ai-toolkit/toolkit/pipelines.py

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

import torch
from diffusers import StableDiffusionXLPipeline, StableDiffusionPipeline
from diffusers.image_processor import VaeImageProcessor
from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl import rescale_noise_cfg

from typing import TYPE_CHECKING

from diffusers.pipelines.stable_diffusion_xl.watermark import StableDiffusionXLWatermarker

if TYPE_CHECKING:
    from diffusers import AutoencoderKL, UNet2DConditionModel
    from transformers import CLIPTextModel, CLIPTokenizer, CLIPTextModelWithProjection
    from diffusers.schedulers import KarrasDiffusionSchedulers


class FakeWatermarker(StableDiffusionXLWatermarker):
    def __init__(self):
        super().__init__()

    def apply_watermark(self, image):
        return image


class HackedStableDiffusionXLPipeline(StableDiffusionXLPipeline):
    def __init__(
            self,
            vae: 'AutoencoderKL',
            text_encoder: 'CLIPTextModel',
            text_encoder_2: 'CLIPTextModelWithProjection',
            tokenizer: 'CLIPTokenizer',
            tokenizer_2: 'CLIPTokenizer',
            unet: 'UNet2DConditionModel',
            scheduler: 'KarrasDiffusionSchedulers',
            force_zeros_for_empty_prompt: bool = True,
            add_watermarker: bool = False,
    ):
        # call parents parent super skipping parent
        super(StableDiffusionXLPipeline, self).__init__()

        self.register_modules(
            vae=vae,
            text_encoder=text_encoder,
            text_encoder_2=text_encoder_2,
            tokenizer=tokenizer,
            tokenizer_2=tokenizer_2,
            unet=unet,
            scheduler=scheduler,
        )
        self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
        self.default_sample_size = 1024

        self.watermark = FakeWatermarker()

    def _get_add_time_ids(self, original_size, crops_coords_top_left, target_size, dtype):
        add_time_ids = list(original_size + crops_coords_top_left + target_size)

        # passed_add_embed_dim = (
        #     self.unet.config.addition_time_embed_dim * len(add_time_ids) + self.text_encoder_2.config.projection_dim
        # )
        # expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features
        #
        # if expected_add_embed_dim != passed_add_embed_dim:
        #     raise ValueError(
        #         f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."
        #     )

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


class CustomStableDiffusionXLPipeline(StableDiffusionXLPipeline):
    # def __init__(self, *args, **kwargs):
    #     super().__init__(*args, **kwargs)

    def predict_noise(
            self,
            prompt: Union[str, List[str]] = None,
            prompt_2: Optional[Union[str, List[str]]] = None,
            num_inference_steps: int = 50,
            guidance_scale: float = 5.0,
            negative_prompt: Optional[Union[str, List[str]]] = None,
            negative_prompt_2: Optional[Union[str, List[str]]] = None,
            num_images_per_prompt: Optional[int] = 1,
            eta: float = 0.0,
            generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
            latents: Optional[torch.FloatTensor] = None,
            prompt_embeds: Optional[torch.FloatTensor] = None,
            negative_prompt_embeds: Optional[torch.FloatTensor] = None,
            pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
            negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
            cross_attention_kwargs: Optional[Dict[str, Any]] = None,
            guidance_rescale: float = 0.0,
            crops_coords_top_left: Tuple[int, int] = (0, 0),
            timestep: Optional[int] = None,
    ):
        r"""
        Function invoked when calling the pipeline for generation.

        Args:
            prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
                instead.
            prompt_2 (`str` or `List[str]`, *optional*):
                The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
                used in both text-encoders
            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                The height in pixels of the generated image.
            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                The width in pixels of the generated image.
            num_inference_steps (`int`, *optional*, defaults to 50):
                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
                expense of slower inference.
            denoising_end (`float`, *optional*):
                When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be
                completed before it is intentionally prematurely terminated. As a result, the returned sample will
                still retain a substantial amount of noise as determined by the discrete timesteps selected by the
                scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a
                "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image
                Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output)
            guidance_scale (`float`, *optional*, defaults to 7.5):
                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
                `guidance_scale` is defined as `w` of equation 2. of [Imagen
                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
                usually at the expense of lower image quality.
            negative_prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation. If not defined, one has to pass
                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
                less than `1`).
            negative_prompt_2 (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
                `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
            num_images_per_prompt (`int`, *optional*, defaults to 1):
                The number of images to generate per prompt.
            eta (`float`, *optional*, defaults to 0.0):
                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
                [`schedulers.DDIMScheduler`], will be ignored for others.
            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
                to make generation deterministic.
            latents (`torch.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.
            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
                If not provided, pooled text embeddings will be generated from `prompt` input argument.
            negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
                input argument.
            output_type (`str`, *optional*, defaults to `"pil"`):
                The output format of the generate image. Choose between
                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead
                of a plain tuple.
            callback (`Callable`, *optional*):
                A function that will be called every `callback_steps` steps during inference. The function will be
                called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
            callback_steps (`int`, *optional*, defaults to 1):
                The frequency at which the `callback` function will be called. If not specified, the callback will be
                called at every step.
            cross_attention_kwargs (`dict`, *optional*):
                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
                `self.processor` in
                [diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py).
            guidance_rescale (`float`, *optional*, defaults to 0.7):
                Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are
                Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of
                [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf).
                Guidance rescale factor should fix overexposure when using zero terminal SNR.
            original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
                If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled.
                `original_size` defaults to `(width, height)` if not specified. Part of SDXL's micro-conditioning as
                explained in section 2.2 of
                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
            crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
                `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
                `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
                `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
            target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
                For most cases, `target_size` should be set to the desired height and width of the generated image. If
                not specified it will default to `(width, height)`. Part of SDXL's micro-conditioning as explained in
                section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).

        Examples:

        Returns:
            [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] or `tuple`:
            [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a
            `tuple`. When returning a tuple, the first element is a list with the generated images.
        """
        # if not predict_noise:
        #     # call parent
        #     return super().__call__(
        #         prompt=prompt,
        #         prompt_2=prompt_2,
        #         height=height,
        #         width=width,
        #         num_inference_steps=num_inference_steps,
        #         denoising_end=denoising_end,
        #         guidance_scale=guidance_scale,
        #         negative_prompt=negative_prompt,
        #         negative_prompt_2=negative_prompt_2,
        #         num_images_per_prompt=num_images_per_prompt,
        #         eta=eta,
        #         generator=generator,
        #         latents=latents,
        #         prompt_embeds=prompt_embeds,
        #         negative_prompt_embeds=negative_prompt_embeds,
        #         pooled_prompt_embeds=pooled_prompt_embeds,
        #         negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
        #         output_type=output_type,
        #         return_dict=return_dict,
        #         callback=callback,
        #         callback_steps=callback_steps,
        #         cross_attention_kwargs=cross_attention_kwargs,
        #         guidance_rescale=guidance_rescale,
        #         original_size=original_size,
        #         crops_coords_top_left=crops_coords_top_left,
        #         target_size=target_size,
        #     )

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

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

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

        device = self._execution_device

        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
        # corresponds to doing no classifier free guidance.
        do_classifier_free_guidance = guidance_scale > 1.0

        # 3. Encode input prompt
        text_encoder_lora_scale = (
            cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
        )
        (
            prompt_embeds,
            negative_prompt_embeds,
            pooled_prompt_embeds,
            negative_pooled_prompt_embeds,
        ) = self.encode_prompt(
            prompt=prompt,
            prompt_2=prompt_2,
            device=device,
            num_images_per_prompt=num_images_per_prompt,
            do_classifier_free_guidance=do_classifier_free_guidance,
            negative_prompt=negative_prompt,
            negative_prompt_2=negative_prompt_2,
            prompt_embeds=prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds,
            pooled_prompt_embeds=pooled_prompt_embeds,
            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
            lora_scale=text_encoder_lora_scale,
        )

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

        # 5. Prepare latent variables
        num_channels_latents = self.unet.config.in_channels
        latents = self.prepare_latents(
            batch_size * num_images_per_prompt,
            num_channels_latents,
            height,
            width,
            prompt_embeds.dtype,
            device,
            generator,
            latents,
        )

        # 7. Prepare added time ids & embeddings
        add_text_embeds = pooled_prompt_embeds
        add_time_ids = self._get_add_time_ids(
            original_size, crops_coords_top_left, target_size, dtype=prompt_embeds.dtype
        ).to(device)  # TODO DOES NOT CAST ORIGINALLY

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

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

        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, timestep)

        # predict the noise residual
        added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
        noise_pred = self.unet(
            latent_model_input,
            timestep,
            encoder_hidden_states=prompt_embeds,
            cross_attention_kwargs=cross_attention_kwargs,
            added_cond_kwargs=added_cond_kwargs,
            return_dict=False,
        )[0]

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

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

        return noise_pred

    def enable_model_cpu_offload(self, gpu_id=0):
        print('Called cpu offload', gpu_id)
        # fuck off
        pass


class CustomStableDiffusionPipeline(StableDiffusionPipeline):

    # replace the call so it matches SDXL call so we can use the same code and also stop early
    def __call__(
            self,
            prompt: Union[str, List[str]] = None,
            prompt_2: Optional[Union[str, List[str]]] = None,
            height: Optional[int] = None,
            width: Optional[int] = None,
            num_inference_steps: int = 50,
            denoising_end: Optional[float] = None,
            guidance_scale: float = 5.0,
            negative_prompt: Optional[Union[str, List[str]]] = None,
            negative_prompt_2: Optional[Union[str, List[str]]] = None,
            num_images_per_prompt: Optional[int] = 1,
            eta: float = 0.0,
            generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
            latents: Optional[torch.FloatTensor] = None,
            prompt_embeds: Optional[torch.FloatTensor] = None,
            negative_prompt_embeds: Optional[torch.FloatTensor] = None,
            pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
            negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
            output_type: Optional[str] = "pil",
            return_dict: bool = True,
            callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
            callback_steps: int = 1,
            cross_attention_kwargs: Optional[Dict[str, Any]] = None,
            guidance_rescale: float = 0.0,
            original_size: Optional[Tuple[int, int]] = None,
            crops_coords_top_left: Tuple[int, int] = (0, 0),
            target_size: Optional[Tuple[int, int]] = None,
    ):
        # 0. Default height and width to unet
        height = height or self.unet.config.sample_size * self.vae_scale_factor
        width = width or self.unet.config.sample_size * self.vae_scale_factor

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

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

        device = self._execution_device
        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
        # corresponds to doing no classifier free guidance.
        do_classifier_free_guidance = guidance_scale > 1.0

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

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

        # 5. Prepare latent variables
        num_channels_latents = self.unet.config.in_channels
        latents = self.prepare_latents(
            batch_size * num_images_per_prompt,
            num_channels_latents,
            height,
            width,
            prompt_embeds.dtype,
            device,
            generator,
            latents,
        )

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

        # 7. Denoising loop
        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order

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

        with self.progress_bar(total=num_inference_steps) as progress_bar:
            for i, t in enumerate(timesteps):
                # expand the latents if we are doing classifier free guidance
                latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)

                # predict the noise residual
                noise_pred = self.unet(
                    latent_model_input,
                    t,
                    encoder_hidden_states=prompt_embeds,
                    cross_attention_kwargs=cross_attention_kwargs,
                    return_dict=False,
                )[0]

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

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

                # compute the previous noisy sample x_t -> x_t-1
                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]

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

        if not output_type == "latent":
            image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
            image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
        else:
            image = latents
            has_nsfw_concept = None

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

        image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)

        # Offload last model to CPU
        if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
            self.final_offload_hook.offload()

        if not return_dict:
            return (image, has_nsfw_concept)

        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)

    # some of the inputs are to keep it compatible with sdx
    def predict_noise(
            self,
            prompt: Union[str, List[str]] = None,
            prompt_2: Optional[Union[str, List[str]]] = None,
            num_inference_steps: int = 50,
            guidance_scale: float = 5.0,
            negative_prompt: Optional[Union[str, List[str]]] = None,
            negative_prompt_2: Optional[Union[str, List[str]]] = None,
            num_images_per_prompt: Optional[int] = 1,
            eta: float = 0.0,
            generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
            latents: Optional[torch.FloatTensor] = None,
            prompt_embeds: Optional[torch.FloatTensor] = None,
            negative_prompt_embeds: Optional[torch.FloatTensor] = None,
            pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
            negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
            cross_attention_kwargs: Optional[Dict[str, Any]] = None,
            guidance_rescale: float = 0.0,
            crops_coords_top_left: Tuple[int, int] = (0, 0),
            timestep: Optional[int] = None,
    ):

        # 0. Default height and width to unet
        height = self.unet.config.sample_size * self.vae_scale_factor
        width = self.unet.config.sample_size * self.vae_scale_factor

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

        device = self._execution_device
        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
        # corresponds to doing no classifier free guidance.
        do_classifier_free_guidance = guidance_scale > 1.0

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

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

        # 5. Prepare latent variables
        num_channels_latents = self.unet.config.in_channels
        latents = self.prepare_latents(
            batch_size * num_images_per_prompt,
            num_channels_latents,
            height,
            width,
            prompt_embeds.dtype,
            device,
            generator,
            latents,
        )

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

        # predict the noise residual
        noise_pred = self.unet(
            latent_model_input,
            timestep,
            encoder_hidden_states=prompt_embeds,
            cross_attention_kwargs=cross_attention_kwargs,
            return_dict=False,
        )[0]

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

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

        return noise_pred