[Community Pipeline] Add multilingual stable diffusion to community pipelines (#1142)

* Add multilingual_stable_diffusion.py file

* Add multilingual stable diffusion to examples README file

* Update examples/community/README.md
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

examples/community/README.md

@@ -18,9 +18,11 @@ If a community doesn't work as expected, please open an issue and ping the autho
 | Composable Stable Diffusion| Stable Diffusion Pipeline that supports prompts that contain "|" in prompts (as an AND condition) and weights (separated by "|" as well) to positively / negatively weight prompts.                                                                                                                                                                                                                                                                                                     | [Composable Stable Diffusion](#composable-stable-diffusion)                                                                 | -                                                                                                                                                                                                                  |                        [Mark Rich](https://github.com/MarkRich) |
 | Seed Resizing Stable Diffusion| Stable Diffusion Pipeline that supports resizing an image and retaining the concepts of the 512 by 512 generation.                                                                                                                                                                                                                                                                                                     | [Seed Resizing](#seed-resizing)                                                                 | -                                                                                                                                                                                                                  |                        [Mark Rich](https://github.com/MarkRich) |
 | Imagic Stable Diffusion | Stable Diffusion Pipeline that enables writing a text prompt to edit an existing image| [Imagic Stable Diffusion](#imagic-stable-diffusion)                                                                 | -                                                                                                                                                                                                                  |                        [Mark Rich](https://github.com/MarkRich) |
+| Multilingual Stable Diffusion| Stable Diffusion Pipeline that supports prompts in 50 different languages.                                                                                                                                                                                                                                                                                                     | [Multilingual Stable Diffusion](#multilingual-stable-diffusion-pipeline)                                                                 | -                                                                                                                                                                                                                  |                        [Juan Carlos Piñeros](https://github.com/juancopi81) |
 | Image to Image Inpainting Stable Diffusion | Stable Diffusion Pipeline that enables the overlaying of two images and subsequent inpainting| [Image to Image Inpainting Stable Diffusion](#image-to-image-inpainting-stable-diffusion)                                                                 | -                                                                                                                                                                                                                  |                        [Alex McKinney](https://github.com/vvvm23) |
 
 
+
 To load a custom pipeline you just need to pass the `custom_pipeline` argument to `DiffusionPipeline`, as one of the files in `diffusers/examples/community`. Feel free to send a PR with your own pipelines, we will merge them quickly.
 ```py
 pipe = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", custom_pipeline="filename_in_the_community_folder")
@@ -502,6 +504,74 @@ image = res.images[0]
 image.save('./seed_resize/seed_resize_{w}_{h}_image_compare.png'.format(w=width, h=height))
 ```
 
+### Multilingual Stable Diffusion Pipeline
+
+The following code can generate an images from texts in different languages using the pre-trained [mBART-50 many-to-one multilingual machine translation model](https://huggingface.co/facebook/mbart-large-50-many-to-one-mmt) and Stable Diffusion.
+
+```python
+from PIL import Image
+
+import torch
+
+from diffusers import DiffusionPipeline
+from transformers import (
+    pipeline,
+    MBart50TokenizerFast,
+    MBartForConditionalGeneration,
+)
+device = "cuda" if torch.cuda.is_available() else "cpu"
+device_dict = {"cuda": 0, "cpu": -1}
+
+# helper function taken from: https://huggingface.co/blog/stable_diffusion
+def image_grid(imgs, rows, cols):
+    assert len(imgs) == rows*cols
+
+    w, h = imgs[0].size
+    grid = Image.new('RGB', size=(cols*w, rows*h))
+    grid_w, grid_h = grid.size
+
+    for i, img in enumerate(imgs):
+        grid.paste(img, box=(i%cols*w, i//cols*h))
+    return grid
+
+# Add language detection pipeline
+language_detection_model_ckpt = "papluca/xlm-roberta-base-language-detection"
+language_detection_pipeline = pipeline("text-classification",
+                                       model=language_detection_model_ckpt,
+                                       device=device_dict[device])
+
+# Add model for language translation
+trans_tokenizer = MBart50TokenizerFast.from_pretrained("facebook/mbart-large-50-many-to-one-mmt")
+trans_model = MBartForConditionalGeneration.from_pretrained("facebook/mbart-large-50-many-to-one-mmt").to(device)
+
+diffuser_pipeline = DiffusionPipeline.from_pretrained(
+    "CompVis/stable-diffusion-v1-4",
+    custom_pipeline="multilingual_stable_diffusion",
+    detection_pipeline=language_detection_pipeline,
+    translation_model=trans_model,
+    translation_tokenizer=trans_tokenizer,
+    revision="fp16",
+    torch_dtype=torch.float16,
+)
+
+diffuser_pipeline.enable_attention_slicing()
+diffuser_pipeline = diffuser_pipeline.to(device)
+
+prompt = ["a photograph of an astronaut riding a horse", 
+          "Una casa en la playa",
+          "Ein Hund, der Orange isst",
+          "Un restaurant parisien"]
+
+output = diffuser_pipeline(prompt)
+
+images = output.images
+
+grid = image_grid(images, rows=2, cols=2)
+```
+
+This example produces the following images:
+![image](https://user-images.githubusercontent.com/4313860/198328706-295824a4-9856-4ce5-8e66-278ceb42fd29.png)
+
 ### Image to Image Inpainting Stable Diffusion
 
 Similar to the standard stable diffusion inpainting example, except with the addition of an `inner_image` argument.
@@ -534,4 +604,4 @@ pipe = pipe.to("cuda")
 
 prompt = "Your prompt here!"
 image = pipe(prompt=prompt, image=init_image, inner_image=inner_image, mask_image=mask_image).images[0]
-```
\ No newline at end of file
+```

examples/community/multilingual_stable_diffusion.py

+import inspect
+from typing import Callable, List, Optional, Union
+
+import torch
+
+from diffusers.configuration_utils import FrozenDict
+from diffusers.models import AutoencoderKL, UNet2DConditionModel
+from diffusers.pipeline_utils import DiffusionPipeline
+from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
+from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+from diffusers.schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
+from diffusers.utils import deprecate, logging
+from transformers import (
+    CLIPFeatureExtractor,
+    CLIPTextModel,
+    CLIPTokenizer,
+    MBart50TokenizerFast,
+    MBartForConditionalGeneration,
+    pipeline,
+)
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+def detect_language(pipe, prompt, batch_size):
+    """helper function to detect language(s) of prompt"""
+
+    if batch_size == 1:
+        preds = pipe(prompt, top_k=1, truncation=True, max_length=128)
+        return preds[0]["label"]
+    else:
+        detected_languages = []
+        for p in prompt:
+            preds = pipe(p, top_k=1, truncation=True, max_length=128)
+            detected_languages.append(preds[0]["label"])
+
+        return detected_languages
+
+
+def translate_prompt(prompt, translation_tokenizer, translation_model, device):
+    """helper function to translate prompt to English"""
+
+    encoded_prompt = translation_tokenizer(prompt, return_tensors="pt").to(device)
+    generated_tokens = translation_model.generate(**encoded_prompt, max_new_tokens=1000)
+    en_trans = translation_tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
+
+    return en_trans[0]
+
+
+class MultilingualStableDiffusion(DiffusionPipeline):
+    r"""
+    Pipeline for text-to-image generation using Stable Diffusion in different languages.
+
+    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:
+        detection_pipeline ([`pipeline`]):
+            Transformers pipeline to detect prompt's language.
+        translation_model ([`MBartForConditionalGeneration`]):
+            Model to translate prompt to English, if necessary. Please refer to the
+            [model card](https://huggingface.co/docs/transformers/model_doc/mbart) for details.
+        translation_tokenizer ([`MBart50TokenizerFast`]):
+            Tokenizer of the translation model.
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`CLIPTextModel`]):
+            Frozen text-encoder. Stable Diffusion uses the text portion of
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latens. Can be one of
+            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+        safety_checker ([`StableDiffusionSafetyChecker`]):
+            Classification module that estimates whether generated images could be considered offensive or harmful.
+            Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
+        feature_extractor ([`CLIPFeatureExtractor`]):
+            Model that extracts features from generated images to be used as inputs for the `safety_checker`.
+    """
+
+    def __init__(
+        self,
+        detection_pipeline: pipeline,
+        translation_model: MBartForConditionalGeneration,
+        translation_tokenizer: MBart50TokenizerFast,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModel,
+        tokenizer: CLIPTokenizer,
+        unet: UNet2DConditionModel,
+        scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
+        safety_checker: StableDiffusionSafetyChecker,
+        feature_extractor: CLIPFeatureExtractor,
+    ):
+        super().__init__()
+
+        if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
+            deprecation_message = (
+                f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
+                f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
+                "to update the config accordingly as leaving `steps_offset` might led to incorrect results"
+                " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
+                " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
+                " file"
+            )
+            deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
+            new_config = dict(scheduler.config)
+            new_config["steps_offset"] = 1
+            scheduler._internal_dict = FrozenDict(new_config)
+
+        if safety_checker is None:
+            logger.warn(
+                f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+                " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+                " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+                " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+                " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+                " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+            )
+
+        self.register_modules(
+            detection_pipeline=detection_pipeline,
+            translation_model=translation_model,
+            translation_tokenizer=translation_tokenizer,
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            unet=unet,
+            scheduler=scheduler,
+            safety_checker=safety_checker,
+            feature_extractor=feature_extractor,
+        )
+
+    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+
+        Args:
+            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
+                `attention_head_dim` must be a multiple of `slice_size`.
+        """
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = self.unet.config.attention_head_dim // 2
+        self.unet.set_attention_slice(slice_size)
+
+    def disable_attention_slicing(self):
+        r"""
+        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
+        back to computing attention in one step.
+        """
+        # set slice_size = `None` to disable `attention slicing`
+        self.enable_attention_slicing(None)
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        prompt: Union[str, List[str]],
+        height: int = 512,
+        width: int = 512,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 7.5,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[torch.Generator] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: Optional[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. Can be in different languages.
+            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.
+            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`).
+            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 (`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`.
+            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.
+
+        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`.
+        """
+        if isinstance(prompt, str):
+            batch_size = 1
+        elif isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        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)}."
+            )
+
+        # detect language and translate if necessary
+        prompt_language = detect_language(self.detection_pipeline, prompt, batch_size)
+        if batch_size == 1 and prompt_language != "en":
+            prompt = translate_prompt(prompt, self.translation_tokenizer, self.translation_model, self.device)
+
+        if isinstance(prompt, list):
+            for index in range(batch_size):
+                if prompt_language[index] != "en":
+                    p = translate_prompt(
+                        prompt[index], self.translation_tokenizer, self.translation_model, self.device
+                    )
+                    prompt[index] = p
+
+        # get prompt text embeddings
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=self.tokenizer.model_max_length,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+
+        if text_input_ids.shape[-1] > self.tokenizer.model_max_length:
+            removed_text = self.tokenizer.batch_decode(text_input_ids[:, self.tokenizer.model_max_length :])
+            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}"
+            )
+            text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length]
+        text_embeddings = self.text_encoder(text_input_ids.to(self.device))[0]
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        bs_embed, seq_len, _ = text_embeddings.shape
+        text_embeddings = text_embeddings.repeat(1, num_images_per_prompt, 1)
+        text_embeddings = text_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+        # 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
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance:
+            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):
+                # detect language and translate it if necessary
+                negative_prompt_language = detect_language(self.detection_pipeline, negative_prompt, batch_size)
+                if negative_prompt_language != "en":
+                    negative_prompt = translate_prompt(
+                        negative_prompt, self.translation_tokenizer, self.translation_model, self.device
+                    )
+                if 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:
+                # detect language and translate it if necessary
+                if isinstance(negative_prompt, list):
+                    negative_prompt_languages = detect_language(self.detection_pipeline, negative_prompt, batch_size)
+                    for index in range(batch_size):
+                        if negative_prompt_languages[index] != "en":
+                            p = translate_prompt(
+                                negative_prompt[index], self.translation_tokenizer, self.translation_model, self.device
+                            )
+                            negative_prompt[index] = p
+                uncond_tokens = negative_prompt
+
+            max_length = text_input_ids.shape[-1]
+            uncond_input = self.tokenizer(
+                uncond_tokens,
+                padding="max_length",
+                max_length=max_length,
+                truncation=True,
+                return_tensors="pt",
+            )
+            uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))[0]
+
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = uncond_embeddings.shape[1]
+            uncond_embeddings = uncond_embeddings.repeat(1, num_images_per_prompt, 1)
+            uncond_embeddings = uncond_embeddings.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
+            text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
+
+        # get the initial random noise unless the user supplied it
+
+        # Unlike in other pipelines, latents need to be generated in the target device
+        # for 1-to-1 results reproducibility with the CompVis implementation.
+        # However this currently doesn't work in `mps`.
+        latents_shape = (batch_size * num_images_per_prompt, self.unet.in_channels, height // 8, width // 8)
+        latents_dtype = text_embeddings.dtype
+        if latents is None:
+            if self.device.type == "mps":
+                # randn does not work reproducibly on mps
+                latents = torch.randn(latents_shape, generator=generator, device="cpu", dtype=latents_dtype).to(
+                    self.device
+                )
+            else:
+                latents = torch.randn(latents_shape, generator=generator, device=self.device, dtype=latents_dtype)
+        else:
+            if latents.shape != latents_shape:
+                raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}")
+            latents = latents.to(self.device)
+
+        # set timesteps
+        self.scheduler.set_timesteps(num_inference_steps)
+
+        # Some schedulers like PNDM have timesteps as arrays
+        # It's more optimized to move all timesteps to correct device beforehand
+        timesteps_tensor = self.scheduler.timesteps.to(self.device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+
+        # 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
+
+        for i, t in enumerate(self.progress_bar(timesteps_tensor)):
+            # 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=text_embeddings).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
+
+            # call the callback, if provided
+            if callback is not None and i % callback_steps == 0:
+                callback(i, t, latents)
+
+        latents = 1 / 0.18215 * 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 bfloa16
+        image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+
+        if self.safety_checker is not None:
+            safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(
+                self.device
+            )
+            image, has_nsfw_concept = self.safety_checker(
+                images=image, clip_input=safety_checker_input.pixel_values.to(text_embeddings.dtype)
+            )
+        else:
+            has_nsfw_concept = None
+
+        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)