Create zero123.py

clip_camera_projection/zero123.py

@@ -0,0 +1,666 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+import math
+import warnings
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import PIL
+import torch
+import torchvision.transforms.functional as TF
+from diffusers.configuration_utils import ConfigMixin, FrozenDict, register_to_config
+from diffusers.image_processor import VaeImageProcessor
+from diffusers.models import AutoencoderKL, UNet2DConditionModel
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
+from diffusers.pipelines.stable_diffusion.safety_checker import (
+    StableDiffusionSafetyChecker,
+)
+from diffusers.schedulers import KarrasDiffusionSchedulers
+from diffusers.utils import deprecate, is_accelerate_available, logging
+from diffusers.utils.torch_utils import randn_tensor
+from packaging import version
+from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class CLIPCameraProjection(ModelMixin, ConfigMixin):
+    """
+    A Projection layer for CLIP embedding and camera embedding.
+
+    Parameters:
+        embedding_dim (`int`, *optional*, defaults to 768): The dimension of the model input `clip_embed`
+        additional_embeddings (`int`, *optional*, defaults to 4): The number of additional tokens appended to the
+            projected `hidden_states`. The actual length of the used `hidden_states` is `num_embeddings +
+            additional_embeddings`.
+    """
+
+    @register_to_config
+    def __init__(self, embedding_dim: int = 768, additional_embeddings: int = 4):
+        super().__init__()
+        self.embedding_dim = embedding_dim
+        self.additional_embeddings = additional_embeddings
+
+        self.input_dim = self.embedding_dim + self.additional_embeddings
+        self.output_dim = self.embedding_dim
+
+        self.proj = torch.nn.Linear(self.input_dim, self.output_dim)
+
+    def forward(
+        self,
+        embedding: torch.FloatTensor,
+    ):
+        """
+        The [`PriorTransformer`] forward method.
+
+        Args:
+            hidden_states (`torch.FloatTensor` of shape `(batch_size, input_dim)`):
+                The currently input embeddings.
+
+        Returns:
+            The output embedding projection (`torch.FloatTensor` of shape `(batch_size, output_dim)`).
+        """
+        proj_embedding = self.proj(embedding)
+        return proj_embedding
+
+
+class Zero123Pipeline(DiffusionPipeline):
+    r"""
+    Pipeline to generate variations from an input image using Stable Diffusion.
+
+    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:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        image_encoder ([`CLIPVisionModelWithProjection`]):
+            Frozen CLIP image-encoder. Stable Diffusion Image Variation uses the vision portion of
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPVisionModelWithProjection),
+            specifically the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+        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 latents. 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 ([`CLIPImageProcessor`]):
+            Model that extracts features from generated images to be used as inputs for the `safety_checker`.
+    """
+    # TODO: feature_extractor is required to encode images (if they are in PIL format),
+    # we should give a descriptive message if the pipeline doesn't have one.
+    _optional_components = ["safety_checker"]
+
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        image_encoder: CLIPVisionModelWithProjection,
+        unet: UNet2DConditionModel,
+        scheduler: KarrasDiffusionSchedulers,
+        safety_checker: StableDiffusionSafetyChecker,
+        feature_extractor: CLIPImageProcessor,
+        clip_camera_projection: CLIPCameraProjection,
+        requires_safety_checker: bool = True,
+    ):
+        super().__init__()
+
+        if safety_checker is None and requires_safety_checker:
+            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 ."
+            )
+
+        if safety_checker is not None and feature_extractor is None:
+            raise ValueError(
+                "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+                " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+            )
+
+        is_unet_version_less_0_9_0 = hasattr(
+            unet.config, "_diffusers_version"
+        ) and version.parse(
+            version.parse(unet.config._diffusers_version).base_version
+        ) < version.parse(
+            "0.9.0.dev0"
+        )
+        is_unet_sample_size_less_64 = (
+            hasattr(unet.config, "sample_size") and unet.config.sample_size < 64
+        )
+        if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64:
+            deprecation_message = (
+                "The configuration file of the unet has set the default `sample_size` to smaller than"
+                " 64 which seems highly unlikely .If you're checkpoint is a fine-tuned version of any of the"
+                " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-"
+                " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5"
+                " \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the"
+                " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`"
+                " in the config might lead 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 `unet/config.json` file"
+            )
+            deprecate(
+                "sample_size<64", "1.0.0", deprecation_message, standard_warn=False
+            )
+            new_config = dict(unet.config)
+            new_config["sample_size"] = 64
+            unet._internal_dict = FrozenDict(new_config)
+
+        self.register_modules(
+            vae=vae,
+            image_encoder=image_encoder,
+            unet=unet,
+            scheduler=scheduler,
+            safety_checker=safety_checker,
+            feature_extractor=feature_extractor,
+            clip_camera_projection=clip_camera_projection,
+        )
+        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.register_to_config(requires_safety_checker=requires_safety_checker)
+
+    def enable_sequential_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
+        text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
+        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
+        """
+        if is_accelerate_available():
+            from accelerate import cpu_offload
+        else:
+            raise ImportError("Please install accelerate via `pip install accelerate`")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        for cpu_offloaded_model in [
+            self.unet,
+            self.image_encoder,
+            self.vae,
+            self.safety_checker,
+        ]:
+            if cpu_offloaded_model is not None:
+                cpu_offload(cpu_offloaded_model, device)
+
+    @property
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device
+    def _execution_device(self):
+        r"""
+        Returns the device on which the pipeline's models will be executed. After calling
+        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
+        hooks.
+        """
+        if not hasattr(self.unet, "_hf_hook"):
+            return self.device
+        for module in self.unet.modules():
+            if (
+                hasattr(module, "_hf_hook")
+                and hasattr(module._hf_hook, "execution_device")
+                and module._hf_hook.execution_device is not None
+            ):
+                return torch.device(module._hf_hook.execution_device)
+        return self.device
+
+    def _encode_image(
+        self,
+        image,
+        elevation,
+        azimuth,
+        distance,
+        device,
+        num_images_per_prompt,
+        do_classifier_free_guidance,
+        clip_image_embeddings=None,
+        image_camera_embeddings=None,
+    ):
+        dtype = next(self.image_encoder.parameters()).dtype
+
+        if image_camera_embeddings is None:
+            if image is None:
+                assert clip_image_embeddings is not None
+                image_embeddings = clip_image_embeddings.to(device=device, dtype=dtype)
+            else:
+                if not isinstance(image, torch.Tensor):
+                    image = self.feature_extractor(
+                        images=image, return_tensors="pt"
+                    ).pixel_values
+
+                image = image.to(device=device, dtype=dtype)
+                image_embeddings = self.image_encoder(image).image_embeds
+                image_embeddings = image_embeddings.unsqueeze(1)
+
+            bs_embed, seq_len, _ = image_embeddings.shape
+
+            if isinstance(elevation, float):
+                elevation = torch.as_tensor(
+                    [elevation] * bs_embed, dtype=dtype, device=device
+                )
+            if isinstance(azimuth, float):
+                azimuth = torch.as_tensor(
+                    [azimuth] * bs_embed, dtype=dtype, device=device
+                )
+            if isinstance(distance, float):
+                distance = torch.as_tensor(
+                    [distance] * bs_embed, dtype=dtype, device=device
+                )
+
+            camera_embeddings = torch.stack(
+                [
+                    torch.deg2rad(elevation),
+                    torch.sin(torch.deg2rad(azimuth)),
+                    torch.cos(torch.deg2rad(azimuth)),
+                    distance,
+                ],
+                dim=-1,
+            )[:, None, :]
+
+            image_embeddings = torch.cat([image_embeddings, camera_embeddings], dim=-1)
+
+            # project (image, camera) embeddings to the same dimension as clip embeddings
+            image_embeddings = self.clip_camera_projection(image_embeddings)
+        else:
+            image_embeddings = image_camera_embeddings.to(device=device, dtype=dtype)
+            bs_embed, seq_len, _ = image_embeddings.shape
+
+        # duplicate image embeddings for each generation per prompt, using mps friendly method
+        image_embeddings = image_embeddings.repeat(1, num_images_per_prompt, 1)
+        image_embeddings = image_embeddings.view(
+            bs_embed * num_images_per_prompt, seq_len, -1
+        )
+
+        if do_classifier_free_guidance:
+            negative_prompt_embeds = torch.zeros_like(image_embeddings)
+
+            # For classifier free guidance, we need to do two forward passes.
+            # Here we concatenate the unconditional and text embeddings into a single batch
+            # to avoid doing two forward passes
+            image_embeddings = torch.cat([negative_prompt_embeds, image_embeddings])
+
+        return image_embeddings
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+    def run_safety_checker(self, image, device, dtype):
+        if self.safety_checker is None:
+            has_nsfw_concept = None
+        else:
+            if torch.is_tensor(image):
+                feature_extractor_input = self.image_processor.postprocess(
+                    image, output_type="pil"
+                )
+            else:
+                feature_extractor_input = self.image_processor.numpy_to_pil(image)
+            safety_checker_input = self.feature_extractor(
+                feature_extractor_input, return_tensors="pt"
+            ).to(device)
+            image, has_nsfw_concept = self.safety_checker(
+                images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+            )
+        return image, has_nsfw_concept
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+    def decode_latents(self, latents):
+        warnings.warn(
+            "The decode_latents method is deprecated and will be removed in a future version. Please"
+            " use VaeImageProcessor instead",
+            FutureWarning,
+        )
+        latents = 1 / self.vae.config.scaling_factor * latents
+        image = self.vae.decode(latents, return_dict=False)[0]
+        image = (image / 2 + 0.5).clamp(0, 1)
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+        image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+        return image
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(
+            inspect.signature(self.scheduler.step).parameters.keys()
+        )
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(
+            inspect.signature(self.scheduler.step).parameters.keys()
+        )
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    def check_inputs(self, image, height, width, callback_steps):
+        # TODO: check image size or adjust image size to (height, width)
+
+        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)}."
+            )
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+    def prepare_latents(
+        self,
+        batch_size,
+        num_channels_latents,
+        height,
+        width,
+        dtype,
+        device,
+        generator,
+        latents=None,
+    ):
+        shape = (
+            batch_size,
+            num_channels_latents,
+            height // self.vae_scale_factor,
+            width // self.vae_scale_factor,
+        )
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        if latents is None:
+            latents = randn_tensor(
+                shape, generator=generator, device=device, dtype=dtype
+            )
+        else:
+            latents = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    def _get_latent_model_input(
+        self,
+        latents: torch.FloatTensor,
+        image: Optional[
+            Union[PIL.Image.Image, List[PIL.Image.Image], torch.FloatTensor]
+        ],
+        num_images_per_prompt: int,
+        do_classifier_free_guidance: bool,
+        image_latents: Optional[torch.FloatTensor] = None,
+    ):
+        if isinstance(image, PIL.Image.Image):
+            image_pt = TF.to_tensor(image).unsqueeze(0).to(latents)
+        elif isinstance(image, list):
+            image_pt = torch.stack([TF.to_tensor(img) for img in image], dim=0).to(
+                latents
+            )
+        elif isinstance(image, torch.Tensor):
+            image_pt = image
+        else:
+            image_pt = None
+
+        if image_pt is None:
+            assert image_latents is not None
+            image_pt = image_latents.repeat_interleave(num_images_per_prompt, dim=0)
+        else:
+            image_pt = image_pt * 2.0 - 1.0  # scale to [-1, 1]
+            # FIXME: encoded latents should be multiplied with self.vae.config.scaling_factor
+            # but zero123 was not trained this way
+            image_pt = self.vae.encode(image_pt).latent_dist.mode()
+            image_pt = image_pt.repeat_interleave(num_images_per_prompt, dim=0)
+        if do_classifier_free_guidance:
+            latent_model_input = torch.cat(
+                [
+                    torch.cat([latents, latents], dim=0),
+                    torch.cat([torch.zeros_like(image_pt), image_pt], dim=0),
+                ],
+                dim=1,
+            )
+        else:
+            latent_model_input = torch.cat([latents, image_pt], dim=1)
+
+        return latent_model_input
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        image: Optional[
+            Union[PIL.Image.Image, List[PIL.Image.Image], torch.FloatTensor]
+        ] = None,
+        elevation: Optional[Union[float, torch.FloatTensor]] = None,
+        azimuth: Optional[Union[float, torch.FloatTensor]] = None,
+        distance: Optional[Union[float, torch.FloatTensor]] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 3.0,
+        num_images_per_prompt: int = 1,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        clip_image_embeddings: Optional[torch.FloatTensor] = None,
+        image_camera_embeddings: Optional[torch.FloatTensor] = None,
+        image_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: int = 1,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            image (`PIL.Image.Image` or `List[PIL.Image.Image]` or `torch.FloatTensor`):
+                The image or images to guide the image generation. If you provide a tensor, it needs to comply with the
+                configuration of
+                [this](https://huggingface.co/lambdalabs/sd-image-variations-diffusers/blob/main/feature_extractor/preprocessor_config.json)
+                `CLIPImageProcessor`
+            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.
+            guidance_scale (`float`, *optional*, defaults to 7.5):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            generator (`torch.Generator`, *optional*):
+                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`.
+            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`.
+        """
+        # 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
+        # TODO: check input elevation, azimuth, and distance
+        # TODO: check image, clip_image_embeddings, image_latents
+        self.check_inputs(image, height, width, callback_steps)
+
+        # 2. Define call parameters
+        if isinstance(image, PIL.Image.Image):
+            batch_size = 1
+        elif isinstance(image, list):
+            batch_size = len(image)
+        elif isinstance(image, torch.Tensor):
+            batch_size = image.shape[0]
+        else:
+            assert image_latents is not None
+            assert (
+                clip_image_embeddings is not None or image_camera_embeddings is not None
+            )
+            batch_size = image_latents.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 image
+        if isinstance(image, PIL.Image.Image) or isinstance(image, list):
+            pil_image = image
+        elif isinstance(image, torch.Tensor):
+            pil_image = [TF.to_pil_image(image[i]) for i in range(image.shape[0])]
+        else:
+            pil_image = None
+        image_embeddings = self._encode_image(
+            pil_image,
+            elevation,
+            azimuth,
+            distance,
+            device,
+            num_images_per_prompt,
+            do_classifier_free_guidance,
+            clip_image_embeddings,
+            image_camera_embeddings,
+        )
+
+        # 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
+        num_channels_latents = 4  # FIXME: hard-coded
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            image_embeddings.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
+        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 = self._get_latent_model_input(
+                    latents,
+                    image,
+                    num_images_per_prompt,
+                    do_classifier_free_guidance,
+                    image_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=image_embeddings,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                ).sample
+
+                # perform guidance
+                if do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + guidance_scale * (
+                        noise_pred_text - noise_pred_uncond
+                    )
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(
+                    noise_pred, t, latents, **extra_step_kwargs
+                ).prev_sample
+
+                # 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, image_embeddings.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
+        )
+
+        if not return_dict:
+            return (image, has_nsfw_concept)
+
+        return StableDiffusionPipelineOutput(
+            images=image, nsfw_content_detected=has_nsfw_concept
+        )