabled=True):
        for module in self.unet.modules():
            if type(module) is GatedSelfAttentionDense:
                module.enabled = enabled
    
    # Copied from SDGligenPipeline            
    def draw_inpaint_mask_from_boxes(self, boxes, size):
        inpaint_mask = torch.ones(size[0], size[1])
        for box in boxes:
            x0, x1 = box[0] * size[0], box[2] * size[0]
            y0, y1 = box[1] * size[1], box[3] * size[1]
            inpaint_mask[int(y0) : int(y1), int(x0) : int(x1)] = 0
        return inpaint_mask
    
    # Copied from SDGligenPipeline            
    def crop(self, im, new_width, new_height):
        width, height = im.size
        left = (width - new_width) / 2
        top = (height - new_height) / 2
        right = (width + new_width) / 2
        bottom = (height + new_height) / 2
        return im.crop((left, top, right, bottom))
    
    # Copied from SDGligenPipeline            
    def target_size_center_crop(self, im, new_hw):
        width, height = im.size
        if width != height:
            im = self.crop(im, min(height, width), min(height, width))
        return im.resize((new_hw, new_hw), PIL.Image.LANCZOS)
    
    # Copied from SDXL
    def _get_add_time_ids(
        self, original_size, crops_coords_top_left, target_size, dtype, text_encoder_projection_dim=None
    ):
        add_time_ids = list(original_size + crops_coords_top_left + target_size)

        passed_add_embed_dim = (
            self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim
        )
        expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features

        if expected_add_embed_dim != passed_add_embed_dim:
            raise ValueError(
                f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."
            )

        add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
        return add_time_ids
    
    # Copied from SDXL
    def upcast_vae(self):
        dtype = self.vae.dtype
        self.vae.to(dtype=torch.float32)
        use_torch_2_0_or_xformers = isinstance(
            self.vae.decoder.mid_block.attentions[0].processor,
            (
                AttnProcessor2_0,
                XFormersAttnProcessor,
                FusedAttnProcessor2_0,
            ),
        )
        # if xformers or torch_2_0 is used attention block does not need
        # to be in float32 which can save lots of memory
        if use_torch_2_0_or_xformers:
            self.vae.post_quant_conv.to(dtype)
            self.vae.decoder.conv_in.to(dtype)
            self.vae.decoder.mid_block.to(dtype)
            
    # Copied from SDXL
    def get_guidance_scale_embedding(
        self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32
    ) -> torch.Tensor:
        """
        See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298

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

        Returns:
            `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`.
        """
        assert len(w.shape) == 1
        w = w * 1000.0

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

    @property
    def guidance_rescale(self):
        return self._guidance_rescale

    @property
    def clip_skip(self):
        return self._clip_skip

    # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
    # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
    # corresponds to doing no classifier free guidance.
    @property
    def do_classifier_free_guidance(self):
        return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None

    @property
    def cross_attention_kwargs(self):
        return self._cross_attention_kwargs

    @property
    def denoising_end(self):
        return self._denoising_end

    @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,
        height: Optional[int] = None,
        width: Optional[int] = None,
        num_inference_steps: int = 50,
        timesteps: List[int] = None,
        sigmas: List[float] = None,
        denoising_end: Optional[float] = None,
        guidance_scale: float = 5.0,
        gligen_scheduled_sampling_beta: float = 0.3,
        gligen_phrases: List[str] = None,
        gligen_boxes: List[List[float]] = None,
        gligen_inpaint_image: Optional[PIL.Image.Image] = None,
        negative_prompt: Optional[Union[str, List[str]]] = None,
        negative_prompt_2: Optional[Union[str, List[str]]] = None,
        num_images_per_prompt: Optional[int] = 1,
        eta: float = 0.0,
        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
        latents: Optional[torch.Tensor] = None,
        prompt_embeds: Optional[torch.Tensor] = None,
        negative_prompt_embeds: Optional[torch.Tensor] = None,
        pooled_prompt_embeds: Optional[torch.Tensor] = None,
        negative_pooled_prompt_embeds: Optional[torch.Tensor] = None,
        ip_adapter_image: Optional[PipelineImageInput] = None,
        ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None,
        output_type: Optional[str] = "pil",
        return_dict: bool = True,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        guidance_rescale: float = 0.0,
        original_size: Optional[Tuple[int, int]] = None,
        crops_coords_top_left: Tuple[int, int] = (0, 0),
        target_size: Optional[Tuple[int, int]] = None,
        negative_original_size: Optional[Tuple[int, int]] = None,
        negative_crops_coords_top_left: Tuple[int, int] = (0, 0),
        negative_target_size: Optional[Tuple[int, int]] = None,
        clip_skip: Optional[int] = None,
        callback_on_step_end: Optional[
            Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
        ] = None,
        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
        **kwargs,
    ):
        r"""
        Function invoked when calling the pipeline for generation.

        Args:
            prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
                instead.
            prompt_2 (`str` or `List[str]`, *optional*):
                The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
                used in both text-encoders
            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.
                Anything below 512 pixels won't work well for
                [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
                and checkpoints that are not specifically fine-tuned on low resolutions.
            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.
                Anything below 512 pixels won't work well for
                [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
                and checkpoints that are not specifically fine-tuned on low resolutions.
            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.
            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.
            denoising_end (`float`, *optional*):
                When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be
                completed before it is intentionally prematurely terminated. As a result, the returned sample will
                still retain a substantial amount of noise as determined by the discrete timesteps selected by the
                scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a
                "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image
                Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output)
            guidance_scale (`float`, *optional*, defaults to 5.0):
                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.
            gligen_phrases (`List[str]`):
                The phrases to guide what to include in each of the regions defined by the corresponding
                `gligen_boxes`. There should only be one phrase per bounding box.
            gligen_boxes (`List[List[float]]`):
                The bounding boxes that identify rectangular regions of the image that are going to be filled with the
                content described by the corresponding `gligen_phrases`. Each rectangular box is defined as a
                `List[float]` of 4 elements `[xmin, ymin, xmax, ymax]` where each value is between [0,1].
            gligen_inpaint_image (`PIL.Image.Image`, *optional*):
                The input image, if provided, is inpainted with objects described by the `gligen_boxes` and
                `gligen_phrases`. Otherwise, it is treated as a generation task on a blank input image.
            gligen_scheduled_sampling_beta (`float`, defaults to 0.3):
                Scheduled Sampling factor from [GLIGEN: Open-Set Grounded Text-to-Image
                Generation](https://arxiv.org/pdf/2301.07093.pdf). Scheduled Sampling factor is only varied for
                scheduled sampling during inference for improved quality and controllability.
            negative_prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation. If not defined, one has to pass
                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
                less than `1`).
            negative_prompt_2 (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
                `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
            num_images_per_prompt (`int`, *optional*, defaults to 1):
                The number of images to generate per prompt.
            eta (`float`, *optional*, defaults to 0.0):
                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
                [`schedulers.DDIMScheduler`], will be ignored for others.
            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.Tensor`, *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.Tensor`, *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.
            negative_prompt_embeds (`torch.Tensor`, *optional*):
                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
                argument.
            pooled_prompt_embeds (`torch.Tensor`, *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.
            negative_pooled_prompt_embeds (`torch.Tensor`, *optional*):
                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
                input argument.
            ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
            ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*):
                Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of
                IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should
                contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not
                provided, embeddings are computed from the `ip_adapter_image` input argument.
            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.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead
                of a plain tuple.
            cross_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).
            guidance_rescale (`float`, *optional*, defaults to 0.0):
                Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are
                Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of
                [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf).
                Guidance rescale factor should fix overexposure when using zero terminal SNR.
            original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
                If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled.
                `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as
                explained in section 2.2 of
                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
            crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
                `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
                `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
                `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
            target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
                For most cases, `target_size` should be set to the desired height and width of the generated image. If
                not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in
                section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
            negative_original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
                To negatively condition the generation process based on a specific image resolution. Part of SDXL's
                micro-conditioning as explained in section 2.2 of
                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
            negative_crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
                To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's
                micro-conditioning as explained in section 2.2 of
                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
            negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
                To negatively condition the generation process based on a target image resolution. It should be as same
                as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of
                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
            callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*):
                A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of
                each denoising step during the inference. with the following arguments: `callback_on_step_end(self:
                DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a
                list of all tensors as specified by `callback_on_step_end_tensor_inputs`.
            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.

        Examples:

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

        callback = kwargs.pop("callback", None)
        callback_steps = kwargs.pop("callback_steps", None)

        if callback is not None:
            deprecate(
                "callback",
                "1.0.0",
                "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
            )
        if callback_steps is not None:
            deprecate(
                "callback_steps",
                "1.0.0",
                "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
            )

        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs

        # 0. Default height and width to unet
        height = height or self.default_sample_size * self.vae_scale_factor
        width = width or self.default_sample_size * self.vae_scale_factor

        original_size = original_size or (height, width)
        target_size = target_size or (height, width)

        # 1. Check inputs. Raise error if not correct
        self.check_inputs(
            prompt,
            prompt_2,
            height,
            width,
            callback_steps,
            gligen_phrases,
            gligen_boxes,
            negative_prompt,
            negative_prompt_2,
            prompt_embeds,
            negative_prompt_embeds,
            pooled_prompt_embeds,
            negative_pooled_prompt_embeds,
            ip_adapter_image,
            ip_adapter_image_embeds,
            callback_on_step_end_tensor_inputs,
        )

        self._guidance_scale = guidance_scale
        self._guidance_rescale = guidance_rescale
        self._clip_skip = clip_skip
        self._cross_attention_kwargs = cross_attention_kwargs
        self._denoising_end = denoising_end
        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

        # 3. Encode input prompt
        lora_scale = (
            self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
        )

        (
            prompt_embeds,
            negative_prompt_embeds,
            pooled_prompt_embeds,
            negative_pooled_prompt_embeds,
        ) = self.encode_prompt(
            prompt=prompt,
            prompt_2=prompt_2,
            device=device,
            num_images_per_prompt=num_images_per_prompt,
            do_classifier_free_guidance=self.do_classifier_free_guidance,
            negative_prompt=negative_prompt,
            negative_prompt_2=negative_prompt_2,
            prompt_embeds=prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds,
            pooled_prompt_embeds=pooled_prompt_embeds,
            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
            lora_scale=lora_scale,
            clip_skip=self.clip_skip,
        )

        # 4. Prepare timesteps
        timesteps, num_inference_steps = retrieve_timesteps(
            self.scheduler, num_inference_steps, device, timesteps, sigmas
        )

        # 5. Prepare latent variables
        num_channels_latents = self.unet.config.in_channels
        latents = self.prepare_latents(
            batch_size * num_images_per_prompt,
            num_channels_latents,
            height,
            width,
            prompt_embeds.dtype,
            device,
            generator,
            latents,
        )
        # 5.1 Prepare GLIGEN variables
        max_objs = 30
        if len(gligen_boxes) > max_objs:
            warnings.warn(
                f"More that {max_objs} objects found. Only first {max_objs} objects will be processed.",
                FutureWarning,
            )
            gligen_phrases = gligen_phrases[:max_objs]
            gligen_boxes = gligen_boxes[:max_objs]
        # prepare batched input to the GLIGENTextBoundingboxProjection (boxes, phrases, mask)
        # obtain its text features for phrases
        (
            gligen_embeds
        ) = self.encode_prompt_gligen(
            prompt=gligen_phrases,
            device=device,
            num_images_per_prompt=1,
            # TODO: whether we had to follow prompt encoding configuration on LoRA and CLIP skip
            lora_scale=lora_scale,
            clip_skip=self.clip_skip,
        )
        
        n_objs = len(gligen_boxes)
        # For each entity, described in phrases, is denoted with a bounding box,
        # we represent the location information as (xmin,ymin,xmax,ymax)
        boxes = torch.zeros(max_objs, 4, device=device, dtype=self.text_encoder.dtype)
        boxes[:n_objs] = torch.tensor(gligen_boxes)
        text_embeddings = torch.zeros(
            max_objs, self.unet.config.cross_attention_dim, device=device, dtype=self.text_encoder.dtype
        )
        text_embeddings[:n_objs] = gligen_embeds
        # Generate a mask for each object that is entity described by phrases
        masks = torch.zeros(max_objs, device=device, dtype=self.text_encoder.dtype)
        masks[:n_objs] = 1
        repeat_batch = batch_size * num_images_per_prompt
        boxes = boxes.unsqueeze(0).expand(repeat_batch, -1, -1).clone()
        text_embeddings = text_embeddings.unsqueeze(0).expand(repeat_batch, -1, -1).clone()
        masks = masks.unsqueeze(0).expand(repeat_batch, -1).clone()
        if self.do_classifier_free_guidance:
            repeat_batch = repeat_batch * 2
            boxes = torch.cat([boxes] * 2)
            text_embeddings = torch.cat([text_embeddings] * 2)
            masks = torch.cat([masks] * 2)
            masks[: repeat_batch // 2] = 0
        if self.cross_attention_kwargs is None:
            self._cross_attention_kwargs = {}
        self.cross_attention_kwargs["gligen"] = {"boxes": boxes, "positive_embeddings": text_embeddings, "masks": masks}
        
         # Prepare latent variables for GLIGEN inpainting
        if gligen_inpaint_image is not None:
            # if the given input image is not of the same size as expected by VAE
            # center crop and resize the input image to expected shape
            if gligen_inpaint_image.size != (self.vae.sample_size, self.vae.sample_size):
                gligen_inpaint_image = self.target_size_center_crop(gligen_inpaint_image, self.vae.sample_size)
            # Convert a single image into a batch of images with a batch size of 1
            # The resulting shape becomes (1, C, H, W), where C is the number of channels,
            # and H and W are the height and width of the image.
            # scales the pixel values to a range [-1, 1]
            gligen_inpaint_image = self.image_processor.preprocess(gligen_inpaint_image)
            gligen_inpaint_image = gligen_inpaint_image.to(dtype=self.vae.dtype, device=self.vae.device)
            # Run AutoEncoder to get corresponding latents
            gligen_inpaint_latent = self.vae.encode(gligen_inpaint_image).latent_dist.sample()
            gligen_inpaint_latent = self.vae.config.scaling_factor * gligen_inpaint_latent
            # Generate an inpainting mask
            # pixel value = 0, where the object is present (defined by bounding boxes above)
            #               1, everywhere else
            gligen_inpaint_mask = self.draw_inpaint_mask_from_boxes(gligen_boxes, gligen_inpaint_latent.shape[2:])
            gligen_inpaint_mask = gligen_inpaint_mask.to(
                dtype=gligen_inpaint_latent.dtype, device=gligen_inpaint_latent.device
            )
            gligen_inpaint_mask = gligen_inpaint_mask[None, None]
            gligen_inpaint_mask_addition = torch.cat(
                (gligen_inpaint_latent * gligen_inpaint_mask, gligen_inpaint_mask), dim=1
            )
            # Convert a single mask into a batch of masks with a batch size of 1
            gligen_inpaint_mask_addition = gligen_inpaint_mask_addition.expand(repeat_batch, -1, -1, -1).clone()
        
        num_grounding_steps = int(gligen_scheduled_sampling_beta * len(timesteps))
        self.enable_fuser(True)
        
        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)

        # 7. Prepare added time ids & embeddings
        add_text_embeds = pooled_prompt_embeds
        if self.text_encoder_2 is None:
            text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
        else:
            text_encoder_projection_dim = self.text_encoder_2.config.projection_dim

        add_time_ids = self._get_add_time_ids(
            original_size,
            crops_coords_top_left,
            target_size,
            dtype=prompt_embeds.dtype,
            text_encoder_projection_dim=text_encoder_projection_dim,
        )
        if negative_original_size is not None and negative_target_size is not None:
            negative_add_time_ids = self._get_add_time_ids(
                negative_original_size,
                negative_crops_coords_top_left,
                negative_target_size,
                dtype=prompt_embeds.dtype,
                text_encoder_projection_dim=text_encoder_projection_dim,
            )
        else:
            negative_add_time_ids = add_time_ids

        if self.do_classifier_free_guidance:
            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
            add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
            add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0)

        prompt_embeds = prompt_embeds.to(device)
        add_text_embeds = add_text_embeds.to(device)
        add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)

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

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

        # 8.1 Apply denoising_end
        if (
            self.denoising_end is not None
            and isinstance(self.denoising_end, float)
            and self.denoising_end > 0
            and self.denoising_end < 1
        ):
            discrete_timestep_cutoff = int(
                round(
                    self.scheduler.config.num_train_timesteps
                    - (self.denoising_end * self.scheduler.config.num_train_timesteps)
                )
            )
            num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps)))
            timesteps = timesteps[:num_inference_steps]

        # 9. Optionally get Guidance Scale Embedding
        timestep_cond = None
        if self.unet.config.time_cond_proj_dim is not None:
            guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
            timestep_cond = self.get_guidance_scale_embedding(
                guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
            ).to(device=device, dtype=latents.dtype)

        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
                
                if i == num_grounding_steps:
                    self.enable_fuser(False)
                    
                if gligen_inpaint_image is not None:
                    gligen_inpaint_latent_with_noise = (
                        self.scheduler.add_noise(
                            gligen_inpaint_latent, torch.randn_like(gligen_inpaint_latent), torch.tensor([t])
                        )
                        .expand(latents.shape[0], -1, -1, -1)
                        .clone()
                    )
                    latents = gligen_inpaint_latent_with_noise * gligen_inpaint_mask + latents * (
                        1 - gligen_inpaint_mask
                    )

                # expand the latents if we are doing classifier free guidance
                latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents

                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
                
                if gligen_inpaint_image is not None:
                    latent_model_input = torch.cat((latent_model_input, gligen_inpaint_mask_addition), dim=1)

                # predict the noise residual
                added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
                if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
                    added_cond_kwargs["image_embeds"] = image_embeds
                noise_pred = self.unet(
                    latent_model_input,
                    t,
                    encoder_hidden_states=prompt_embeds,
                    timestep_cond=timestep_cond,
                    cross_attention_kwargs=self.cross_attention_kwargs,
                    added_cond_kwargs=added_cond_kwargs,
                    return_dict=False,
                )[0]

                # perform guidance
                if self.do_classifier_free_guidance:
                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)

                if self.do_classifier_free_guidance and self.guidance_rescale > 0.0:
                    # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
                    noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale)

                # compute the previous noisy sample x_t -> x_t-1
                latents_dtype = latents.dtype
                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, 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)
                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
                    add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds)
                    negative_pooled_prompt_embeds = callback_outputs.pop(
                        "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds
                    )
                    add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids)
                    negative_add_time_ids = callback_outputs.pop("negative_add_time_ids", negative_add_time_ids)

                # call the callback, if provided
                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                    progress_bar.update()
                    if callback is not None and i % callback_steps == 0:
                        step_idx = i // getattr(self.scheduler, "order", 1)
                        callback(step_idx, t, latents)

                if XLA_AVAILABLE:
                    xm.mark_step()

        if not output_type == "latent":
            # make sure the VAE is in float32 mode, as it overflows in float16
            needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast

            if needs_upcasting:
                self.upcast_vae()
                latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
            elif latents.dtype != self.vae.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
                    self.vae = self.vae.to(latents.dtype)

            # unscale/denormalize the latents
            # denormalize with the mean and std if available and not None
            has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None
            has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None
            if has_latents_mean and has_latents_std:
                latents_mean = (
                    torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype)
                )
                latents_std = (
                    torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype)
                )
                latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean
            else:
                latents = latents / self.vae.config.scaling_factor

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

            # cast back to fp16 if needed
            if needs_upcasting:
                self.vae.to(dtype=torch.float16)
        else:
            image = latents

        if not output_type == "latent":
            # apply watermark if available
            if self.watermark is not None:
                image = self.watermark.apply_watermark(image)

            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 StableDiffusionXLPipelineOutput(images=image)