diffusers/pipelines/stable_diffusion/pipeline_stable_unclip.py

# Copyright 2023 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

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

import torch
from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer
from transformers.models.clip.modeling_clip import CLIPTextModelOutput

from ...models import AutoencoderKL, PriorTransformer, UNet2DConditionModel
from ...models.embeddings import get_timestep_embedding
from ...schedulers import KarrasDiffusionSchedulers
from ...utils import is_accelerate_available, logging, randn_tensor, replace_example_docstring
from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
from .stable_unclip_image_normalizer import StableUnCLIPImageNormalizer


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

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

        >>> pipe = StableUnCLIPPipeline.from_pretrained(
        ...     "fusing/stable-unclip-2-1-l", torch_dtype=torch.float16
        ... )  # TODO update model path
        >>> pipe = pipe.to("cuda")

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


class StableUnCLIPPipeline(DiffusionPipeline):
    """
    Pipeline for text-to-image generation using stable unCLIP.

    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.)

    Args:
        prior_tokenizer ([`CLIPTokenizer`]):
            Tokenizer of class
            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
        prior_text_encoder ([`CLIPTextModelWithProjection`]):
            Frozen text-encoder.
        prior ([`PriorTransformer`]):
            The canonincal unCLIP prior to approximate the image embedding from the text embedding.
        prior_scheduler ([`KarrasDiffusionSchedulers`]):
            Scheduler used in the prior denoising process.
        image_normalizer ([`StableUnCLIPImageNormalizer`]):
            Used to normalize the predicted image embeddings before the noise is applied and un-normalize the image
            embeddings after the noise has been applied.
        image_noising_scheduler ([`KarrasDiffusionSchedulers`]):
            Noise schedule for adding noise to the predicted image embeddings. The amount of noise to add is determined
            by `noise_level` in `StableUnCLIPPipeline.__call__`.
        tokenizer (`CLIPTokenizer`):
            Tokenizer of class
            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
        text_encoder ([`CLIPTextModel`]):
            Frozen text-encoder.
        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
        scheduler ([`KarrasDiffusionSchedulers`]):
            A scheduler to be used in combination with `unet` to denoise the encoded image latents.
        vae ([`AutoencoderKL`]):
            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
    """

    # prior components
    prior_tokenizer: CLIPTokenizer
    prior_text_encoder: CLIPTextModelWithProjection
    prior: PriorTransformer
    prior_scheduler: KarrasDiffusionSchedulers

    # image noising components
    image_normalizer: StableUnCLIPImageNormalizer
    image_noising_scheduler: KarrasDiffusionSchedulers

    # regular denoising components
    tokenizer: CLIPTokenizer
    text_encoder: CLIPTextModel
    unet: UNet2DConditionModel
    scheduler: KarrasDiffusionSchedulers

    vae: AutoencoderKL

    def __init__(
        self,
        # prior components
        prior_tokenizer: CLIPTokenizer,
        prior_text_encoder: CLIPTextModelWithProjection,
        prior: PriorTransformer,
        prior_scheduler: KarrasDiffusionSchedulers,
        # image noising components
        image_normalizer: StableUnCLIPImageNormalizer,
        image_noising_scheduler: KarrasDiffusionSchedulers,
        # regular denoising components
        tokenizer: CLIPTokenizer,
        text_encoder: CLIPTextModelWithProjection,
        unet: UNet2DConditionModel,
        scheduler: KarrasDiffusionSchedulers,
        # vae
        vae: AutoencoderKL,
    ):
        super().__init__()

        self.register_modules(
            prior_tokenizer=prior_tokenizer,
            prior_text_encoder=prior_text_encoder,
            prior=prior,
            prior_scheduler=prior_scheduler,
            image_normalizer=image_normalizer,
            image_noising_scheduler=image_noising_scheduler,
            tokenizer=tokenizer,
            text_encoder=text_encoder,
            unet=unet,
            scheduler=scheduler,
            vae=vae,
        )

        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
    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()

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

    def enable_sequential_cpu_offload(self, gpu_id=0):
        r"""
        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, the pipeline's
        models have their state dicts saved to CPU and then are moved to a `torch.device('meta') and loaded to GPU only
        when their specific submodule has its `forward` method called.
        """
        if is_accelerate_available():
            from accelerate import cpu_offload
        else:
            raise ImportError("Please install accelerate via `pip install accelerate`")

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

        # TODO: self.prior.post_process_latents and self.image_noiser.{scale,unscale} are not covered by the offload hooks, so they fails if added to the list
        models = [
            self.prior_text_encoder,
            self.text_encoder,
            self.unet,
            self.vae,
        ]
        for cpu_offloaded_model in models:
            if cpu_offloaded_model is not None:
                cpu_offload(cpu_offloaded_model, device)

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

    # Copied from diffusers.pipelines.unclip.pipeline_unclip.UnCLIPPipeline._encode_prompt with _encode_prompt->_encode_prior_prompt, tokenizer->prior_tokenizer, text_encoder->prior_text_encoder
    def _encode_prior_prompt(
        self,
        prompt,
        device,
        num_images_per_prompt,
        do_classifier_free_guidance,
        text_model_output: Optional[Union[CLIPTextModelOutput, Tuple]] = None,
        text_attention_mask: Optional[torch.Tensor] = None,
    ):
        if text_model_output is None:
            batch_size = len(prompt) if isinstance(prompt, list) else 1
            # get prompt text embeddings
            text_inputs = self.prior_tokenizer(
                prompt,
                padding="max_length",
                max_length=self.prior_tokenizer.model_max_length,
                truncation=True,
                return_tensors="pt",
            )
            text_input_ids = text_inputs.input_ids
            text_mask = text_inputs.attention_mask.bool().to(device)

            untruncated_ids = self.prior_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.prior_tokenizer.batch_decode(
                    untruncated_ids[:, self.prior_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.prior_tokenizer.model_max_length} tokens: {removed_text}"
                )
                text_input_ids = text_input_ids[:, : self.prior_tokenizer.model_max_length]

            prior_text_encoder_output = self.prior_text_encoder(text_input_ids.to(device))

            prompt_embeds = prior_text_encoder_output.text_embeds
            prior_text_encoder_hidden_states = prior_text_encoder_output.last_hidden_state

        else:
            batch_size = text_model_output[0].shape[0]
            prompt_embeds, prior_text_encoder_hidden_states = text_model_output[0], text_model_output[1]
            text_mask = text_attention_mask

        prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0)
        prior_text_encoder_hidden_states = prior_text_encoder_hidden_states.repeat_interleave(
            num_images_per_prompt, dim=0
        )
        text_mask = text_mask.repeat_interleave(num_images_per_prompt, dim=0)

        if do_classifier_free_guidance:
            uncond_tokens = [""] * batch_size

            uncond_input = self.prior_tokenizer(
                uncond_tokens,
                padding="max_length",
                max_length=self.prior_tokenizer.model_max_length,
                truncation=True,
                return_tensors="pt",
            )
            uncond_text_mask = uncond_input.attention_mask.bool().to(device)
            negative_prompt_embeds_prior_text_encoder_output = self.prior_text_encoder(
                uncond_input.input_ids.to(device)
            )

            negative_prompt_embeds = negative_prompt_embeds_prior_text_encoder_output.text_embeds
            uncond_prior_text_encoder_hidden_states = (
                negative_prompt_embeds_prior_text_encoder_output.last_hidden_state
            )

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

            seq_len = uncond_prior_text_encoder_hidden_states.shape[1]
            uncond_prior_text_encoder_hidden_states = uncond_prior_text_encoder_hidden_states.repeat(
                1, num_images_per_prompt, 1
            )
            uncond_prior_text_encoder_hidden_states = uncond_prior_text_encoder_hidden_states.view(
                batch_size * num_images_per_prompt, seq_len, -1
            )
            uncond_text_mask = uncond_text_mask.repeat_interleave(num_images_per_prompt, dim=0)

            # done duplicates

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

            text_mask = torch.cat([uncond_text_mask, text_mask])

        return prompt_embeds, prior_text_encoder_hidden_states, text_mask

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
    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,
    ):
        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. 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.
        """
        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:
            text_inputs = self.tokenizer(
                prompt,
                padding="max_length",
                max_length=self.tokenizer.model_max_length,
                truncation=True,
                return_tensors="pt",
            )
            text_input_ids = text_inputs.input_ids
            untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids

            if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
                text_input_ids, untruncated_ids
            ):
                removed_text = self.tokenizer.batch_decode(
                    untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
                )
                logger.warning(
                    "The following part of your input was truncated because CLIP can only handle sequences up to"
                    f" {self.tokenizer.model_max_length} tokens: {removed_text}"
                )

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

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

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

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

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

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

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

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

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

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

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

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

        return prompt_embeds

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
    def decode_latents(self, latents):
        latents = 1 / self.vae.config.scaling_factor * latents
        image = self.vae.decode(latents).sample
        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

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs with prepare_extra_step_kwargs->prepare_prior_extra_step_kwargs, scheduler->prior_scheduler
    def prepare_prior_extra_step_kwargs(self, generator, eta):
        # prepare extra kwargs for the prior_scheduler step, since not all prior_schedulers have the same signature
        # eta (η) is only used with the DDIMScheduler, it will be ignored for other prior_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.prior_scheduler.step).parameters.keys())
        extra_step_kwargs = {}
        if accepts_eta:
            extra_step_kwargs["eta"] = eta

        # check if the prior_scheduler accepts generator
        accepts_generator = "generator" in set(inspect.signature(self.prior_scheduler.step).parameters.keys())
        if accepts_generator:
            extra_step_kwargs["generator"] = generator
        return extra_step_kwargs

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
    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,
        noise_level,
        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(
                "Provide either `prompt` or `prompt_embeds`. Please make sure to define only one of the two."
            )

        if prompt is None and prompt_embeds is None:
            raise ValueError(
                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
            )

        if 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(
                "Provide either `negative_prompt` or `negative_prompt_embeds`. Cannot leave both `negative_prompt` and `negative_prompt_embeds` undefined."
            )

        if prompt is not None and negative_prompt is not None:
            if 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)}."
                )

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

        if noise_level < 0 or noise_level >= self.image_noising_scheduler.config.num_train_timesteps:
            raise ValueError(
                f"`noise_level` must be between 0 and {self.image_noising_scheduler.config.num_train_timesteps - 1}, inclusive."
            )

    # Copied from diffusers.pipelines.unclip.pipeline_unclip.UnCLIPPipeline.prepare_latents
    def prepare_latents(self, shape, dtype, device, generator, latents, scheduler):
        if latents is None:
            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
        else:
            if latents.shape != shape:
                raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
            latents = latents.to(device)

        latents = latents * scheduler.init_noise_sigma
        return latents

    def noise_image_embeddings(
        self,
        image_embeds: torch.Tensor,
        noise_level: int,
        noise: Optional[torch.FloatTensor] = None,
        generator: Optional[torch.Generator] = None,
    ):
        """
        Add noise to the image embeddings. The amount of noise is controlled by a `noise_level` input. A higher
        `noise_level` increases the variance in the final un-noised images.

        The noise is applied in two ways
        1. A noise schedule is applied directly to the embeddings
        2. A vector of sinusoidal time embeddings are appended to the output.

        In both cases, the amount of noise is controlled by the same `noise_level`.

        The embeddings are normalized before the noise is applied and un-normalized after the noise is applied.
        """
        if noise is None:
            noise = randn_tensor(
                image_embeds.shape, generator=generator, device=image_embeds.device, dtype=image_embeds.dtype
            )

        noise_level = torch.tensor([noise_level] * image_embeds.shape[0], device=image_embeds.device)

        image_embeds = self.image_normalizer.scale(image_embeds)

        image_embeds = self.image_noising_scheduler.add_noise(image_embeds, timesteps=noise_level, noise=noise)

        image_embeds = self.image_normalizer.unscale(image_embeds)

        noise_level = get_timestep_embedding(
            timesteps=noise_level, embedding_dim=image_embeds.shape[-1], flip_sin_to_cos=True, downscale_freq_shift=0
        )

        # `get_timestep_embeddings` does not contain any weights and will always return f32 tensors,
        # but we might actually be running in fp16. so we need to cast here.
        # there might be better ways to encapsulate this.
        noise_level = noise_level.to(image_embeds.dtype)

        image_embeds = torch.cat((image_embeds, noise_level), 1)

        return image_embeds

    @torch.no_grad()
    @replace_example_docstring(EXAMPLE_DOC_STRING)
    def __call__(
        self,
        # regular denoising process args
        prompt: Optional[Union[str, List[str]]] = None,
        height: Optional[int] = None,
        width: Optional[int] = None,
        num_inference_steps: int = 20,
        guidance_scale: float = 10.0,
        negative_prompt: Optional[Union[str, List[str]]] = None,
        num_images_per_prompt: Optional[int] = 1,
        eta: float = 0.0,
        generator: Optional[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,
        noise_level: int = 0,
        # prior args
        prior_num_inference_steps: int = 25,
        prior_guidance_scale: float = 4.0,
        prior_latents: Optional[torch.FloatTensor] = None,
    ):
        """
        Function invoked when calling the pipeline for generation.

        Args:
            prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
                instead.
            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                The height in pixels of the generated image.
            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                The width in pixels of the generated image.
            num_inference_steps (`int`, *optional*, defaults to 20):
                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 10.0):
                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
                `guidance_scale` is defined as `w` of equation 2. of [Imagen
                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
                usually at the expense of lower image quality.
            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. If not defined, one has to pass `negative_prompt_embeds`. instead.
                Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`).
            num_images_per_prompt (`int`, *optional*, defaults to 1):
                The number of images to generate per prompt.
            eta (`float`, *optional*, defaults to 0.0):
                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
                [`schedulers.DDIMScheduler`], will be ignored for others.
            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
                to make generation deterministic.
            latents (`torch.FloatTensor`, *optional*):
                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
                tensor will ge generated by sampling using the supplied random `generator`.
            prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
                provided, text embeddings will be generated from `prompt` input argument.
            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
                argument.
            output_type (`str`, *optional*, defaults to `"pil"`):
                The output format of the generate image. Choose between
                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
                plain tuple.
            callback (`Callable`, *optional*):
                A function that will be called every `callback_steps` steps during inference. The function will be
                called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
            callback_steps (`int`, *optional*, defaults to 1):
                The frequency at which the `callback` function will be called. If not specified, the callback will be
                called at every step.
            cross_attention_kwargs (`dict`, *optional*):
                A kwargs dictionary that if specified is passed along to the `AttnProcessor` as defined under
                `self.processor` in
                [diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py).
            noise_level (`int`, *optional*, defaults to `0`):
                The amount of noise to add to the image embeddings. A higher `noise_level` increases the variance in
                the final un-noised images. See `StableUnCLIPPipeline.noise_image_embeddings` for details.
            prior_num_inference_steps (`int`, *optional*, defaults to 25):
                The number of denoising steps in the prior denoising process. More denoising steps usually lead to a
                higher quality image at the expense of slower inference.
            prior_guidance_scale (`float`, *optional*, defaults to 4.0):
                Guidance scale for the prior denoising process as defined in [Classifier-Free Diffusion
                Guidance](https://arxiv.org/abs/2207.12598). `prior_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.
            prior_latents (`torch.FloatTensor`, *optional*):
                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
                embedding generation in the prior denoising process. 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`.

        Examples:

        Returns:
            [`~pipelines.ImagePipelineOutput`] or `tuple`: [`~ pipeline_utils.ImagePipelineOutput`] if `return_dict` is
            True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated images.
        """
        # 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=prompt,
            height=height,
            width=width,
            callback_steps=callback_steps,
            noise_level=noise_level,
            negative_prompt=negative_prompt,
            prompt_embeds=prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds,
        )

        # 2. Define call parameters
        if prompt is not None and isinstance(prompt, str):
            batch_size = 1
        elif prompt is not None and isinstance(prompt, list):
            batch_size = len(prompt)
        else:
            batch_size = prompt_embeds.shape[0]

        batch_size = batch_size * num_images_per_prompt

        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.
        prior_do_classifier_free_guidance = prior_guidance_scale > 1.0

        # 3. Encode input prompt
        prior_prompt_embeds, prior_text_encoder_hidden_states, prior_text_mask = self._encode_prior_prompt(
            prompt=prompt,
            device=device,
            num_images_per_prompt=num_images_per_prompt,
            do_classifier_free_guidance=prior_do_classifier_free_guidance,
        )

        # 4. Prepare prior timesteps
        self.prior_scheduler.set_timesteps(prior_num_inference_steps, device=device)
        prior_timesteps_tensor = self.prior_scheduler.timesteps

        # 5. Prepare prior latent variables
        embedding_dim = self.prior.config.embedding_dim
        prior_latents = self.prepare_latents(
            (batch_size, embedding_dim),
            prior_prompt_embeds.dtype,
            device,
            generator,
            prior_latents,
            self.prior_scheduler,
        )

        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
        prior_extra_step_kwargs = self.prepare_prior_extra_step_kwargs(generator, eta)

        # 7. Prior denoising loop
        for i, t in enumerate(self.progress_bar(prior_timesteps_tensor)):
            # expand the latents if we are doing classifier free guidance
            latent_model_input = torch.cat([prior_latents] * 2) if prior_do_classifier_free_guidance else prior_latents
            latent_model_input = self.prior_scheduler.scale_model_input(latent_model_input, t)

            predicted_image_embedding = self.prior(
                latent_model_input,
                timestep=t,
                proj_embedding=prior_prompt_embeds,
                encoder_hidden_states=prior_text_encoder_hidden_states,
                attention_mask=prior_text_mask,
            ).predicted_image_embedding

            if prior_do_classifier_free_guidance:
                predicted_image_embedding_uncond, predicted_image_embedding_text = predicted_image_embedding.chunk(2)
                predicted_image_embedding = predicted_image_embedding_uncond + prior_guidance_scale * (
                    predicted_image_embedding_text - predicted_image_embedding_uncond
                )

            prior_latents = self.prior_scheduler.step(
                predicted_image_embedding,
                timestep=t,
                sample=prior_latents,
                **prior_extra_step_kwargs,
            ).prev_sample

            if callback is not None and i % callback_steps == 0:
                callback(i, t, prior_latents)

        prior_latents = self.prior.post_process_latents(prior_latents)

        image_embeds = prior_latents

        # done prior

        # 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

        # 8. Encode input prompt
        prompt_embeds = 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,
        )

        # 9. Prepare image embeddings
        image_embeds = self.noise_image_embeddings(
            image_embeds=image_embeds,
            noise_level=noise_level,
            generator=generator,
        )

        if do_classifier_free_guidance:
            negative_prompt_embeds = torch.zeros_like(image_embeds)

            # 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
            image_embeds = torch.cat([negative_prompt_embeds, image_embeds])

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

        # 11. Prepare latent variables
        num_channels_latents = self.unet.in_channels
        shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
        latents = self.prepare_latents(
            shape=shape,
            dtype=prompt_embeds.dtype,
            device=device,
            generator=generator,
            latents=latents,
            scheduler=self.scheduler,
        )

        # 12. 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)

        # 13. Denoising loop
        for i, t in enumerate(self.progress_bar(timesteps)):
            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)

            # predict the noise residual
            noise_pred = self.unet(
                latent_model_input,
                t,
                encoder_hidden_states=prompt_embeds,
                class_labels=image_embeds,
                cross_attention_kwargs=cross_attention_kwargs,
            ).sample

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

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

            if callback is not None and i % callback_steps == 0:
                callback(i, t, latents)

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

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

        if not return_dict:
            return (image,)

        return ImagePipelineOutput(images=image)