ashawkey/zero123-xl-diffusers

1# Copyright 2023 The HuggingFace Team. All rights reserved.2#3# Licensed under the Apache License, Version 2.0 (the "License");4# you may not use this file except in compliance with the License.5# You may obtain a copy of the License at6#7#     http://www.apache.org/licenses/LICENSE-2.08#9# Unless required by applicable law or agreed to in writing, software10# distributed under the License is distributed on an "AS IS" BASIS,11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.12# See the License for the specific language governing permissions and13# limitations under the License.14 15import inspect16import math17import warnings18from typing import Any, Callable, Dict, List, Optional, Union19 20import PIL21import torch22import torchvision.transforms.functional as TF23from diffusers.configuration_utils import ConfigMixin, FrozenDict, register_to_config24from diffusers.image_processor import VaeImageProcessor25from diffusers.models import AutoencoderKL, UNet2DConditionModel26from diffusers.models.modeling_utils import ModelMixin27from diffusers.pipelines.pipeline_utils import DiffusionPipeline28from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput29from diffusers.pipelines.stable_diffusion.safety_checker import (30    StableDiffusionSafetyChecker,31)32from diffusers.schedulers import KarrasDiffusionSchedulers33from diffusers.utils import deprecate, is_accelerate_available, logging34from diffusers.utils.torch_utils import randn_tensor35from packaging import version36from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection37 38logger = logging.get_logger(__name__)  # pylint: disable=invalid-name39 40 41class CLIPCameraProjection(ModelMixin, ConfigMixin):42    """43    A Projection layer for CLIP embedding and camera embedding.44 45    Parameters:46        embedding_dim (`int`, *optional*, defaults to 768): The dimension of the model input `clip_embed`47        additional_embeddings (`int`, *optional*, defaults to 4): The number of additional tokens appended to the48            projected `hidden_states`. The actual length of the used `hidden_states` is `num_embeddings +49            additional_embeddings`.50    """51 52    @register_to_config53    def __init__(self, embedding_dim: int = 768, additional_embeddings: int = 4):54        super().__init__()55        self.embedding_dim = embedding_dim56        self.additional_embeddings = additional_embeddings57 58        self.input_dim = self.embedding_dim + self.additional_embeddings59        self.output_dim = self.embedding_dim60 61        self.proj = torch.nn.Linear(self.input_dim, self.output_dim)62 63    def forward(64        self,65        embedding: torch.FloatTensor,66    ):67        """68        The [`PriorTransformer`] forward method.69 70        Args:71            hidden_states (`torch.FloatTensor` of shape `(batch_size, input_dim)`):72                The currently input embeddings.73 74        Returns:75            The output embedding projection (`torch.FloatTensor` of shape `(batch_size, output_dim)`).76        """77        proj_embedding = self.proj(embedding)78        return proj_embedding79 80 81class Zero123Pipeline(DiffusionPipeline):82    r"""83    Pipeline to generate variations from an input image using Stable Diffusion.84 85    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the86    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)87 88    Args:89        vae ([`AutoencoderKL`]):90            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.91        image_encoder ([`CLIPVisionModelWithProjection`]):92            Frozen CLIP image-encoder. Stable Diffusion Image Variation uses the vision portion of93            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPVisionModelWithProjection),94            specifically the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.95        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.96        scheduler ([`SchedulerMixin`]):97            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of98            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].99        safety_checker ([`StableDiffusionSafetyChecker`]):100            Classification module that estimates whether generated images could be considered offensive or harmful.101            Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.102        feature_extractor ([`CLIPImageProcessor`]):103            Model that extracts features from generated images to be used as inputs for the `safety_checker`.104    """105    # TODO: feature_extractor is required to encode images (if they are in PIL format),106    # we should give a descriptive message if the pipeline doesn't have one.107    _optional_components = ["safety_checker"]108 109    def __init__(110        self,111        vae: AutoencoderKL,112        image_encoder: CLIPVisionModelWithProjection,113        unet: UNet2DConditionModel,114        scheduler: KarrasDiffusionSchedulers,115        safety_checker: StableDiffusionSafetyChecker,116        feature_extractor: CLIPImageProcessor,117        clip_camera_projection: CLIPCameraProjection,118        requires_safety_checker: bool = True,119    ):120        super().__init__()121 122        if safety_checker is None and requires_safety_checker:123            logger.warn(124                f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"125                " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"126                " results in services or applications open to the public. Both the diffusers team and Hugging Face"127                " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"128                " it only for use-cases that involve analyzing network behavior or auditing its results. For more"129                " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."130            )131 132        if safety_checker is not None and feature_extractor is None:133            raise ValueError(134                "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"135                " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."136            )137 138        is_unet_version_less_0_9_0 = hasattr(139            unet.config, "_diffusers_version"140        ) and version.parse(141            version.parse(unet.config._diffusers_version).base_version142        ) < version.parse(143            "0.9.0.dev0"144        )145        is_unet_sample_size_less_64 = (146            hasattr(unet.config, "sample_size") and unet.config.sample_size < 64147        )148        if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64:149            deprecation_message = (150                "The configuration file of the unet has set the default `sample_size` to smaller than"151                " 64 which seems highly unlikely .If you're checkpoint is a fine-tuned version of any of the"152                " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-"153                " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5"154                " \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the"155                " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`"156                " in the config might lead to incorrect results in future versions. If you have downloaded this"157                " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for"158                " the `unet/config.json` file"159            )160            deprecate(161                "sample_size<64", "1.0.0", deprecation_message, standard_warn=False162            )163            new_config = dict(unet.config)164            new_config["sample_size"] = 64165            unet._internal_dict = FrozenDict(new_config)166 167        self.register_modules(168            vae=vae,169            image_encoder=image_encoder,170            unet=unet,171            scheduler=scheduler,172            safety_checker=safety_checker,173            feature_extractor=feature_extractor,174            clip_camera_projection=clip_camera_projection,175        )176        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)177        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)178        self.register_to_config(requires_safety_checker=requires_safety_checker)179 180    def enable_sequential_cpu_offload(self, gpu_id=0):181        r"""182        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,183        text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a184        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.185        """186        if is_accelerate_available():187            from accelerate import cpu_offload188        else:189            raise ImportError("Please install accelerate via `pip install accelerate`")190 191        device = torch.device(f"cuda:{gpu_id}")192 193        for cpu_offloaded_model in [194            self.unet,195            self.image_encoder,196            self.vae,197            self.safety_checker,198        ]:199            if cpu_offloaded_model is not None:200                cpu_offload(cpu_offloaded_model, device)201 202    @property203    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device204    def _execution_device(self):205        r"""206        Returns the device on which the pipeline's models will be executed. After calling207        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module208        hooks.209        """210        if not hasattr(self.unet, "_hf_hook"):211            return self.device212        for module in self.unet.modules():213            if (214                hasattr(module, "_hf_hook")215                and hasattr(module._hf_hook, "execution_device")216                and module._hf_hook.execution_device is not None217            ):218                return torch.device(module._hf_hook.execution_device)219        return self.device220 221    def _encode_image(222        self,223        image,224        elevation,225        azimuth,226        distance,227        device,228        num_images_per_prompt,229        do_classifier_free_guidance,230        clip_image_embeddings=None,231        image_camera_embeddings=None,232    ):233        dtype = next(self.image_encoder.parameters()).dtype234 235        if image_camera_embeddings is None:236            if image is None:237                assert clip_image_embeddings is not None238                image_embeddings = clip_image_embeddings.to(device=device, dtype=dtype)239            else:240                if not isinstance(image, torch.Tensor):241                    image = self.feature_extractor(242                        images=image, return_tensors="pt"243                    ).pixel_values244 245                image = image.to(device=device, dtype=dtype)246                image_embeddings = self.image_encoder(image).image_embeds247                image_embeddings = image_embeddings.unsqueeze(1)248 249            bs_embed, seq_len, _ = image_embeddings.shape250 251            if isinstance(elevation, float):252                elevation = torch.as_tensor(253                    [elevation] * bs_embed, dtype=dtype, device=device254                )255            if isinstance(azimuth, float):256                azimuth = torch.as_tensor(257                    [azimuth] * bs_embed, dtype=dtype, device=device258                )259            if isinstance(distance, float):260                distance = torch.as_tensor(261                    [distance] * bs_embed, dtype=dtype, device=device262                )263 264            camera_embeddings = torch.stack(265                [266                    torch.deg2rad(elevation),267                    torch.sin(torch.deg2rad(azimuth)),268                    torch.cos(torch.deg2rad(azimuth)),269                    distance,270                ],271                dim=-1,272            )[:, None, :]273 274            image_embeddings = torch.cat([image_embeddings, camera_embeddings], dim=-1)275 276            # project (image, camera) embeddings to the same dimension as clip embeddings277            image_embeddings = self.clip_camera_projection(image_embeddings)278        else:279            image_embeddings = image_camera_embeddings.to(device=device, dtype=dtype)280            bs_embed, seq_len, _ = image_embeddings.shape281 282        # duplicate image embeddings for each generation per prompt, using mps friendly method283        image_embeddings = image_embeddings.repeat(1, num_images_per_prompt, 1)284        image_embeddings = image_embeddings.view(285            bs_embed * num_images_per_prompt, seq_len, -1286        )287 288        if do_classifier_free_guidance:289            negative_prompt_embeds = torch.zeros_like(image_embeddings)290 291            # For classifier free guidance, we need to do two forward passes.292            # Here we concatenate the unconditional and text embeddings into a single batch293            # to avoid doing two forward passes294            image_embeddings = torch.cat([negative_prompt_embeds, image_embeddings])295 296        return image_embeddings297 298    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker299    def run_safety_checker(self, image, device, dtype):300        if self.safety_checker is None:301            has_nsfw_concept = None302        else:303            if torch.is_tensor(image):304                feature_extractor_input = self.image_processor.postprocess(305                    image, output_type="pil"306                )307            else:308                feature_extractor_input = self.image_processor.numpy_to_pil(image)309            safety_checker_input = self.feature_extractor(310                feature_extractor_input, return_tensors="pt"311            ).to(device)312            image, has_nsfw_concept = self.safety_checker(313                images=image, clip_input=safety_checker_input.pixel_values.to(dtype)314            )315        return image, has_nsfw_concept316 317    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents318    def decode_latents(self, latents):319        warnings.warn(320            "The decode_latents method is deprecated and will be removed in a future version. Please"321            " use VaeImageProcessor instead",322            FutureWarning,323        )324        latents = 1 / self.vae.config.scaling_factor * latents325        image = self.vae.decode(latents, return_dict=False)[0]326        image = (image / 2 + 0.5).clamp(0, 1)327        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16328        image = image.cpu().permute(0, 2, 3, 1).float().numpy()329        return image330 331    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs332    def prepare_extra_step_kwargs(self, generator, eta):333        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature334        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.335        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502336        # and should be between [0, 1]337 338        accepts_eta = "eta" in set(339            inspect.signature(self.scheduler.step).parameters.keys()340        )341        extra_step_kwargs = {}342        if accepts_eta:343            extra_step_kwargs["eta"] = eta344 345        # check if the scheduler accepts generator346        accepts_generator = "generator" in set(347            inspect.signature(self.scheduler.step).parameters.keys()348        )349        if accepts_generator:350            extra_step_kwargs["generator"] = generator351        return extra_step_kwargs352 353    def check_inputs(self, image, height, width, callback_steps):354        # TODO: check image size or adjust image size to (height, width)355 356        if height % 8 != 0 or width % 8 != 0:357            raise ValueError(358                f"`height` and `width` have to be divisible by 8 but are {height} and {width}."359            )360 361        if (callback_steps is None) or (362            callback_steps is not None363            and (not isinstance(callback_steps, int) or callback_steps <= 0)364        ):365            raise ValueError(366                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"367                f" {type(callback_steps)}."368            )369 370    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents371    def prepare_latents(372        self,373        batch_size,374        num_channels_latents,375        height,376        width,377        dtype,378        device,379        generator,380        latents=None,381    ):382        shape = (383            batch_size,384            num_channels_latents,385            height // self.vae_scale_factor,386            width // self.vae_scale_factor,387        )388        if isinstance(generator, list) and len(generator) != batch_size:389            raise ValueError(390                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"391                f" size of {batch_size}. Make sure the batch size matches the length of the generators."392            )393 394        if latents is None:395            latents = randn_tensor(396                shape, generator=generator, device=device, dtype=dtype397            )398        else:399            latents = latents.to(device)400 401        # scale the initial noise by the standard deviation required by the scheduler402        latents = latents * self.scheduler.init_noise_sigma403        return latents404 405    def _get_latent_model_input(406        self,407        latents: torch.FloatTensor,408        image: Optional[409            Union[PIL.Image.Image, List[PIL.Image.Image], torch.FloatTensor]410        ],411        num_images_per_prompt: int,412        do_classifier_free_guidance: bool,413        image_latents: Optional[torch.FloatTensor] = None,414    ):415        if isinstance(image, PIL.Image.Image):416            image_pt = TF.to_tensor(image).unsqueeze(0).to(latents)417        elif isinstance(image, list):418            image_pt = torch.stack([TF.to_tensor(img) for img in image], dim=0).to(419                latents420            )421        elif isinstance(image, torch.Tensor):422            image_pt = image423        else:424            image_pt = None425 426        if image_pt is None:427            assert image_latents is not None428            image_pt = image_latents.repeat_interleave(num_images_per_prompt, dim=0)429        else:430            image_pt = image_pt * 2.0 - 1.0  # scale to [-1, 1]431            # FIXME: encoded latents should be multiplied with self.vae.config.scaling_factor432            # but zero123 was not trained this way433            image_pt = self.vae.encode(image_pt).latent_dist.mode()434            image_pt = image_pt.repeat_interleave(num_images_per_prompt, dim=0)435        if do_classifier_free_guidance:436            latent_model_input = torch.cat(437                [438                    torch.cat([latents, latents], dim=0),439                    torch.cat([torch.zeros_like(image_pt), image_pt], dim=0),440                ],441                dim=1,442            )443        else:444            latent_model_input = torch.cat([latents, image_pt], dim=1)445 446        return latent_model_input447 448    @torch.no_grad()449    def __call__(450        self,451        image: Optional[452            Union[PIL.Image.Image, List[PIL.Image.Image], torch.FloatTensor]453        ] = None,454        elevation: Optional[Union[float, torch.FloatTensor]] = None,455        azimuth: Optional[Union[float, torch.FloatTensor]] = None,456        distance: Optional[Union[float, torch.FloatTensor]] = None,457        height: Optional[int] = None,458        width: Optional[int] = None,459        num_inference_steps: int = 50,460        guidance_scale: float = 3.0,461        num_images_per_prompt: int = 1,462        eta: float = 0.0,463        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,464        latents: Optional[torch.FloatTensor] = None,465        clip_image_embeddings: Optional[torch.FloatTensor] = None,466        image_camera_embeddings: Optional[torch.FloatTensor] = None,467        image_latents: Optional[torch.FloatTensor] = None,468        output_type: Optional[str] = "pil",469        return_dict: bool = True,470        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,471        callback_steps: int = 1,472        cross_attention_kwargs: Optional[Dict[str, Any]] = None,473    ):474        r"""475        Function invoked when calling the pipeline for generation.476 477        Args:478            image (`PIL.Image.Image` or `List[PIL.Image.Image]` or `torch.FloatTensor`):479                The image or images to guide the image generation. If you provide a tensor, it needs to comply with the480                configuration of481                [this](https://huggingface.co/lambdalabs/sd-image-variations-diffusers/blob/main/feature_extractor/preprocessor_config.json)482                `CLIPImageProcessor`483            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):484                The height in pixels of the generated image.485            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):486                The width in pixels of the generated image.487            num_inference_steps (`int`, *optional*, defaults to 50):488                The number of denoising steps. More denoising steps usually lead to a higher quality image at the489                expense of slower inference.490            guidance_scale (`float`, *optional*, defaults to 7.5):491                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).492                `guidance_scale` is defined as `w` of equation 2. of [Imagen493                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >494                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,495                usually at the expense of lower image quality.496            num_images_per_prompt (`int`, *optional*, defaults to 1):497                The number of images to generate per prompt.498            eta (`float`, *optional*, defaults to 0.0):499                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to500                [`schedulers.DDIMScheduler`], will be ignored for others.501            generator (`torch.Generator`, *optional*):502                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)503                to make generation deterministic.504            latents (`torch.FloatTensor`, *optional*):505                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image506                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents507                tensor will ge generated by sampling using the supplied random `generator`.508            output_type (`str`, *optional*, defaults to `"pil"`):509                The output format of the generate image. Choose between510                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.511            return_dict (`bool`, *optional*, defaults to `True`):512                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a513                plain tuple.514            callback (`Callable`, *optional*):515                A function that will be called every `callback_steps` steps during inference. The function will be516                called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.517            callback_steps (`int`, *optional*, defaults to 1):518                The frequency at which the `callback` function will be called. If not specified, the callback will be519                called at every step.520 521        Returns:522            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:523            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.524            When returning a tuple, the first element is a list with the generated images, and the second element is a525            list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"526            (nsfw) content, according to the `safety_checker`.527        """528        # 0. Default height and width to unet529        height = height or self.unet.config.sample_size * self.vae_scale_factor530        width = width or self.unet.config.sample_size * self.vae_scale_factor531 532        # 1. Check inputs. Raise error if not correct533        # TODO: check input elevation, azimuth, and distance534        # TODO: check image, clip_image_embeddings, image_latents535        self.check_inputs(image, height, width, callback_steps)536 537        # 2. Define call parameters538        if isinstance(image, PIL.Image.Image):539            batch_size = 1540        elif isinstance(image, list):541            batch_size = len(image)542        elif isinstance(image, torch.Tensor):543            batch_size = image.shape[0]544        else:545            assert image_latents is not None546            assert (547                clip_image_embeddings is not None or image_camera_embeddings is not None548            )549            batch_size = image_latents.shape[0]550 551        device = self._execution_device552        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)553        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`554        # corresponds to doing no classifier free guidance.555        do_classifier_free_guidance = guidance_scale > 1.0556 557        # 3. Encode input image558        if isinstance(image, PIL.Image.Image) or isinstance(image, list):559            pil_image = image560        elif isinstance(image, torch.Tensor):561            pil_image = [TF.to_pil_image(image[i]) for i in range(image.shape[0])]562        else:563            pil_image = None564        image_embeddings = self._encode_image(565            pil_image,566            elevation,567            azimuth,568            distance,569            device,570            num_images_per_prompt,571            do_classifier_free_guidance,572            clip_image_embeddings,573            image_camera_embeddings,574        )575 576        # 4. Prepare timesteps577        self.scheduler.set_timesteps(num_inference_steps, device=device)578        timesteps = self.scheduler.timesteps579 580        # 5. Prepare latent variables581        # num_channels_latents = self.unet.config.in_channels582        num_channels_latents = 4  # FIXME: hard-coded583        latents = self.prepare_latents(584            batch_size * num_images_per_prompt,585            num_channels_latents,586            height,587            width,588            image_embeddings.dtype,589            device,590            generator,591            latents,592        )593 594        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline595        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)596 597        # 7. Denoising loop598        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order599        with self.progress_bar(total=num_inference_steps) as progress_bar:600            for i, t in enumerate(timesteps):601                # expand the latents if we are doing classifier free guidance602                latent_model_input = self._get_latent_model_input(603                    latents,604                    image,605                    num_images_per_prompt,606                    do_classifier_free_guidance,607                    image_latents,608                )609                latent_model_input = self.scheduler.scale_model_input(610                    latent_model_input, t611                )612 613                # predict the noise residual614                noise_pred = self.unet(615                    latent_model_input,616                    t,617                    encoder_hidden_states=image_embeddings,618                    cross_attention_kwargs=cross_attention_kwargs,619                ).sample620 621                # perform guidance622                if do_classifier_free_guidance:623                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)624                    noise_pred = noise_pred_uncond + guidance_scale * (625                        noise_pred_text - noise_pred_uncond626                    )627 628                # compute the previous noisy sample x_t -> x_t-1629                latents = self.scheduler.step(630                    noise_pred, t, latents, **extra_step_kwargs631                ).prev_sample632 633                # call the callback, if provided634                if i == len(timesteps) - 1 or (635                    (i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0636                ):637                    progress_bar.update()638                    if callback is not None and i % callback_steps == 0:639                        callback(i, t, latents)640 641        if not output_type == "latent":642            image = self.vae.decode(643                latents / self.vae.config.scaling_factor, return_dict=False644            )[0]645            image, has_nsfw_concept = self.run_safety_checker(646                image, device, image_embeddings.dtype647            )648        else:649            image = latents650            has_nsfw_concept = None651 652        if has_nsfw_concept is None:653            do_denormalize = [True] * image.shape[0]654        else:655            do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]656 657        image = self.image_processor.postprocess(658            image, output_type=output_type, do_denormalize=do_denormalize659        )660 661        if not return_dict:662            return (image, has_nsfw_concept)663 664        return StableDiffusionPipelineOutput(665            images=image, nsfw_content_detected=has_nsfw_concept666        )