lpw_stable_diffusion_onnx.py

import inspect
import re
from typing import Callable, List, Optional, Union

import numpy as np
import PIL.Image
import torch
from packaging import version
from transformers import CLIPImageProcessor, CLIPTokenizer

import diffusers
from diffusers import OnnxRuntimeModel, OnnxStableDiffusionPipeline, SchedulerMixin
from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
from diffusers.utils import logging


try:
    from diffusers.pipelines.onnx_utils import ORT_TO_NP_TYPE
except ImportError:
    ORT_TO_NP_TYPE = {
        "tensor(bool)": np.bool_,
        "tensor(int8)": np.int8,
        "tensor(uint8)": np.uint8,
        "tensor(int16)": np.int16,
        "tensor(uint16)": np.uint16,
        "tensor(int32)": np.int32,
        "tensor(uint32)": np.uint32,
        "tensor(int64)": np.int64,
        "tensor(uint64)": np.uint64,
        "tensor(float16)": np.float16,
        "tensor(float)": np.float32,
        "tensor(double)": np.float64,
    }

try:
    from diffusers.utils import PIL_INTERPOLATION
except ImportError:
    if version.parse(version.parse(PIL.__version__).base_version) >= version.parse("9.1.0"):
        PIL_INTERPOLATION = {
            "linear": PIL.Image.Resampling.BILINEAR,
            "bilinear": PIL.Image.Resampling.BILINEAR,
            "bicubic": PIL.Image.Resampling.BICUBIC,
            "lanczos": PIL.Image.Resampling.LANCZOS,
            "nearest": PIL.Image.Resampling.NEAREST,
        }
    else:
        PIL_INTERPOLATION = {
            "linear": PIL.Image.LINEAR,
            "bilinear": PIL.Image.BILINEAR,
            "bicubic": PIL.Image.BICUBIC,
            "lanczos": PIL.Image.LANCZOS,
            "nearest": PIL.Image.NEAREST,
        }
# ------------------------------------------------------------------------------

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

re_attention = re.compile(
    r"""
\\\(|
\\\)|
\\\[|
\\]|
\\\\|
\\|
\(|
\[|
:([+-]?[.\d]+)\)|
\)|
]|
[^\\()\[\]:]+|
:
""",
    re.X,
)


def parse_prompt_attention(text):
    """
    Parses a string with attention tokens and returns a list of pairs: text and its associated weight.
    Accepted tokens are:
      (abc) - increases attention to abc by a multiplier of 1.1
      (abc:3.12) - increases attention to abc by a multiplier of 3.12
      [abc] - decreases attention to abc by a multiplier of 1.1
      \\( - literal character '('
      \\[ - literal character '['
      \\) - literal character ')'
      \\] - literal character ']'
      \\ - literal character '\'
      anything else - just text
    >>> parse_prompt_attention('normal text')
    [['normal text', 1.0]]
    >>> parse_prompt_attention('an (important) word')
    [['an ', 1.0], ['important', 1.1], [' word', 1.0]]
    >>> parse_prompt_attention('(unbalanced')
    [['unbalanced', 1.1]]
    >>> parse_prompt_attention('\\(literal\\]')
    [['(literal]', 1.0]]
    >>> parse_prompt_attention('(unnecessary)(parens)')
    [['unnecessaryparens', 1.1]]
    >>> parse_prompt_attention('a (((house:1.3)) [on] a (hill:0.5), sun, (((sky))).')
    [['a ', 1.0],
     ['house', 1.5730000000000004],
     [' ', 1.1],
     ['on', 1.0],
     [' a ', 1.1],
     ['hill', 0.55],
     [', sun, ', 1.1],
     ['sky', 1.4641000000000006],
     ['.', 1.1]]
    """

    res = []
    round_brackets = []
    square_brackets = []

    round_bracket_multiplier = 1.1
    square_bracket_multiplier = 1 / 1.1

    def multiply_range(start_position, multiplier):
        for p in range(start_position, len(res)):
            res[p][1] *= multiplier

    for m in re_attention.finditer(text):
        text = m.group(0)
        weight = m.group(1)

        if text.startswith("\\"):
            res.append([text[1:], 1.0])
        elif text == "(":
            round_brackets.append(len(res))
        elif text == "[":
            square_brackets.append(len(res))
        elif weight is not None and len(round_brackets) > 0:
            multiply_range(round_brackets.pop(), float(weight))
        elif text == ")" and len(round_brackets) > 0:
            multiply_range(round_brackets.pop(), round_bracket_multiplier)
        elif text == "]" and len(square_brackets) > 0:
            multiply_range(square_brackets.pop(), square_bracket_multiplier)
        else:
            res.append([text, 1.0])

    for pos in round_brackets:
        multiply_range(pos, round_bracket_multiplier)

    for pos in square_brackets:
        multiply_range(pos, square_bracket_multiplier)

    if len(res) == 0:
        res = [["", 1.0]]

    # merge runs of identical weights
    i = 0
    while i + 1 < len(res):
        if res[i][1] == res[i + 1][1]:
            res[i][0] += res[i + 1][0]
            res.pop(i + 1)
        else:
            i += 1

    return res


def get_prompts_with_weights(pipe, prompt: List[str], max_length: int):
    r"""
    Tokenize a list of prompts and return its tokens with weights of each token.

    No padding, starting or ending token is included.
    """
    tokens = []
    weights = []
    truncated = False
    for text in prompt:
        texts_and_weights = parse_prompt_attention(text)
        text_token = []
        text_weight = []
        for word, weight in texts_and_weights:
            # tokenize and discard the starting and the ending token
            token = pipe.tokenizer(word, return_tensors="np").input_ids[0, 1:-1]
            text_token += list(token)
            # copy the weight by length of token
            text_weight += [weight] * len(token)
            # stop if the text is too long (longer than truncation limit)
            if len(text_token) > max_length:
                truncated = True
                break
        # truncate
        if len(text_token) > max_length:
            truncated = True
            text_token = text_token[:max_length]
            text_weight = text_weight[:max_length]
        tokens.append(text_token)
        weights.append(text_weight)
    if truncated:
        logger.warning("Prompt was truncated. Try to shorten the prompt or increase max_embeddings_multiples")
    return tokens, weights


def pad_tokens_and_weights(tokens, weights, max_length, bos, eos, pad, no_boseos_middle=True, chunk_length=77):
    r"""
    Pad the tokens (with starting and ending tokens) and weights (with 1.0) to max_length.
    """
    max_embeddings_multiples = (max_length - 2) // (chunk_length - 2)
    weights_length = max_length if no_boseos_middle else max_embeddings_multiples * chunk_length
    for i in range(len(tokens)):
        tokens[i] = [bos] + tokens[i] + [pad] * (max_length - 1 - len(tokens[i]) - 1) + [eos]
        if no_boseos_middle:
            weights[i] = [1.0] + weights[i] + [1.0] * (max_length - 1 - len(weights[i]))
        else:
            w = []
            if len(weights[i]) == 0:
                w = [1.0] * weights_length
            else:
                for j in range(max_embeddings_multiples):
                    w.append(1.0)  # weight for starting token in this chunk
                    w += weights[i][j * (chunk_length - 2) : min(len(weights[i]), (j + 1) * (chunk_length - 2))]
                    w.append(1.0)  # weight for ending token in this chunk
                w += [1.0] * (weights_length - len(w))
            weights[i] = w[:]

    return tokens, weights


def get_unweighted_text_embeddings(
    pipe,
    text_input: np.array,
    chunk_length: int,
    no_boseos_middle: Optional[bool] = True,
):
    """
    When the length of tokens is a multiple of the capacity of the text encoder,
    it should be split into chunks and sent to the text encoder individually.
    """
    max_embeddings_multiples = (text_input.shape[1] - 2) // (chunk_length - 2)
    if max_embeddings_multiples > 1:
        text_embeddings = []
        for i in range(max_embeddings_multiples):
            # extract the i-th chunk
            text_input_chunk = text_input[:, i * (chunk_length - 2) : (i + 1) * (chunk_length - 2) + 2].copy()

            # cover the head and the tail by the starting and the ending tokens
            text_input_chunk[:, 0] = text_input[0, 0]
            text_input_chunk[:, -1] = text_input[0, -1]

            text_embedding = pipe.text_encoder(input_ids=text_input_chunk)[0]

            if no_boseos_middle:
                if i == 0:
                    # discard the ending token
                    text_embedding = text_embedding[:, :-1]
                elif i == max_embeddings_multiples - 1:
                    # discard the starting token
                    text_embedding = text_embedding[:, 1:]
                else:
                    # discard both starting and ending tokens
                    text_embedding = text_embedding[:, 1:-1]

            text_embeddings.append(text_embedding)
        text_embeddings = np.concatenate(text_embeddings, axis=1)
    else:
        text_embeddings = pipe.text_encoder(input_ids=text_input)[0]
    return text_embeddings


def get_weighted_text_embeddings(
    pipe,
    prompt: Union[str, List[str]],
    uncond_prompt: Optional[Union[str, List[str]]] = None,
    max_embeddings_multiples: Optional[int] = 4,
    no_boseos_middle: Optional[bool] = False,
    skip_parsing: Optional[bool] = False,
    skip_weighting: Optional[bool] = False,
    **kwargs,
):
    r"""
    Prompts can be assigned with local weights using brackets. For example,
    prompt 'A (very beautiful) masterpiece' highlights the words 'very beautiful',
    and the embedding tokens corresponding to the words get multiplied by a constant, 1.1.

    Also, to regularize of the embedding, the weighted embedding would be scaled to preserve the original mean.

    Args:
        pipe (`OnnxStableDiffusionPipeline`):
            Pipe to provide access to the tokenizer and the text encoder.
        prompt (`str` or `List[str]`):
            The prompt or prompts to guide the image generation.
        uncond_prompt (`str` or `List[str]`):
            The unconditional prompt or prompts for guide the image generation. If unconditional prompt
            is provided, the embeddings of prompt and uncond_prompt are concatenated.
        max_embeddings_multiples (`int`, *optional*, defaults to `1`):
            The max multiple length of prompt embeddings compared to the max output length of text encoder.
        no_boseos_middle (`bool`, *optional*, defaults to `False`):
            If the length of text token is multiples of the capacity of text encoder, whether reserve the starting and
            ending token in each of the chunk in the middle.
        skip_parsing (`bool`, *optional*, defaults to `False`):
            Skip the parsing of brackets.
        skip_weighting (`bool`, *optional*, defaults to `False`):
            Skip the weighting. When the parsing is skipped, it is forced True.
    """
    max_length = (pipe.tokenizer.model_max_length - 2) * max_embeddings_multiples + 2
    if isinstance(prompt, str):
        prompt = [prompt]

    if not skip_parsing:
        prompt_tokens, prompt_weights = get_prompts_with_weights(pipe, prompt, max_length - 2)
        if uncond_prompt is not None:
            if isinstance(uncond_prompt, str):
                uncond_prompt = [uncond_prompt]
            uncond_tokens, uncond_weights = get_prompts_with_weights(pipe, uncond_prompt, max_length - 2)
    else:
        prompt_tokens = [
            token[1:-1]
            for token in pipe.tokenizer(prompt, max_length=max_length, truncation=True, return_tensors="np").input_ids
        ]
        prompt_weights = [[1.0] * len(token) for token in prompt_tokens]
        if uncond_prompt is not None:
            if isinstance(uncond_prompt, str):
                uncond_prompt = [uncond_prompt]
            uncond_tokens = [
                token[1:-1]
                for token in pipe.tokenizer(
                    uncond_prompt,
                    max_length=max_length,
                    truncation=True,
                    return_tensors="np",
                ).input_ids
            ]
            uncond_weights = [[1.0] * len(token) for token in uncond_tokens]

    # round up the longest length of tokens to a multiple of (model_max_length - 2)
    max_length = max([len(token) for token in prompt_tokens])
    if uncond_prompt is not None:
        max_length = max(max_length, max([len(token) for token in uncond_tokens]))

    max_embeddings_multiples = min(
        max_embeddings_multiples,
        (max_length - 1) // (pipe.tokenizer.model_max_length - 2) + 1,
    )
    max_embeddings_multiples = max(1, max_embeddings_multiples)
    max_length = (pipe.tokenizer.model_max_length - 2) * max_embeddings_multiples + 2

    # pad the length of tokens and weights
    bos = pipe.tokenizer.bos_token_id
    eos = pipe.tokenizer.eos_token_id
    pad = getattr(pipe.tokenizer, "pad_token_id", eos)
    prompt_tokens, prompt_weights = pad_tokens_and_weights(
        prompt_tokens,
        prompt_weights,
        max_length,
        bos,
        eos,
        pad,
        no_boseos_middle=no_boseos_middle,
        chunk_length=pipe.tokenizer.model_max_length,
    )
    prompt_tokens = np.array(prompt_tokens, dtype=np.int32)
    if uncond_prompt is not None:
        uncond_tokens, uncond_weights = pad_tokens_and_weights(
            uncond_tokens,
            uncond_weights,
            max_length,
            bos,
            eos,
            pad,
            no_boseos_middle=no_boseos_middle,
            chunk_length=pipe.tokenizer.model_max_length,
        )
        uncond_tokens = np.array(uncond_tokens, dtype=np.int32)

    # get the embeddings
    text_embeddings = get_unweighted_text_embeddings(
        pipe,
        prompt_tokens,
        pipe.tokenizer.model_max_length,
        no_boseos_middle=no_boseos_middle,
    )
    prompt_weights = np.array(prompt_weights, dtype=text_embeddings.dtype)
    if uncond_prompt is not None:
        uncond_embeddings = get_unweighted_text_embeddings(
            pipe,
            uncond_tokens,
            pipe.tokenizer.model_max_length,
            no_boseos_middle=no_boseos_middle,
        )
        uncond_weights = np.array(uncond_weights, dtype=uncond_embeddings.dtype)

    # assign weights to the prompts and normalize in the sense of mean
    # TODO: should we normalize by chunk or in a whole (current implementation)?
    if (not skip_parsing) and (not skip_weighting):
        previous_mean = text_embeddings.mean(axis=(-2, -1))
        text_embeddings *= prompt_weights[:, :, None]
        text_embeddings *= (previous_mean / text_embeddings.mean(axis=(-2, -1)))[:, None, None]
        if uncond_prompt is not None:
            previous_mean = uncond_embeddings.mean(axis=(-2, -1))
            uncond_embeddings *= uncond_weights[:, :, None]
            uncond_embeddings *= (previous_mean / uncond_embeddings.mean(axis=(-2, -1)))[:, None, 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 uncond_prompt is not None:
        return text_embeddings, uncond_embeddings

    return text_embeddings


def preprocess_image(image):
    w, h = image.size
    w, h = (x - x % 32 for x in (w, h))  # resize to integer multiple of 32
    image = image.resize((w, h), resample=PIL_INTERPOLATION["lanczos"])
    image = np.array(image).astype(np.float32) / 255.0
    image = image[None].transpose(0, 3, 1, 2)
    return 2.0 * image - 1.0


def preprocess_mask(mask, scale_factor=8):
    mask = mask.convert("L")
    w, h = mask.size
    w, h = (x - x % 32 for x in (w, h))  # resize to integer multiple of 32
    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 = mask[None].transpose(0, 1, 2, 3)  # what does this step do?
    mask = 1 - mask  # repaint white, keep black
    return mask


class OnnxStableDiffusionLongPromptWeightingPipeline(OnnxStableDiffusionPipeline):
    r"""
    Pipeline for text-to-image generation using Stable Diffusion without tokens length limit, and support parsing
    weighting in prompt.

    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.)
    """

    if version.parse(version.parse(diffusers.__version__).base_version) >= version.parse("0.9.0"):

        def __init__(
            self,
            vae_encoder: OnnxRuntimeModel,
            vae_decoder: OnnxRuntimeModel,
            text_encoder: OnnxRuntimeModel,
            tokenizer: CLIPTokenizer,
            unet: OnnxRuntimeModel,
            scheduler: SchedulerMixin,
            safety_checker: OnnxRuntimeModel,
            feature_extractor: CLIPImageProcessor,
            requires_safety_checker: bool = True,
        ):
            super().__init__(
                vae_encoder=vae_encoder,
                vae_decoder=vae_decoder,
                text_encoder=text_encoder,
                tokenizer=tokenizer,
                unet=unet,
                scheduler=scheduler,
                safety_checker=safety_checker,
                feature_extractor=feature_extractor,
                requires_safety_checker=requires_safety_checker,
            )
            self.__init__additional__()

    else:

        def __init__(
            self,
            vae_encoder: OnnxRuntimeModel,
            vae_decoder: OnnxRuntimeModel,
            text_encoder: OnnxRuntimeModel,
            tokenizer: CLIPTokenizer,
            unet: OnnxRuntimeModel,
            scheduler: SchedulerMixin,
            safety_checker: OnnxRuntimeModel,
            feature_extractor: CLIPImageProcessor,
        ):
            super().__init__(
                vae_encoder=vae_encoder,
                vae_decoder=vae_decoder,
                text_encoder=text_encoder,
                tokenizer=tokenizer,
                unet=unet,
                scheduler=scheduler,
                safety_checker=safety_checker,
                feature_extractor=feature_extractor,
            )
            self.__init__additional__()

    def __init__additional__(self):
        self.unet.config.in_channels = 4
        self.vae_scale_factor = 8

    def _encode_prompt(
        self,
        prompt,
        num_images_per_prompt,
        do_classifier_free_guidance,
        negative_prompt,
        max_embeddings_multiples,
    ):
        r"""
        Encodes the prompt into text encoder hidden states.

        Args:
            prompt (`str` or `list(int)`):
                prompt to be encoded
            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]`):
                The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
                if `guidance_scale` is less than `1`).
            max_embeddings_multiples (`int`, *optional*, defaults to `3`):
                The max multiple length of prompt embeddings compared to the max output length of text encoder.
        """
        batch_size = len(prompt) if isinstance(prompt, list) else 1

        if negative_prompt is None:
            negative_prompt = [""] * batch_size
        elif isinstance(negative_prompt, str):
            negative_prompt = [negative_prompt] * batch_size
        if 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`."
            )

        text_embeddings, uncond_embeddings = get_weighted_text_embeddings(
            pipe=self,
            prompt=prompt,
            uncond_prompt=negative_prompt if do_classifier_free_guidance else None,
            max_embeddings_multiples=max_embeddings_multiples,
        )

        text_embeddings = text_embeddings.repeat(num_images_per_prompt, 0)
        if do_classifier_free_guidance:
            uncond_embeddings = uncond_embeddings.repeat(num_images_per_prompt, 0)
            text_embeddings = np.concatenate([uncond_embeddings, text_embeddings])

        return text_embeddings

    def check_inputs(self, prompt, height, width, strength, callback_steps):
        if 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 strength < 0 or strength > 1:
            raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")

        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)}."
            )

    def get_timesteps(self, num_inference_steps, strength, is_text2img):
        if is_text2img:
            return self.scheduler.timesteps, num_inference_steps
        else:
            # get the original timestep using init_timestep
            offset = self.scheduler.config.get("steps_offset", 0)
            init_timestep = int(num_inference_steps * strength) + offset
            init_timestep = min(init_timestep, num_inference_steps)

            t_start = max(num_inference_steps - init_timestep + offset, 0)
            timesteps = self.scheduler.timesteps[t_start:]
            return timesteps, num_inference_steps - t_start

    def run_safety_checker(self, image):
        if self.safety_checker is not None:
            safety_checker_input = self.feature_extractor(
                self.numpy_to_pil(image), return_tensors="np"
            ).pixel_values.astype(image.dtype)
            # There will throw an error if use safety_checker directly and batchsize>1
            images, has_nsfw_concept = [], []
            for i in range(image.shape[0]):
                image_i, has_nsfw_concept_i = self.safety_checker(
                    clip_input=safety_checker_input[i : i + 1], images=image[i : i + 1]
                )
                images.append(image_i)
                has_nsfw_concept.append(has_nsfw_concept_i[0])
            image = np.concatenate(images)
        else:
            has_nsfw_concept = None
        return image, has_nsfw_concept

    def decode_latents(self, latents):
        latents = 1 / 0.18215 * latents
        # image = self.vae_decoder(latent_sample=latents)[0]
        # it seems likes there is a strange result for using half-precision vae decoder if batchsize>1
        image = np.concatenate(
            [self.vae_decoder(latent_sample=latents[i : i + 1])[0] for i in range(latents.shape[0])]
        )
        image = np.clip(image / 2 + 0.5, 0, 1)
        image = image.transpose((0, 2, 3, 1))
        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 prepare_latents(self, image, timestep, batch_size, height, width, dtype, generator, latents=None):
        if image is None:
            shape = (
                batch_size,
                self.unet.config.in_channels,
                height // self.vae_scale_factor,
                width // self.vae_scale_factor,
            )

            if latents is None:
                latents = torch.randn(shape, generator=generator, device="cpu").numpy().astype(dtype)
            else:
                if latents.shape != shape:
                    raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")

            # scale the initial noise by the standard deviation required by the scheduler
            latents = (torch.from_numpy(latents) * self.scheduler.init_noise_sigma).numpy()
            return latents, None, None
        else:
            init_latents = self.vae_encoder(sample=image)[0]
            init_latents = 0.18215 * init_latents
            init_latents = np.concatenate([init_latents] * batch_size, axis=0)
            init_latents_orig = init_latents
            shape = init_latents.shape

            # add noise to latents using the timesteps
            noise = torch.randn(shape, generator=generator, device="cpu").numpy().astype(dtype)
            latents = self.scheduler.add_noise(
                torch.from_numpy(init_latents), torch.from_numpy(noise), timestep
            ).numpy()
            return latents, init_latents_orig, noise

    @torch.no_grad()
    def __call__(
        self,
        prompt: Union[str, List[str]],
        negative_prompt: Optional[Union[str, List[str]]] = None,
        image: Union[np.ndarray, PIL.Image.Image] = None,
        mask_image: Union[np.ndarray, PIL.Image.Image] = None,
        height: int = 512,
        width: int = 512,
        num_inference_steps: int = 50,
        guidance_scale: float = 7.5,
        strength: float = 0.8,
        num_images_per_prompt: Optional[int] = 1,
        eta: float = 0.0,
        generator: Optional[torch.Generator] = None,
        latents: Optional[np.ndarray] = None,
        max_embeddings_multiples: Optional[int] = 3,
        output_type: Optional[str] = "pil",
        return_dict: bool = True,
        callback: Optional[Callable[[int, int, np.ndarray], None]] = None,
        is_cancelled_callback: Optional[Callable[[], bool]] = None,
        callback_steps: int = 1,
        **kwargs,
    ):
        r"""
        Function invoked when calling the pipeline for generation.

        Args:
            prompt (`str` or `List[str]`):
                The prompt or prompts to guide the image generation.
            negative_prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
                if `guidance_scale` is less than `1`).
            image (`np.ndarray` or `PIL.Image.Image`):
                `Image`, or tensor representing an image batch, that will be used as the starting point for the
                process.
            mask_image (`np.ndarray` or `PIL.Image.Image`):
                `Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be
                replaced by noise and therefore repainted, while black pixels will be preserved. If `mask_image` is a
                PIL image, it will be converted to a single channel (luminance) before use. If it's a tensor, it should
                contain one color channel (L) instead of 3, so the expected shape would be `(B, H, W, 1)`.
            height (`int`, *optional*, defaults to 512):
                The height in pixels of the generated image.
            width (`int`, *optional*, defaults to 512):
                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.
            strength (`float`, *optional*, defaults to 0.8):
                Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1.
                `image` will be used as a starting point, adding more noise to it the larger the `strength`. The
                number of denoising steps depends on the amount of noise initially added. When `strength` is 1, added
                noise will be maximum and the denoising process will run for the full number of iterations specified in
                `num_inference_steps`. A value of 1, therefore, essentially ignores `image`.
            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`, *optional*):
                A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation
                deterministic.
            latents (`np.ndarray`, *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`.
            max_embeddings_multiples (`int`, *optional*, defaults to `3`):
                The max multiple length of prompt embeddings compared to the max output length of text encoder.
            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: np.ndarray)`.
            is_cancelled_callback (`Callable`, *optional*):
                A function that will be called every `callback_steps` steps during inference. If the function returns
                `True`, the inference will be cancelled.
            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.

        Returns:
            `None` if cancelled by `is_cancelled_callback`,
            [`~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

        # 1. Check inputs. Raise error if not correct
        self.check_inputs(prompt, height, width, strength, callback_steps)

        # 2. Define call parameters
        batch_size = 1 if isinstance(prompt, str) else len(prompt)
        # 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_embeddings = self._encode_prompt(
            prompt,
            num_images_per_prompt,
            do_classifier_free_guidance,
            negative_prompt,
            max_embeddings_multiples,
        )
        dtype = text_embeddings.dtype

        # 4. Preprocess image and mask
        if isinstance(image, PIL.Image.Image):
            image = preprocess_image(image)
        if image is not None:
            image = image.astype(dtype)
        if isinstance(mask_image, PIL.Image.Image):
            mask_image = preprocess_mask(mask_image, self.vae_scale_factor)
        if mask_image is not None:
            mask = mask_image.astype(dtype)
            mask = np.concatenate([mask] * batch_size * num_images_per_prompt)
        else:
            mask = None

        # 5. set timesteps
        self.scheduler.set_timesteps(num_inference_steps)
        timestep_dtype = next(
            (input.type for input in self.unet.model.get_inputs() if input.name == "timestep"), "tensor(float)"
        )
        timestep_dtype = ORT_TO_NP_TYPE[timestep_dtype]
        timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, image is None)
        latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)

        # 6. Prepare latent variables
        latents, init_latents_orig, noise = self.prepare_latents(
            image,
            latent_timestep,
            batch_size * num_images_per_prompt,
            height,
            width,
            dtype,
            generator,
            latents,
        )

        # 7. 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)

        # 8. Denoising loop
        for i, t in enumerate(self.progress_bar(timesteps)):
            # expand the latents if we are doing classifier free guidance
            latent_model_input = np.concatenate([latents] * 2) if do_classifier_free_guidance else latents
            latent_model_input = self.scheduler.scale_model_input(torch.from_numpy(latent_model_input), t)
            latent_model_input = latent_model_input.numpy()

            # predict the noise residual
            noise_pred = self.unet(
                sample=latent_model_input,
                timestep=np.array([t], dtype=timestep_dtype),
                encoder_hidden_states=text_embeddings,
            )
            noise_pred = noise_pred[0]

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

            # compute the previous noisy sample x_t -> x_t-1
            scheduler_output = self.scheduler.step(
                torch.from_numpy(noise_pred), t, torch.from_numpy(latents), **extra_step_kwargs
            )
            latents = scheduler_output.prev_sample.numpy()

            if mask is not None:
                # masking
                init_latents_proper = self.scheduler.add_noise(
                    torch.from_numpy(init_latents_orig),
                    torch.from_numpy(noise),
                    t,
                ).numpy()
                latents = (init_latents_proper * mask) + (latents * (1 - mask))

            # call the callback, if provided
            if i % callback_steps == 0:
                if callback is not None:
                    step_idx = i // getattr(self.scheduler, "order", 1)
                    callback(step_idx, t, latents)
                if is_cancelled_callback is not None and is_cancelled_callback():
                    return None

        # 9. Post-processing
        image = self.decode_latents(latents)

        # 10. Run safety checker
        image, has_nsfw_concept = self.run_safety_checker(image)

        # 11. Convert to PIL
        if output_type == "pil":
            image = self.numpy_to_pil(image)

        if not return_dict:
            return image, has_nsfw_concept

        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)

    def text2img(
        self,
        prompt: Union[str, List[str]],
        negative_prompt: Optional[Union[str, List[str]]] = None,
        height: int = 512,
        width: int = 512,
        num_inference_steps: int = 50,
        guidance_scale: float = 7.5,
        num_images_per_prompt: Optional[int] = 1,
        eta: float = 0.0,
        generator: Optional[torch.Generator] = None,
        latents: Optional[np.ndarray] = None,
        max_embeddings_multiples: Optional[int] = 3,
        output_type: Optional[str] = "pil",
        return_dict: bool = True,
        callback: Optional[Callable[[int, int, np.ndarray], None]] = None,
        callback_steps: int = 1,
        **kwargs,
    ):
        r"""
        Function for text-to-image generation.
        Args:
            prompt (`str` or `List[str]`):
                The prompt or prompts to guide the image generation.
            negative_prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
                if `guidance_scale` is less than `1`).
            height (`int`, *optional*, defaults to 512):
                The height in pixels of the generated image.
            width (`int`, *optional*, defaults to 512):
                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.
            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`, *optional*):
                A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation
                deterministic.
            latents (`np.ndarray`, *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`.
            max_embeddings_multiples (`int`, *optional*, defaults to `3`):
                The max multiple length of prompt embeddings compared to the max output length of text encoder.
            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: np.ndarray)`.
            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.
        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`.
        """
        return self.__call__(
            prompt=prompt,
            negative_prompt=negative_prompt,
            height=height,
            width=width,
            num_inference_steps=num_inference_steps,
            guidance_scale=guidance_scale,
            num_images_per_prompt=num_images_per_prompt,
            eta=eta,
            generator=generator,
            latents=latents,
            max_embeddings_multiples=max_embeddings_multiples,
            output_type=output_type,
            return_dict=return_dict,
            callback=callback,
            callback_steps=callback_steps,
            **kwargs,
        )

    def img2img(
        self,
        image: Union[np.ndarray, PIL.Image.Image],
        prompt: Union[str, List[str]],
        negative_prompt: Optional[Union[str, List[str]]] = None,
        strength: float = 0.8,
        num_inference_steps: Optional[int] = 50,
        guidance_scale: Optional[float] = 7.5,
        num_images_per_prompt: Optional[int] = 1,
        eta: Optional[float] = 0.0,
        generator: Optional[torch.Generator] = None,
        max_embeddings_multiples: Optional[int] = 3,
        output_type: Optional[str] = "pil",
        return_dict: bool = True,
        callback: Optional[Callable[[int, int, np.ndarray], None]] = None,
        callback_steps: int = 1,
        **kwargs,
    ):
        r"""
        Function for image-to-image generation.
        Args:
            image (`np.ndarray` or `PIL.Image.Image`):
                `Image`, or ndarray representing an image batch, that will be used as the starting point for the
                process.
            prompt (`str` or `List[str]`):
                The prompt or prompts to guide the image generation.
            negative_prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
                if `guidance_scale` is less than `1`).
            strength (`float`, *optional*, defaults to 0.8):
                Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1.
                `image` will be used as a starting point, adding more noise to it the larger the `strength`. The
                number of denoising steps depends on the amount of noise initially added. When `strength` is 1, added
                noise will be maximum and the denoising process will run for the full number of iterations specified in
                `num_inference_steps`. A value of 1, therefore, essentially ignores `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. This parameter will be modulated by `strength`.
            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.
            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`, *optional*):
                A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation
                deterministic.
            max_embeddings_multiples (`int`, *optional*, defaults to `3`):
                The max multiple length of prompt embeddings compared to the max output length of text encoder.
            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: np.ndarray)`.
            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.
        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`.
        """
        return self.__call__(
            prompt=prompt,
            negative_prompt=negative_prompt,
            image=image,
            num_inference_steps=num_inference_steps,
            guidance_scale=guidance_scale,
            strength=strength,
            num_images_per_prompt=num_images_per_prompt,
            eta=eta,
            generator=generator,
            max_embeddings_multiples=max_embeddings_multiples,
            output_type=output_type,
            return_dict=return_dict,
            callback=callback,
            callback_steps=callback_steps,
            **kwargs,
        )

    def inpaint(
        self,
        image: Union[np.ndarray, PIL.Image.Image],
        mask_image: Union[np.ndarray, PIL.Image.Image],
        prompt: Union[str, List[str]],
        negative_prompt: Optional[Union[str, List[str]]] = None,
        strength: float = 0.8,
        num_inference_steps: Optional[int] = 50,
        guidance_scale: Optional[float] = 7.5,
        num_images_per_prompt: Optional[int] = 1,
        eta: Optional[float] = 0.0,
        generator: Optional[torch.Generator] = None,
        max_embeddings_multiples: Optional[int] = 3,
        output_type: Optional[str] = "pil",
        return_dict: bool = True,
        callback: Optional[Callable[[int, int, np.ndarray], None]] = None,
        callback_steps: int = 1,
        **kwargs,
    ):
        r"""
        Function for inpaint.
        Args:
            image (`np.ndarray` or `PIL.Image.Image`):
                `Image`, or tensor representing an image batch, that will be used as the starting point for the
                process. This is the image whose masked region will be inpainted.
            mask_image (`np.ndarray` or `PIL.Image.Image`):
                `Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be
                replaced by noise and therefore repainted, while black pixels will be preserved. If `mask_image` is a
                PIL image, it will be converted to a single channel (luminance) before use. If it's a tensor, it should
                contain one color channel (L) instead of 3, so the expected shape would be `(B, H, W, 1)`.
            prompt (`str` or `List[str]`):
                The prompt or prompts to guide the image generation.
            negative_prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
                if `guidance_scale` is less than `1`).
            strength (`float`, *optional*, defaults to 0.8):
                Conceptually, indicates how much to inpaint the masked area. Must be between 0 and 1. When `strength`
                is 1, the denoising process will be run on the masked area for the full number of iterations specified
                in `num_inference_steps`. `image` will be used as a reference for the masked area, adding more
                noise to that region the larger the `strength`. If `strength` is 0, no inpainting will occur.
            num_inference_steps (`int`, *optional*, defaults to 50):
                The reference number of denoising steps. More denoising steps usually lead to a higher quality image at
                the expense of slower inference. This parameter will be modulated by `strength`, as explained above.
            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.
            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`, *optional*):
                A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation
                deterministic.
            max_embeddings_multiples (`int`, *optional*, defaults to `3`):
                The max multiple length of prompt embeddings compared to the max output length of text encoder.
            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: np.ndarray)`.
            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.
        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`.
        """
        return self.__call__(
            prompt=prompt,
            negative_prompt=negative_prompt,
            image=image,
            mask_image=mask_image,
            num_inference_steps=num_inference_steps,
            guidance_scale=guidance_scale,
            strength=strength,
            num_images_per_prompt=num_images_per_prompt,
            eta=eta,
            generator=generator,
            max_embeddings_multiples=max_embeddings_multiples,
            output_type=output_type,
            return_dict=return_dict,
            callback=callback,
            callback_steps=callback_steps,
            **kwargs,
        )