pipeline_stable_diffusion.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 numpy  as np
import pickle
import inspect
import math
from typing import Any, Callable, Dict, List, Optional, Union
from PIL import Image
import torch
import torch.nn.functional as F
import torch.optim as optim

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 .unet_2d_condition import UNet2DConditionModel
from diffusers.models.lora import adjust_lora_scale_text_encoder
from diffusers.schedulers import KarrasDiffusionSchedulers
from diffusers.utils import (
    deprecate,
    logging,
    replace_example_docstring,
    PIL_INTERPOLATION
)
from diffusers.utils.torch_utils import randn_tensor
from diffusers.pipeline_utils import DiffusionPipeline
from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
import torch.optim as optim

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 preprocess_image(image, batch_size):
    w, h = image.size
    w, h = (x - x % 8 for x in (w, h))  # resize to integer multiple of 8
    image = image.resize((w, h), resample=PIL_INTERPOLATION["lanczos"])
    image = np.array(image).astype(np.float32) / 255.0
    image = np.vstack([image[None].transpose(0, 3, 1, 2)] * batch_size)
    image = torch.from_numpy(image)
    return 2.0 * image - 1.0

def preprocess_mask(mask, batch_size, scale_factor=8):
    if not isinstance(mask, torch.FloatTensor):
        mask = mask.convert("L")
        w, h = mask.size
        w, h = (x - x % 8 for x in (w, h))  # resize to integer multiple of 8
        mask = mask.resize((w // scale_factor, h // scale_factor), resample=PIL_INTERPOLATION["nearest"])
        mask = np.array(mask).astype(np.float32) / 255.0
        mask = np.tile(mask, (4, 1, 1))
        mask = np.vstack([mask[None]] * batch_size)
        mask = 1 - mask  # repaint white, keep black
        mask = torch.from_numpy(mask)
        return mask

    else:
        valid_mask_channel_sizes = [1, 3]
        # if mask channel is fourth tensor dimension, permute dimensions to pytorch standard (B, C, H, W)
        if mask.shape[3] in valid_mask_channel_sizes:
            mask = mask.permute(0, 3, 1, 2)
        elif mask.shape[1] not in valid_mask_channel_sizes:
            raise ValueError(
                f"Mask channel dimension of size in {valid_mask_channel_sizes} should be second or fourth dimension,"
                f" but received mask of shape {tuple(mask.shape)}"
            )
        # (potentially) reduce mask channel dimension from 3 to 1 for broadcasting to latent shape
        mask = mask.mean(dim=1, keepdim=True)
        h, w = mask.shape[-2:]
        h, w = (x - x % 8 for x in (h, w))  # resize to integer multiple of 8
        mask = torch.nn.functional.interpolate(mask, (h // scale_factor, w // scale_factor))
        return mask
    
def scale_proportion(obj_box, H, W, use_legacy=False):
    if use_legacy:
        # Bias towards the top-left corner
        x_min, y_min, x_max, y_max = int(obj_box[0] * W), int(obj_box[1] * H), int(obj_box[2] * W), int(obj_box[3] * H)
    else:
        # Separately rounding box_w and box_h to allow shift invariant box sizes. Otherwise box sizes may change when both coordinates being rounded end with ".5".
        x_min, y_min = round(obj_box[0] * W), round(obj_box[1] * H)
        box_w, box_h = round((obj_box[2] - obj_box[0]) * W), round((obj_box[3] - obj_box[1]) * H)
        x_max, y_max = x_min + box_w, y_min + box_h
        
        x_min, y_min = max(x_min, 0), max(y_min, 0)
        x_max, y_max = min(x_max, W), min(y_max, H)
        
    return x_min, y_min, x_max, y_max


def add_ca_loss_per_attn_map_to_loss(loss, attn_map, object_number, bboxes, object_positions, use_ratio_based_loss=False, fg_top_p=0.2, bg_top_p=0.2, fg_weight=1.0, bg_weight=4.0, verbose=False):

    b, i, j = attn_map.shape
    H = W = int(math.sqrt(i))
    for obj_idx in range(object_number):
        obj_loss = 0
        mask = torch.zeros(size=(H, W), device="cuda")
        obj_boxes = bboxes[obj_idx]

        for obj_box in obj_boxes:

            x_min, y_min, x_max, y_max = scale_proportion(obj_box, H=H, W=W)
            mask[y_min: y_max, x_min: x_max] = 1

        for obj_position in object_positions[obj_idx]:
            
            ca_map_obj = attn_map[:, :, obj_position].reshape(b, H, W)

            if use_ratio_based_loss:                
                activation_value = (ca_map_obj * mask).reshape(b, -1).sum(dim=-1)/ca_map_obj.reshape(b, -1).sum(dim=-1)
                obj_loss += torch.mean((1 - activation_value) ** 2)

            else:
                ca_map_obj = attn_map[:, :, obj_position] # .reshape(b, H, W)
                k_fg = (mask.sum() * fg_top_p).long().clamp_(min=1)
                k_bg = ((1 - mask).sum() * bg_top_p).long().clamp_(min=1)
                
                mask_1d = mask.view(1, -1)
                
                obj_loss += (1 - (ca_map_obj * mask_1d).topk(k=k_fg).values.mean(dim=1)).sum(dim=0) * fg_weight
                obj_loss += ((ca_map_obj * (1 - mask_1d)).topk(k=k_bg).values.mean(dim=1)).sum(dim=0) * bg_weight    

        loss += obj_loss / len(object_positions[obj_idx])
        
    return loss



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):
    
    model_cpu_offload_seq = "text_encoder->unet->vae"
    _optional_components = ["safety_checker", "feature_extractor"]
    _exclude_from_cpu_offload = ["safety_checker"]

    def __init__(
        self,
        vae: AutoencoderKL,
        text_encoder: CLIPTextModel,
        tokenizer: CLIPTokenizer,
        unet: 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)

        self.register_modules(
            vae=vae,
            text_encoder=text_encoder,
            tokenizer=tokenizer,
            unet=unet,
            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 enabled, 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 for saving a large amount of memory and to allow
        processing larger images.
        """
        self.vae.enable_tiling()

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

    def _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,
    ):
        deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
        deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)

        prompt_embeds_tuple = self.encode_prompt(
            prompt=prompt,
            device=device,
            num_images_per_prompt=num_images_per_prompt,
            do_classifier_free_guidance=do_classifier_free_guidance,
            negative_prompt=negative_prompt,
            prompt_embeds=prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds,
            lora_scale=lora_scale,
        )

        # concatenate for backwards comp
        prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])

        return prompt_embeds

    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

            # dynamically adjust the LoRA scale
            adjust_lora_scale_text_encoder(self.text_encoder, 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 #[1,77]
            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]

        if self.text_encoder is not None:
            prompt_embeds_dtype = self.text_encoder.dtype
        elif self.unet is not None:
            prompt_embeds_dtype = self.unet.dtype
        else:
            prompt_embeds_dtype = prompt_embeds.dtype

        prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_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=prompt_embeds_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)

        return prompt_embeds, negative_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):
        deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
        deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)

        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."
            )
        
    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 prepare_inpainting_latents(self, image, timestep, num_images_per_prompt, dtype, device, generator):
        image = image.to(device=device, dtype=dtype)
        init_latent_dist = self.vae.encode(image).latent_dist
        init_latents = init_latent_dist.sample(generator=generator)
        init_latents = self.vae.config.scaling_factor * init_latents

        # Expand init_latents for batch_size and num_images_per_prompt
        init_latents = torch.cat([init_latents] * num_images_per_prompt, dim=0)
        init_latents_orig = init_latents

        # add noise to latents using the timesteps
        noise = randn_tensor(init_latents.shape, generator=generator, device=device, dtype=dtype)
        init_latents = self.scheduler.add_noise(init_latents, noise, timestep)
        latents = init_latents
        return latents, init_latents_orig, noise
    
    def get_timesteps(self, num_inference_steps, strength, device):
        # get the original timestep using init_timestep
        init_timestep = min(int(num_inference_steps * strength), num_inference_steps)

        t_start = max(num_inference_steps - init_timestep, 0)
        timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]

        return timesteps, num_inference_steps - t_start
    
    #@torch.no_grad()
    #@replace_example_docstring(EXAMPLE_DOC_STRING)
    def __call__(
        self,
        prompt: Union[str, List[str]] = None,
        height: int = 512,
        width: int = 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,
        prompt_book_info: list = None,
        do_latent_guidance: bool = False,
        mid_state_img: Image.Image = None,
        mid_state_mask: np.ndarray = None,
        layout_mask: np.ndarray = None,
        refine_step: int = 60,
        bboxes: List = [],
        object_positions: List = [],
        return_latents: bool = False,
        is_editing: bool = False,
        repeat_ind: int = 0,
    ):
      
        # 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
        
        # 1. Check inputs. Raise error if not correct
        self.check_inputs(
            prompt, height, width, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds
        )
        if layout_mask is None:
            raise ValueError("layout_mask couldn't be None!")

        # 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
        # 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
        )

        # 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
        if do_classifier_free_guidance:
            prompt_embeds = torch.cat([prompt_embeds, negative_prompt_embeds])

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

        # 4.5 Prepare inpaintings
        if is_editing:
            if not isinstance(is_editing[0], torch.FloatTensor):
                edit_before = preprocess_image(is_editing[0], 1)
            else:
                edit_before = is_editing[0]
            edit_msk = preprocess_mask(is_editing[1], 1, self.vae_scale_factor)
            timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, is_editing[2], device)
            latent_timestep = timesteps[:1].repeat(1 * 1)

            latents, init_latents_orig, noise = self.prepare_inpainting_latents(
                edit_before, latent_timestep, 1, prompt_embeds.dtype, device, generator
            )

            mask = edit_msk.to(device=device, dtype=latents.dtype)
            mask = torch.cat([mask] * 1)

        else:
            num_channels_latents = self.unet.config.in_channels
            latents = self.prepare_latents( 
                1 * num_images_per_prompt,
                num_channels_latents,
                height,
                width,
                prompt_embeds.dtype,
                device,
                generator,
                latents,
            )

        # 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)

        
        # 6.2 Prepare mid stage image latents for loss guidance
        if do_latent_guidance:
            inp_img = np.array(mid_state_img).astype(np.float32) / 255.0
            inp_img = np.vstack([inp_img[None].transpose(0, 3, 1, 2)] * 1)
            inp_img = torch.from_numpy(inp_img)
            inp_img = 2.0 * inp_img - 1.0

            myimage = inp_img.to(device=device, dtype=latents.dtype)
            init_latent_dist = self.vae.encode(myimage).latent_dist
            init_latents = init_latent_dist.sample(generator=generator)
            init_latents = self.vae.config.scaling_factor * init_latents
            init_latents = torch.cat([init_latents] * 1, dim=0)

            noise = randn_tensor(init_latents.shape, generator=generator, device=torch.device(device), dtype=latents.dtype)
            init_latents = self.scheduler.add_noise(init_latents, noise, self.scheduler.timesteps)
            mid_state_latents = init_latents.unsqueeze(1)

        

        loss = torch.tensor(10000.)
        # 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):
                if do_latent_guidance and i < 10:
                    latents = latents * (1 - mid_state_mask) + (latents * mid_state_mask * 0.2 + mid_state_latents[i] * mid_state_mask * 0.8)
                    latents = latents.half()

                with torch.no_grad():
                    # 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 prompt_book_info[1] and i >= refine_step:
                        print('Stop patallel cross attention by step:',i)
                        cond, negative = prompt_embeds[0:1,:,:], prompt_embeds[-1:,:,:] #global, bg, cha1, cha2, ..., negative
                        prompt_embeds = torch.cat([cond,negative])
                        print(cond.size(), negative.size())
                        prompt_book_info[1] = False

                    # predict the noise residual
                    noise_pred, _ = self.unet(
                        latent_model_input,
                        t,
                        encoder_hidden_states=prompt_embeds,
                        cross_attention_kwargs=cross_attention_kwargs,
                        return_dict=True,
                        prompt_book_info=prompt_book_info,
                        layout_mask=layout_mask,
                    )

                # perform guidance
                if do_classifier_free_guidance:
                    noise_pred_text, noise_pred_uncond = 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)

                # 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]

                # inpainting
                if is_editing:
                    init_latents_proper = self.scheduler.add_noise(init_latents_orig, noise, torch.tensor([t]))
                    latents = (init_latents_proper * mask) + (latents * (1 - mask))


                # 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)
        
        
        # use original latents corresponding to unmasked portions of the image                
        if is_editing:
            latents = (init_latents_orig * mask) + (latents * (1 - mask))

        # save generate image
        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)
        else:
            image = 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 = image.detach()
        image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)

        # Offload all models
        self.maybe_free_model_hooks()

        if not return_dict:
            return (image, has_nsfw_concept)

        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)