pipelines/pipeline_stable_diffusion_spade_timemoe3.py

# Copyright 2023 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
import warnings
from typing import Any, Callable, Dict, List, Optional, Tuple, Union

import torch
import torch.nn.functional as F
from packaging import version
from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer

from diffusers.configuration_utils import FrozenDict
from diffusers.image_processor import VaeImageProcessor
from diffusers.loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin
from diffusers.models import AutoencoderKL, UNet2DConditionModel
from diffusers.schedulers import KarrasDiffusionSchedulers
from diffusers.utils import (
    deprecate,
    is_accelerate_available,
    is_accelerate_version,
    logging,
    randn_tensor,
    replace_example_docstring,
)
from diffusers.pipelines.pipeline_utils import DiffusionPipeline
from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker


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

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

        >>> pipe = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16)
        >>> pipe = pipe.to("cuda")

        >>> prompt = "a photo of an astronaut riding a horse on mars"
        >>> image = pipe(prompt).images[0]
        ```
"""


def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
    """
    Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
    Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
    """
    std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
    std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
    # rescale the results from guidance (fixes overexposure)
    noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
    # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
    noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
    return noise_cfg


class StableDiffusionPipeline(DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, FromSingleFileMixin):
    r"""
    Pipeline for text-to-image generation using Stable Diffusion.

    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)

    In addition the pipeline inherits the following loading methods:
        - *Textual-Inversion*: [`loaders.TextualInversionLoaderMixin.load_textual_inversion`]
        - *LoRA*: [`loaders.LoraLoaderMixin.load_lora_weights`]
        - *Ckpt*: [`loaders.FromCkptMixin.from_ckpt`]

    as well as the following saving methods:
        - *LoRA*: [`loaders.LoraLoaderMixin.save_lora_weights`]

    Args:
        vae ([`AutoencoderKL`]):
            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
        text_encoder ([`CLIPTextModel`]):
            Frozen text-encoder. Stable Diffusion uses the text portion of
            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
        tokenizer (`CLIPTokenizer`):
            Tokenizer of class
            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
        scheduler ([`SchedulerMixin`]):
            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
        safety_checker ([`StableDiffusionSafetyChecker`]):
            Classification module that estimates whether generated images could be considered offensive or harmful.
            Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
        feature_extractor ([`CLIPImageProcessor`]):
            Model that extracts features from generated images to be used as inputs for the `safety_checker`.
    """
    _optional_components = ["safety_checker", "feature_extractor"]

    def __init__(
        self,
        vae: AutoencoderKL,
        text_encoder: CLIPTextModel,
        tokenizer: CLIPTokenizer,
        unet: UNet2DConditionModel,
        unet2: UNet2DConditionModel,
        unet3: UNet2DConditionModel,
        scheduler: KarrasDiffusionSchedulers,
        safety_checker: StableDiffusionSafetyChecker,
        feature_extractor: CLIPImageProcessor,
        requires_safety_checker: bool = True,
    ):
        super().__init__()

        if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
            deprecation_message = (
                f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
                f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
                "to update the config accordingly as leaving `steps_offset` might led to incorrect results"
                " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
                " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
                " file"
            )
            deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
            new_config = dict(scheduler.config)
            new_config["steps_offset"] = 1
            scheduler._internal_dict = FrozenDict(new_config)

        if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is True:
            deprecation_message = (
                f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`."
                " `clip_sample` should be set to False in the configuration file. Please make sure to update the"
                " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in"
                " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very"
                " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file"
            )
            deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False)
            new_config = dict(scheduler.config)
            new_config["clip_sample"] = False
            scheduler._internal_dict = FrozenDict(new_config)

        if safety_checker is None and requires_safety_checker:
            logger.warning(
                f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
                " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
                " results in services or applications open to the public. Both the diffusers team and Hugging Face"
                " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
                " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
                " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
            )

        if safety_checker is not None and feature_extractor is None:
            raise ValueError(
                "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
                " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
            )

        is_unet_version_less_0_9_0 = hasattr(unet.config, "_diffusers_version") and version.parse(
            version.parse(unet.config._diffusers_version).base_version
        ) < version.parse("0.9.0.dev0")
        is_unet_sample_size_less_64 = hasattr(unet.config, "sample_size") and unet.config.sample_size < 64
        if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64:
            deprecation_message = (
                "The configuration file of the unet has set the default `sample_size` to smaller than"
                " 64 which seems highly unlikely. If your checkpoint is a fine-tuned version of any of the"
                " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-"
                " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5"
                " \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the"
                " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`"
                " in the config might lead to incorrect results in future versions. If you have downloaded this"
                " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for"
                " the `unet/config.json` file"
            )
            deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False)
            new_config = dict(unet.config)
            new_config["sample_size"] = 64
            unet._internal_dict = FrozenDict(new_config)
            new_config = dict(unet2.config)
            new_config["sample_size"] = 64
            unet2._internal_dict = FrozenDict(new_config)
            new_config = dict(unet3.config)
            new_config["sample_size"] = 64
            unet3._internal_dict = FrozenDict(new_config)
            # new_config = dict(unet4.config)
            # new_config["sample_size"] = 64
            # unet4._internal_dict = FrozenDict(new_config)

        self.register_modules(
            vae=vae,
            text_encoder=text_encoder,
            tokenizer=tokenizer,
            unet=unet,
            unet2=unet2,
            unet3=unet3,
            scheduler=scheduler,
            safety_checker=safety_checker,
            feature_extractor=feature_extractor,
        )
        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.register_to_config(requires_safety_checker=requires_safety_checker)

    def enable_vae_slicing(self):
        r"""
        Enable sliced VAE decoding.

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

    def disable_vae_slicing(self):
        r"""
        Disable sliced VAE decoding. If `enable_vae_slicing` was previously invoked, this method will go back to
        computing decoding in one step.
        """
        self.vae.disable_slicing()

    def enable_vae_tiling(self):
        r"""
        Enable tiled VAE decoding.

        When this option is enabled, the VAE will split the input tensor into tiles to compute decoding and encoding in
        several steps. This is useful to save a large amount of memory and to allow the processing of larger images.
        """
        self.vae.enable_tiling()

    def disable_vae_tiling(self):
        r"""
        Disable tiled VAE decoding. If `enable_vae_tiling` was previously invoked, this method will go back to
        computing decoding in one step.
        """
        self.vae.disable_tiling()

    def enable_sequential_cpu_offload(self, gpu_id=0):
        r"""
        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
        text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
        Note that offloading happens on a submodule basis. Memory savings are higher than with
        `enable_model_cpu_offload`, but performance is lower.
        """
        if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
            from accelerate import cpu_offload
        else:
            raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")

        device = torch.device(f"cuda:{gpu_id}")

        if self.device.type != "cpu":
            self.to("cpu", silence_dtype_warnings=True)
            torch.cuda.empty_cache()  # otherwise we don't see the memory savings (but they probably exist)

        for cpu_offloaded_model in [self.unet, self.unet2, self.unet3, self.text_encoder, self.vae]:
            cpu_offload(cpu_offloaded_model, device)

        if self.safety_checker is not None:
            cpu_offload(self.safety_checker, execution_device=device, offload_buffers=True)

    def enable_model_cpu_offload(self, gpu_id=0):
        r"""
        Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
        to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
        method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
        `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
        """
        if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
            from accelerate import cpu_offload_with_hook
        else:
            raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.")

        device = torch.device(f"cuda:{gpu_id}")

        if self.device.type != "cpu":
            self.to("cpu", silence_dtype_warnings=True)
            torch.cuda.empty_cache()  # otherwise we don't see the memory savings (but they probably exist)

        hook = None
        for cpu_offloaded_model in [self.text_encoder, self.unet, self.unet2, self.unet3, self.vae]:
            _, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook)

        if self.safety_checker is not None:
            _, hook = cpu_offload_with_hook(self.safety_checker, device, prev_module_hook=hook)

        # We'll offload the last model manually.
        self.final_offload_hook = hook

    @property
    def _execution_device(self):
        r"""
        Returns the device on which the pipeline's models will be executed. After calling
        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
        hooks.
        """
        if not hasattr(self.unet, "_hf_hook"):
            return self.device
        for module in self.unet.modules():
            if (
                hasattr(module, "_hf_hook")
                and hasattr(module._hf_hook, "execution_device")
                and module._hf_hook.execution_device is not None
            ):
                return torch.device(module._hf_hook.execution_device)
        return self.device

    def _encode_prompt(
        self,
        prompt,
        device,
        num_images_per_prompt,
        do_classifier_free_guidance,
        negative_prompt=None,
        prompt_embeds: Optional[torch.FloatTensor] = None,
        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
        lora_scale: Optional[float] = None,
    ):
        r"""
        Encodes the prompt into text encoder hidden states.

        Args:
             prompt (`str` or `List[str]`, *optional*):
                prompt to be encoded
            device: (`torch.device`):
                torch device
            num_images_per_prompt (`int`):
                number of images that should be generated per prompt
            do_classifier_free_guidance (`bool`):
                whether to use classifier free guidance or not
            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`).
            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.
            negative_prompt_embeds (`torch.FloatTensor`, *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.
            lora_scale (`float`, *optional*):
                A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
        """
        # 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, LoraLoaderMixin):
            self._lora_scale = lora_scale

        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]

        if prompt_embeds is None:
            # textual inversion: procecss multi-vector tokens if necessary
            if isinstance(self, TextualInversionLoaderMixin):
                prompt = self.maybe_convert_prompt(prompt, self.tokenizer)

            text_inputs = self.tokenizer(
                prompt,
                padding="max_length",
                max_length=self.tokenizer.model_max_length,
                truncation=True,
                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.model_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.model_max_length} tokens: {removed_text}"
                )

            if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
                attention_mask = text_inputs.attention_mask.to(device)
            else:
                attention_mask = None

            prompt_embeds = self.text_encoder(
                text_input_ids.to(device),
                attention_mask=attention_mask,
            )
            prompt_embeds = prompt_embeds[0]

        prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.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)

        # get unconditional embeddings for classifier free guidance
        if do_classifier_free_guidance and negative_prompt_embeds is None:
            uncond_tokens: List[str]
            if negative_prompt is None:
                uncond_tokens = [""] * batch_size
            elif 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 isinstance(negative_prompt, str):
                uncond_tokens = [negative_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

            # textual inversion: procecss multi-vector tokens if necessary
            if isinstance(self, TextualInversionLoaderMixin):
                uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)

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

            if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
                attention_mask = uncond_input.attention_mask.to(device)
            else:
                attention_mask = None

            negative_prompt_embeds = self.text_encoder(
                uncond_input.input_ids.to(device),
                attention_mask=attention_mask,
            )
            negative_prompt_embeds = negative_prompt_embeds[0]

        if do_classifier_free_guidance:
            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
            seq_len = negative_prompt_embeds.shape[1]

            negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder.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)

            # For classifier free guidance, we need to do two forward passes.
            # Here we concatenate the unconditional and text embeddings into a single batch
            # to avoid doing two forward passes
            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])

        return prompt_embeds

    def run_safety_checker(self, image, device, dtype):
        if self.safety_checker is None:
            has_nsfw_concept = None
        else:
            if torch.is_tensor(image):
                feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
            else:
                feature_extractor_input = self.image_processor.numpy_to_pil(image)
            safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
            image, has_nsfw_concept = self.safety_checker(
                images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
            )
        return image, has_nsfw_concept

    def decode_latents(self, latents):
        warnings.warn(
            "The decode_latents method is deprecated and will be removed in a future version. Please"
            " use VaeImageProcessor instead",
            FutureWarning,
        )
        latents = 1 / self.vae.config.scaling_factor * latents
        image = self.vae.decode(latents, return_dict=False)[0]
        image = (image / 2 + 0.5).clamp(0, 1)
        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
        image = image.cpu().permute(0, 2, 3, 1).float().numpy()
        return image

    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

    def check_inputs(
        self,
        prompt,
        height,
        width,
        callback_steps,
        negative_prompt=None,
        prompt_embeds=None,
        negative_prompt_embeds=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 None) or (
            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 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 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)}")

        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."
            )

        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}."
                )

    def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
        shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, 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
    
    def _get_add_time_ids(self, original_size, crops_coords_top_left, target_size, dtype):
        add_time_ids = list(original_size + crops_coords_top_left + target_size)

        add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
        return add_time_ids

    @torch.no_grad()
    @replace_example_docstring(EXAMPLE_DOC_STRING)
    def __call__(
        self,
        prompt: Union[str, List[str]] = None,
        height: Optional[int] = None,
        width: Optional[int] = None,
        # h = 512,
        # w = 512,
        num_inference_steps: int = 50,
        guidance_scale: float = 7.5,
        negative_prompt: 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.FloatTensor] = None,
        prompt_embeds: Optional[torch.FloatTensor] = None,
        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
        output_type: Optional[str] = "pil",
        return_dict: bool = True,
        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
        callback_steps: int = 1,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        guidance_rescale: float = 0.0,
        args=None,
        batch=None,
        depth_embedder=None, 
        normal_embedder=None,
        canny_embedder=None,
        body_embedder=None,
        face_embedder=None,
        hand_embedder=None,
        original_size: Optional[Tuple[int, int]] = None,
        crops_coords_top_left: Tuple[int, int] = (0, 0),
        target_size: Optional[Tuple[int, int]] = None,
    ):
        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.
            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                The height in pixels of the generated image.
            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                The width in pixels of the generated image.
            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.
            guidance_scale (`float`, *optional*, defaults to 7.5):
                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.
            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`).
            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.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.
            negative_prompt_embeds (`torch.FloatTensor`, *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.
            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.StableDiffusionPipelineOutput`] instead of a
                plain tuple.
            callback (`Callable`, *optional*):
                A function that will be called every `callback_steps` steps during inference. The function will be
                called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
            callback_steps (`int`, *optional*, defaults to 1):
                The frequency at which the `callback` function will be called. If not specified, the callback will be
                called at every step.
            cross_attention_kwargs (`dict`, *optional*):
                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
                `self.processor` in
                [diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py).
            guidance_rescale (`float`, *optional*, defaults to 0.7):
                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.

        Examples:

        Returns:
            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
            When returning a tuple, the first element is a list with the generated images, and the second element is a
            list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
            (nsfw) content, according to the `safety_checker`.
        """
        # 0. Default height and width to unet
        height = height or self.unet.config.sample_size * self.vae_scale_factor
        width = width or self.unet.config.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, height, width, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds
        )

        # 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]
            
        # H = torch.tensor([h * batch_size]).cuda()
        # W = torch.tensor([w * batch_size]).cuda()

        device = self._execution_device
        # 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.
        do_classifier_free_guidance = guidance_scale > 1.0

        # 3. Encode input prompt
        text_encoder_lora_scale = (
            cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
        )
        prompt_embeds = self._encode_prompt(
            prompt,
            device,
            num_images_per_prompt,
            do_classifier_free_guidance,
            negative_prompt,
            prompt_embeds=prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds,
            lora_scale=text_encoder_lora_scale,
        )

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

        # 5. Prepare latent variables
        # num_channels_latents = self.unet.config.in_channels
        num_channels_latents = 4
        latents = self.prepare_latents(
            batch_size * num_images_per_prompt,
            num_channels_latents,
            height,
            width,
            prompt_embeds.dtype,
            device,
            generator,
            latents,
        )
        
        _, c, h, w = latents.shape
        
        shape = (batch_size * num_images_per_prompt, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
        
        if "depth" in args.cond_type:
            if args.cond_reshape == "vae":
                if batch is None:
                    depth_latents = torch.zeros((batch_size, c, h, w)).to(self.unet.device)
                else:
                    depth_latents = self.vae.encode(batch["depth"].to(latents.dtype)).latent_dist.sample()
                    depth_latents = depth_latents * self.vae.config.scaling_factor
                depth_latents_input = torch.cat([depth_latents] * 2) if do_classifier_free_guidance else depth_latents
            elif args.cond_reshape == "resize":
                if batch is None:
                    depth_latents = torch.zeros((batch_size, 3, h, w)).to(self.unet.device)
                else:
                    depth_latents = F.interpolate(batch['depth'], (h,w))
                depth_latents_input = torch.cat([depth_latents] * 2) if do_classifier_free_guidance else depth_latents
            elif args.cond_reshape == "learn_conv":
                if batch is None:
                    depth_latents = torch.zeros((batch_size, args.embedder_channel, h, w)).to(self.unet.device)
                else:    
                    depth_latents = depth_embedder(batch['depth'])
                depth_latents_input = torch.cat([depth_latents] * 2) if do_classifier_free_guidance else depth_latents
            else:
                assert False, "unknown condition reshape type"
              
            if "depth" in args.noisy_cond:    
                # depth_latents = randn_tensor(shape, generator=generator, device=device, dtype=prompt_embeds.dtype)
                # depth_latents = depth_latents * self.scheduler.init_noise_sigma
                depth_latents = latents.clone()
                
        if "normal" in args.cond_type:
            if args.cond_reshape == "vae":
                if batch is None:
                    normal_latents = torch.zeros((batch_size, c, h, w)).to(self.unet.device)
                else:
                    normal_latents = self.vae.encode(batch["normal"].to(latents.dtype)).latent_dist.sample()
                    normal_latents = normal_latents * self.vae.config.scaling_factor
                normal_latents_input = torch.cat([normal_latents] * 2) if do_classifier_free_guidance else normal_latents
            elif args.cond_reshape == "resize":
                if batch is None:
                    normal_latents = torch.zeros((batch_size, 3, h, w)).to(self.unet.device)
                else:
                    normal_latents = F.interpolate(batch['normal'], (h,w))
                normal_latents_input = torch.cat([normal_latents] * 2) if do_classifier_free_guidance else normal_latents
            elif args.cond_reshape == "learn_conv":
                if batch is None:
                    normal_latents = torch.zeros((batch_size, args.embedder_channel, h, w)).to(self.unet.device)
                else:
                    normal_latents = normal_embedder(batch['normal'])
                normal_latents_input = torch.cat([normal_latents] * 2) if do_classifier_free_guidance else normal_latents
            else:
                assert False, "unknown condition reshape type"
                
            if "normal" in args.noisy_cond:    
                # normal_latents = randn_tensor(shape, generator=generator, device=device, dtype=prompt_embeds.dtype)
                # normal_latents = normal_latents * self.scheduler.init_noise_sigma
                normal_latents = latents.clone()
                
        if "canny" in args.cond_type:
            if args.cond_reshape == "vae":
                if batch is None:
                    canny_latents = torch.zeros((batch_size, c, h, w)).to(self.unet.device)
                else:
                    canny_latents = self.vae.encode(batch["canny"].to(latents.dtype)).latent_dist.sample()
                    canny_latents = canny_latents * self.vae.config.scaling_factor
                canny_latents_input = torch.cat([canny_latents] * 2) if do_classifier_free_guidance else canny_latents
            elif args.cond_reshape == "resize":
                if batch is None:
                    canny_latents = torch.zeros((batch_size, 3, h, w)).to(self.unet.device)
                else:
                    canny_latents = F.interpolate(batch['canny'], (h,w))
                canny_latents_input = torch.cat([canny_latents] * 2) if do_classifier_free_guidance else canny_latents
            elif args.cond_reshape == "learn_conv":
                if batch is None:
                    canny_latents = torch.zeros((batch_size, args.embedder_channel, h, w)).to(self.unet.device)
                else:
                    canny_latents = canny_embedder(batch['canny'])
                canny_latents_input = torch.cat([canny_latents] * 2) if do_classifier_free_guidance else canny_latents
            else:
                assert False, "unknown condition reshape type"
                
            if "canny" in args.noisy_cond:    
                # canny_latents = randn_tensor(shape, generator=generator, device=device, dtype=prompt_embeds.dtype)
                # canny_latents = canny_latents * self.scheduler.init_noise_sigma
                canny_latents = latents.clone()
                
        if "body" in args.cond_type:
            if args.cond_reshape == "vae":
                if batch is None:
                    body_latents = torch.zeros((batch_size, c, h, w)).to(self.unet.device)
                else:
                    body_latents = self.vae.encode(batch["body"].to(latents.dtype)).latent_dist.sample()
                    body_latents = body_latents * self.vae.config.scaling_factor
                body_latents_input = torch.cat([body_latents] * 2) if do_classifier_free_guidance else body_latents
            elif args.cond_reshape == "resize":
                if batch is None:
                    body_latents = torch.zeros((batch_size, 3, h, w)).to(self.unet.device)
                else:
                    body_latents = F.interpolate(batch['body'], (h,w))
                body_latents_input = torch.cat([body_latents] * 2) if do_classifier_free_guidance else body_latents
            elif args.cond_reshape == "learn_conv":
                if batch is None:
                    body_latents = torch.zeros((batch_size, args.embedder_channel, h, w)).to(self.unet.device)
                else:
                    body_latents = body_embedder(batch['body'])
                body_latents_input = torch.cat([body_latents] * 2) if do_classifier_free_guidance else body_latents
            else:
                assert False, "unknown condition reshape type"
                
            if "body" in args.noisy_cond:    
                # body_latents = randn_tensor(shape, generator=generator, device=device, dtype=prompt_embeds.dtype)
                # body_latents = body_latents * self.scheduler.init_noise_sigma
                body_latents = latents.clone()
                
        if "face" in args.cond_type:
            if args.cond_reshape == "vae":
                if batch is None:
                    face_latents = torch.zeros((batch_size, c, h, w)).to(self.unet.device)
                else:
                    face_latents = self.vae.encode(batch["face"].to(latents.dtype)).latent_dist.sample()
                    face_latents = face_latents * self.vae.config.scaling_factor
                face_latents_input = torch.cat([face_latents] * 2) if do_classifier_free_guidance else face_latents
            elif args.cond_reshape == "resize":
                if batch is None:
                    face_latents = torch.zeros((batch_size, 3, h, w)).to(self.unet.device)
                else:
                    face_latents = F.interpolate(batch['face'], (h,w))
                face_latents_input = torch.cat([face_latents] * 2) if do_classifier_free_guidance else face_latents
            elif args.cond_reshape == "learn_conv":
                if batch is None:
                    face_latents = torch.zeros((batch_size, args.embedder_channel, h, w)).to(self.unet.device)
                else:
                    face_latents = face_embedder(batch['face'])
                face_latents_input = torch.cat([face_latents] * 2) if do_classifier_free_guidance else face_latents
            else:
                assert False, "unknown condition reshape type"
                
            if "face" in args.noisy_cond:    
                # face_latents = randn_tensor(shape, generator=generator, device=device, dtype=prompt_embeds.dtype)
                # face_latents = face_latents * self.scheduler.init_noise_sigma
                face_latents = latents.clone()
                
        if "hand" in args.cond_type:
            if args.cond_reshape == "vae":
                if batch is None:
                    hand_latents = torch.zeros((batch_size, c, h, w)).to(self.unet.device)
                else:
                    hand_latents = self.vae.encode(batch["hand"].to(latents.dtype)).latent_dist.sample()
                    hand_latents = hand_latents * self.vae.config.scaling_factor
                hand_latents_input = torch.cat([hand_latents] * 2) if do_classifier_free_guidance else hand_latents
            elif args.cond_reshape == "resize":
                if batch is None:
                    hand_latents = torch.zeros((batch_size, 3, h, w)).to(self.unet.device)
                else:
                    hand_latents = F.interpolate(batch['hand'], (h,w))
                hand_latents_input = torch.cat([hand_latents] * 2) if do_classifier_free_guidance else hand_latents
            elif args.cond_reshape == "learn_conv":
                if batch is None:
                    hand_latents = torch.zeros((batch_size, args.embedder_channel, h, w)).to(self.unet.device)
                else:
                    hand_latents = hand_embedder(batch['hand'])
                hand_latents_input = torch.cat([hand_latents] * 2) if do_classifier_free_guidance else hand_latents
            else:
                assert False, "unknown condition reshape type"
                
            if "hand" in args.noisy_cond:    
                # hand_latents = randn_tensor(shape, generator=generator, device=device, dtype=prompt_embeds.dtype)
                # hand_latents = hand_latents * self.scheduler.init_noise_sigma
                hand_latents = latents.clone()
                
        if "ldmk" in args.cond_type:
            if args.cond_reshape == "vae":
                if batch is None:
                    ldmk_latents = torch.zeros((batch_size, c, h, w)).to(self.unet.device)
                else:
                    ldmk_latents = self.vae.encode(batch["ldmk"].to(latents.dtype)).latent_dist.sample()
                    ldmk_latents = ldmk_latents * self.vae.config.scaling_factor
                ldmk_latents_input = torch.cat([ldmk_latents] * 2) if do_classifier_free_guidance else ldmk_latents
            elif args.cond_reshape == "resize":
                if batch is None:
                    ldmk_latents = torch.zeros((batch_size, 3, h, w)).to(self.unet.device)
                else:
                    ldmk_latents = F.interpolate(batch['ldmk'], (h,w))
                ldmk_latents_input = torch.cat([ldmk_latents] * 2) if do_classifier_free_guidance else ldmk_latents
            elif args.cond_reshape == "learn_conv":
                if batch is None:
                    ldmk_latents = torch.zeros((batch_size, args.embedder_channel, h, w)).to(self.unet.device)
                else:
                    ldmk_latents = hand_embedder(batch['ldmk'])
                ldmk_latents_input = torch.cat([ldmk_latents] * 2) if do_classifier_free_guidance else ldmk_latents
            else:
                assert False, "unknown condition reshape type"
                
            if "ldmk" in args.noisy_cond:    
                # hand_latents = randn_tensor(shape, generator=generator, device=device, dtype=prompt_embeds.dtype)
                # hand_latents = hand_latents * self.scheduler.init_noise_sigma
                ldmk_latents = latents.clone()

        # 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)
        
        add_time_ids = self._get_add_time_ids(
            original_size, crops_coords_top_left, target_size, dtype=prompt_embeds.dtype
        )
        if do_classifier_free_guidance:
            add_time_ids = torch.cat([add_time_ids, add_time_ids], dim=0)
        add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)

        # 7. Denoising loop
        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
        with self.progress_bar(total=num_inference_steps) as progress_bar:
            for i, t in enumerate(timesteps):
                # expand the latents if we are doing classifier free guidance
                latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
                
                if "depth" in args.noisy_cond:
                    depth_latents_input = torch.cat([depth_latents] * 2) if do_classifier_free_guidance else depth_latents
                    depth_latents_input = self.scheduler.scale_model_input(depth_latents_input, t)
                    
                if "normal" in args.noisy_cond:
                    normal_latents_input = torch.cat([normal_latents] * 2) if do_classifier_free_guidance else normal_latents
                    normal_latents_input = self.scheduler.scale_model_input(normal_latents_input, t)
                    
                if "canny" in args.noisy_cond:
                    canny_latents_input = torch.cat([canny_latents] * 2) if do_classifier_free_guidance else canny_latents
                    canny_latents_input = self.scheduler.scale_model_input(canny_latents_input, t)
                    
                if "body" in args.noisy_cond:
                    body_latents_input = torch.cat([body_latents] * 2) if do_classifier_free_guidance else body_latents
                    body_latents_input = self.scheduler.scale_model_input(body_latents_input, t)
                    
                if "face" in args.noisy_cond:
                    face_latents_input = torch.cat([face_latents] * 2) if do_classifier_free_guidance else face_latents
                    face_latents_input = self.scheduler.scale_model_input(face_latents_input, t)
                    
                if "hand" in args.noisy_cond:
                    hand_latents_input = torch.cat([hand_latents] * 2) if do_classifier_free_guidance else hand_latents
                    hand_latents_input = self.scheduler.scale_model_input(hand_latents_input, t)
                    
                if "ldmk" in args.noisy_cond:
                    ldmk_latents_input = torch.cat([ldmk_latents] * 2) if do_classifier_free_guidance else ldmk_latents
                    ldmk_latents_input = self.scheduler.scale_model_input(ldmk_latents_input, t)
                
                _, c, h, w = latent_model_input.shape
                        
                if args.cond_inject == "concat":
                    latent_model_input = torch.cat([latent_model_input, depth_latents_input], dim=1) if "depth" in args.cond_type else latent_model_input
                    latent_model_input = torch.cat([latent_model_input, normal_latents_input], dim=1) if "normal" in args.cond_type else latent_model_input
                    latent_model_input = torch.cat([latent_model_input, canny_latents_input], dim=1) if "canny" in args.cond_type else latent_model_input
                    latent_model_input = torch.cat([latent_model_input, body_latents_input], dim=1) if "body" in args.cond_type else latent_model_input
                    latent_model_input = torch.cat([latent_model_input, face_latents_input], dim=1) if "face" in args.cond_type else latent_model_input
                    latent_model_input = torch.cat([latent_model_input, hand_latents_input], dim=1) if "hand" in args.cond_type else latent_model_input
                    latent_model_input = torch.cat([latent_model_input, ldmk_latents_input], dim=1) if "ldmk" in args.cond_type else latent_model_input
                elif args.cond_inject == "sum":
                    if len(args.cond_type) == 0:
                        pass
                    else:
                        if args.cond_reshape == "vae":
                            channel_dim = 4
                        elif args.cond_reshape == "resize":
                            channel_dim = 3
                        elif args.cond_reshape == "learn_conv":
                            channel_dim = args.embedder_channel
                        sum_latents = torch.zeros((latent_model_input.shape[0], channel_dim, h, w)).to(self.unet.device)
                        sum_latents = sum_latents + depth_latents_input if "depth" in args.cond_type else sum_latents
                        sum_latents = sum_latents + normal_latents_input if "normal" in args.cond_type else sum_latents
                        sum_latents = sum_latents + canny_latents_input if "canny" in args.cond_type else sum_latents
                        sum_latents = sum_latents + body_latents_input if "body" in args.cond_type else sum_latents
                        sum_latents = sum_latents + face_latents_input if "face" in args.cond_type else sum_latents
                        sum_latents = sum_latents + hand_latents_input if "hand" in args.cond_type else sum_latents
                        latent_model_input = torch.cat([latent_model_input, sum_latents], dim=1)

                added_cond_kwargs = {"time_ids": add_time_ids}

                # predict the noise residual
                if args.cond_inject == "spade":
                    if batch is None:
                        num_cond = 0
                        if "depth" in args.cond_type: num_cond += 1
                        if "normal" in args.cond_type: num_cond += 1
                        if "canny" in args.cond_type: num_cond += 1
                        if "body" in args.cond_type: num_cond += 1
                        if "face" in args.cond_type: num_cond += 1
                        if "hand" in args.cond_type: num_cond += 1
                        if "ldmk" in args.cond_type: num_cond += 1
                        label_channels = num_cond * 3
                        structural_cond = torch.zeros((batch_size, label_channels, h, w)).to(self.unet.device)
                    else:
                        structural_cond = []
                        if "depth" in args.cond_type:
                            structural_cond.append(batch["depth"])
                        if "normal" in args.cond_type:
                            structural_cond.append(batch["normal"])
                        if "canny" in args.cond_type:
                            structural_cond.append(batch["canny"])
                        if "body" in args.cond_type:
                            structural_cond.append(batch["body"])
                        if "face" in args.cond_type:
                            structural_cond.append(batch["face"])
                        if "hand" in args.cond_type:
                            structural_cond.append(batch["hand"])
                        if "ldmk" in args.cond_type:
                            structural_cond.append(batch["ldmk"])
                        structural_cond = torch.cat(structural_cond, dim=1)
                    structural_cond = torch.cat([structural_cond] * 2) if do_classifier_free_guidance else structural_cond
                    noise_pred = self.unet(
                        latent_model_input,
                        structural_cond,
                        t,
                        encoder_hidden_states=prompt_embeds,
                        cross_attention_kwargs=cross_attention_kwargs,
                        return_dict=False,
                    )[0]
                else:
                    if t <= self.scheduler.config.num_train_timesteps // 4:
                        noise_pred = self.unet(
                            latent_model_input,
                            t,
                            added_cond_kwargs=added_cond_kwargs,
                            encoder_hidden_states=prompt_embeds,
                            cross_attention_kwargs=cross_attention_kwargs,
                            return_dict=False,
                        )[0]
                    elif t >= self.scheduler.config.num_train_timesteps // 4 and t <= self.scheduler.config.num_train_timesteps // 4 * 2:
                        noise_pred = self.unet2(
                            latent_model_input,
                            t,
                            added_cond_kwargs=added_cond_kwargs,
                            encoder_hidden_states=prompt_embeds,
                            cross_attention_kwargs=cross_attention_kwargs,
                            return_dict=False,
                        )[0]
                    else:
                        noise_pred = self.unet3(
                            latent_model_input,
                            t,
                            added_cond_kwargs=added_cond_kwargs,
                            encoder_hidden_states=prompt_embeds,
                            cross_attention_kwargs=cross_attention_kwargs,
                            return_dict=False,
                        )[0]

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

                if do_classifier_free_guidance and 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=guidance_rescale)
                
                if noise_pred.shape[1] > 4:
                    cond_pred = noise_pred[:, 4:]
                    noise_pred = noise_pred[:, :4]
                    if "depth" in args.cond_type:
                        depth_pred = cond_pred[:, :4]
                        if cond_pred.shape[1] > 4:
                            cond_pred = cond_pred[:, 4:]
                    if "normal" in args.cond_type:
                        normal_pred = cond_pred[:, :4]
                        if cond_pred.shape[1] > 4:
                            cond_pred = cond_pred[:, 4:]
                    if "canny" in args.cond_type:
                        canny_pred = cond_pred[:, :4]
                        if cond_pred.shape[1] > 4:
                            cond_pred = cond_pred[:, 4:]
                    if "body" in args.cond_type:
                        body_pred = cond_pred[:, :4]
                        if cond_pred.shape[1] > 4:
                            cond_pred = cond_pred[:, 4:]
                    if "face" in args.cond_type:
                        face_pred = cond_pred[:, :4]
                        if cond_pred.shape[1] > 4:
                            cond_pred = cond_pred[:, 4:]
                    if "hand" in args.cond_type:
                        hand_pred = cond_pred[:, :4]
                        if cond_pred.shape[1] > 4:
                            cond_pred = cond_pred[:, 4:] 

                # compute the previous noisy sample x_t -> x_t-1
                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
                
                if "depth" in args.noisy_cond:
                    depth_latents = self.scheduler.step(depth_pred, t, depth_latents, **extra_step_kwargs, return_dict=False)[0]
                if "normal" in args.noisy_cond:
                    normal_latents = self.scheduler.step(normal_pred, t, normal_latents, **extra_step_kwargs, return_dict=False)[0]
                if "canny" in args.noisy_cond:
                    canny_latents = self.scheduler.step(canny_pred, t, canny_latents, **extra_step_kwargs, return_dict=False)[0]
                if "body" in args.noisy_cond:
                    body_latents = self.scheduler.step(body_pred, t, body_latents, **extra_step_kwargs, return_dict=False)[0]
                if "face" in args.noisy_cond:
                    face_latents = self.scheduler.step(face_pred, t, face_latents, **extra_step_kwargs, return_dict=False)[0]
                if "hand" in args.noisy_cond:
                    hand_latents = self.scheduler.step(hand_pred, t, hand_latents, **extra_step_kwargs, return_dict=False)[0]

                # 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:
                        callback(i, t, latents)

        if not output_type == "latent":
            image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
            image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
            if "depth" in args.noisy_cond:
                depth_image = self.vae.decode(depth_latents / self.vae.config.scaling_factor, return_dict=False)[0]
            if "normal" in args.noisy_cond:
                normal_image = self.vae.decode(normal_latents / self.vae.config.scaling_factor, return_dict=False)[0]
            if "canny" in args.noisy_cond:
                canny_image = self.vae.decode(canny_latents / self.vae.config.scaling_factor, return_dict=False)[0]
            if "body" in args.noisy_cond:
                body_image = self.vae.decode(body_latents / self.vae.config.scaling_factor, return_dict=False)[0]
            if "face" in args.noisy_cond:
                face_image = self.vae.decode(face_latents / self.vae.config.scaling_factor, return_dict=False)[0]
            if "hand" in args.noisy_cond:
                hand_image = self.vae.decode(hand_latents / self.vae.config.scaling_factor, return_dict=False)[0]
        else:
            image = latents
            if "depth" in args.noisy_cond:
                depth_image = depth_latents
            if "normal" in args.noisy_cond:
                normal_image = normal_latents
            if "canny" in args.noisy_cond:
                canny_image = canny_latents
            if "body" in args.noisy_cond:
                body_image = body_latents
            if "face" in args.noisy_cond:
                face_image = face_latents
            if "hand" in args.noisy_cond:
                hand_image = hand_latents
            has_nsfw_concept = None

        if has_nsfw_concept is None:
            do_denormalize = [True] * image.shape[0]
        else:
            do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]

        image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
        if "depth" in args.noisy_cond:
            depth_image = self.image_processor.postprocess(depth_image, output_type=output_type, do_denormalize=do_denormalize)
        if "normal" in args.noisy_cond:
            normal_image = self.image_processor.postprocess(normal_image, output_type=output_type, do_denormalize=do_denormalize)
        if "canny" in args.noisy_cond:
            canny_image = self.image_processor.postprocess(canny_image, output_type=output_type, do_denormalize=do_denormalize)
        if "body" in args.noisy_cond:
            body_image = self.image_processor.postprocess(body_image, output_type=output_type, do_denormalize=do_denormalize)
        if "face" in args.noisy_cond:
            face_image = self.image_processor.postprocess(face_image, output_type=output_type, do_denormalize=do_denormalize)
        if "hand" in args.noisy_cond:
            hand_image = self.image_processor.postprocess(hand_image, output_type=output_type, do_denormalize=do_denormalize)
            
        # Offload last model to CPU
        if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
            self.final_offload_hook.offload()

        if not return_dict:
            output_tuple = (image)
            if "depth" in args.noisy_cond:
                output_tuple = output_tuple + (depth_image)
            if "normal" in args.noisy_cond:
                output_tuple = output_tuple + (normal_image)
            if "canny" in args.noisy_cond:
                output_tuple = output_tuple + (canny_image)
            if "body" in args.noisy_cond:
                output_tuple = output_tuple + (body_image)
            if "face" in args.noisy_cond:
                output_tuple = output_tuple + (face_image)
            if "hand" in args.noisy_cond:
                output_tuple = output_tuple + (hand_image)
            return output_tuple + (has_nsfw_concept)
        
        output = StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
        
        if "depth" in args.noisy_cond:
            output["depth_image"] = depth_image
        if "normal" in args.noisy_cond:
            output["normal_image"] = normal_image
        if "canny" in args.noisy_cond:
            output["canny_image"] = canny_image
        if "body" in args.noisy_cond:
            output["body_image"] = body_image
        if "face" in args.noisy_cond:
            output["face_image"] = face_image
        if "hand" in args.noisy_cond:
            output["hand_image"] = hand_image
        
        return output