Create pipeline.py

analogy_encoder/pipeline.py

@@ -0,0 +1,741 @@
+"""
+ADOBE CONFIDENTIAL
+Copyright 2024 Adobe
+All Rights Reserved.
+NOTICE: All information contained herein is, and remains
+the property of Adobe and its suppliers, if any. The intellectual
+and technical concepts contained herein are proprietary to Adobe
+and its suppliers and are protected by all applicable intellectual
+property laws, including trade secret and copyright laws.
+Dissemination of this information or reproduction of this material
+is strictly forbidden unless prior written permission is obtained
+from Adobe.
+"""
+
+from typing import Callable, List, Optional, Union
+import inspect
+import einops
+import PIL.Image
+import numpy as np
+import torch as th
+import torch.nn as nn
+from torchvision import transforms
+
+from diffusers import ModelMixin
+from transformers import AutoModel, AutoConfig, SiglipVisionConfig, Dinov2Config, Dinov2Model
+from transformers import SiglipVisionModel
+from diffusers import DiffusionPipeline
+from diffusers.image_processor import VaeImageProcessor
+from diffusers.models import AutoencoderKL, UNet2DConditionModel
+from diffusers.schedulers import KarrasDiffusionSchedulers
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+# REf: https://github.com/tatp22/multidim-positional-encoding/tree/master
+
+    
+OUT_SIZE = 768
+IN_SIZE = 2048
+
+DINO_SIZE = 224
+DINO_MEAN = [0.485, 0.456, 0.406]
+DINO_STD = [0.229, 0.224, 0.225]
+
+SIGLIP_SIZE = 256
+SIGLIP_MEAN = [0.5]
+SIGLIP_STD = [0.5]
+
+
+def get_emb(sin_inp):
+    """
+    Gets a base embedding for one dimension with sin and cos intertwined
+    """
+    emb = th.stack((sin_inp.sin(), sin_inp.cos()), dim=-1)
+    return th.flatten(emb, -2, -1)
+
+
+class PositionalEncoding1D(nn.Module):
+    def __init__(self, channels):
+        """
+        :param channels: The last dimension of the tensor you want to apply pos emb to.
+        """
+        super(PositionalEncoding1D, self).__init__()
+        self.org_channels = channels
+        channels = int(np.ceil(channels / 2) * 2)
+        self.channels = channels
+        inv_freq = 1.0 / (10000 ** (th.arange(0, channels, 2).float() / channels))
+        self.register_buffer("inv_freq", inv_freq)
+        self.register_buffer("cached_penc", None, persistent=False)
+
+    def forward(self, tensor):
+        """
+        :param tensor: A 3d tensor of size (batch_size, x, ch)
+        :return: Positional Encoding Matrix of size (batch_size, x, ch)
+        """
+        if len(tensor.shape) != 3:
+            raise RuntimeError("The input tensor has to be 3d!")
+
+        if self.cached_penc is not None and self.cached_penc.shape == tensor.shape:
+            return self.cached_penc
+
+        self.cached_penc = None
+        batch_size, x, orig_ch = tensor.shape
+        pos_x = th.arange(x, device=tensor.device, dtype=self.inv_freq.dtype)
+        sin_inp_x = th.einsum("i,j->ij", pos_x, self.inv_freq)
+        emb_x = get_emb(sin_inp_x)
+        emb = th.zeros((x, self.channels), device=tensor.device, dtype=tensor.dtype)
+        emb[:, : self.channels] = emb_x
+
+        self.cached_penc = emb[None, :, :orig_ch].repeat(batch_size, 1, 1)
+        return self.cached_penc
+
+
+
+class PositionalEncoding3D(nn.Module):
+    def __init__(self, channels):
+        """
+        :param channels: The last dimension of the tensor you want to apply pos emb to.
+        """
+        super(PositionalEncoding3D, self).__init__()
+        self.org_channels = channels
+        channels = int(np.ceil(channels / 6) * 2)
+        if channels % 2:
+            channels += 1
+        self.channels = channels
+        inv_freq = 1.0 / (10000 ** (th.arange(0, channels, 2).float() / channels))
+        self.register_buffer("inv_freq", inv_freq)
+        self.register_buffer("cached_penc", None, persistent=False)
+
+    def forward(self, tensor):
+        """
+        :param tensor: A 5d tensor of size (batch_size, x, y, z, ch)
+        :return: Positional Encoding Matrix of size (batch_size, x, y, z, ch)
+        """
+        if len(tensor.shape) != 5:
+            raise RuntimeError("The input tensor has to be 5d!")
+
+        if self.cached_penc is not None and self.cached_penc.shape == tensor.shape:
+            return self.cached_penc
+
+        self.cached_penc = None
+        batch_size, x, y, z, orig_ch = tensor.shape
+        pos_x = th.arange(x, device=tensor.device, dtype=self.inv_freq.dtype)
+        pos_y = th.arange(y, device=tensor.device, dtype=self.inv_freq.dtype)
+        pos_z = th.arange(z, device=tensor.device, dtype=self.inv_freq.dtype)
+        sin_inp_x = th.einsum("i,j->ij", pos_x, self.inv_freq)
+        sin_inp_y = th.einsum("i,j->ij", pos_y, self.inv_freq)
+        sin_inp_z = th.einsum("i,j->ij", pos_z, self.inv_freq)
+        emb_x = get_emb(sin_inp_x).unsqueeze(1).unsqueeze(1)
+        emb_y = get_emb(sin_inp_y).unsqueeze(1)
+        emb_z = get_emb(sin_inp_z)
+        emb = th.zeros(
+            (x, y, z, self.channels * 3),
+            device=tensor.device,
+            dtype=tensor.dtype,
+        )
+        emb[:, :, :, : self.channels] = emb_x
+        emb[:, :, :, self.channels : 2 * self.channels] = emb_y
+        emb[:, :, :, 2 * self.channels :] = emb_z
+
+        self.cached_penc = emb[None, :, :, :, :orig_ch].repeat(batch_size, 1, 1, 1, 1)
+        return self.cached_penc
+
+class AnalogyInputProcessor(ModelMixin, ConfigMixin):
+    
+    @register_to_config
+    def __init__(self,):
+        super(AnalogyInputProcessor, self).__init__()
+        
+        self.dino_transform = transforms.Compose(
+            [
+                transforms.Resize((DINO_SIZE, DINO_SIZE)),
+                transforms.ToTensor(),
+                transforms.Normalize(DINO_MEAN, DINO_STD), # SIGLIP normalization
+            ]
+        )
+        
+        self.siglip_transform = transforms.Compose(
+            [
+                transforms.Resize((SIGLIP_SIZE, SIGLIP_SIZE)),
+                transforms.ToTensor(),
+                transforms.Normalize(SIGLIP_MEAN, SIGLIP_STD), # SIGLIP normalization
+            ]
+        )
+        
+        dino_mean = th.tensor(DINO_MEAN).view(1, 3, 1, 1)
+        dino_std = th.tensor(DINO_STD).view(1, 3, 1, 1)
+        siglip_mean = [SIGLIP_MEAN[0],] * 3
+        siglip_std = [SIGLIP_STD[0],] * 3
+        siglip_mean = th.tensor(siglip_mean).view(1, 3, 1, 1)
+        siglip_std = th.tensor(siglip_std).view(1, 3, 1, 1)
+        self.register_buffer("dino_mean", dino_mean)
+        self.register_buffer("dino_std", dino_std)
+        self.register_buffer("siglip_mean", siglip_mean)
+        self.register_buffer("siglip_std", siglip_std)
+        
+    def __call__(self, analogy_prompt):
+        # List of tuples of (A, A*, B)
+        img_a_dino = []
+        img_a_siglip = []
+        img_a_star_dino = []
+        img_a_star_siglip = []
+        img_b_dino = []
+        img_b_siglip = []
+        
+        for im_set in analogy_prompt:
+            img_a, img_a_star, img_b = im_set
+            img_a_dino.append(self.dino_transform(img_a))
+            img_a_siglip.append(self.siglip_transform(img_a))
+            img_a_star_dino.append(self.dino_transform(img_a_star))
+            img_a_star_siglip.append(self.siglip_transform(img_a_star))
+            img_b_dino.append(self.dino_transform(img_b))
+            img_b_siglip.append(self.siglip_transform(img_b))
+        
+        img_a_dino = th.stack(img_a_dino, 0)
+        img_a_siglip = th.stack(img_a_siglip, 0)
+        img_a_star_dino = th.stack(img_a_star_dino, 0)
+        img_a_star_siglip = th.stack(img_a_star_siglip, 0)
+        img_b_dino = th.stack(img_b_dino, 0)
+        img_b_siglip = th.stack(img_b_siglip, 0)
+        
+        dino_combined_input = th.stack([img_b_dino, img_a_dino, img_a_star_dino], 0)
+        siglip_combined_input = th.stack([img_b_siglip, img_a_siglip, img_a_star_siglip], 0)
+        
+        return dino_combined_input, siglip_combined_input
+    def get_negative(self, dino_in, siglip_in):
+        
+        dino_i = ((dino_in * 0 + 0.5) - self.dino_mean) / self.dino_std
+        siglip_i = ((siglip_in * 0 + 0.5) - self.siglip_mean) / self.siglip_std
+        return dino_i, siglip_i
+            
+
+class AnalogyProjector(ModelMixin, ConfigMixin):
+    
+    @register_to_config
+    def __init__(self):
+        super(AnalogyProjector, self).__init__()
+        self.projector = DinoSiglipMixer()
+        self.pos_embd_1D = PositionalEncoding1D(OUT_SIZE)
+        self.pos_embd_3D = PositionalEncoding3D(OUT_SIZE)
+        
+
+    def forward(self, dino_in, siglip_in, batch_size):
+        
+        image_embeddings = self.projector(dino_in, siglip_in)
+        
+        image_embeddings = einops.rearrange(image_embeddings, '(k b) t d -> b k t d', b=batch_size)
+        image_embeddings = self.position_embd(image_embeddings)
+        return image_embeddings
+
+    def position_embd(self, image_embeddings, concat=False):
+        canvas_embd = image_embeddings[:, :, 1:, :]
+        batch_size = canvas_embd.shape[0]
+        type_size = canvas_embd.shape[1]
+        xy_size = canvas_embd.shape[2]
+        
+        x_size = int(xy_size ** 0.5)
+
+        canvas_embd = canvas_embd.reshape(batch_size, type_size, x_size, x_size, -1)
+        if concat:
+            canvas_embd = th.cat([canvas_embd, self.pos_embd_3D(canvas_embd)], -1)
+        else:
+            canvas_embd = self.pos_embd_3D(canvas_embd) + canvas_embd
+        canvas_embd = canvas_embd.reshape(batch_size, type_size, xy_size, -1)
+
+        class_embd = image_embeddings[:, :, 0, :]
+        if concat:
+            class_embd = th.cat([class_embd, self.pos_embd_1D(class_embd)], -1)
+        else:    
+            class_embd = self.pos_embd_1D(class_embd) + class_embd
+        all_embd_list = []
+        for i in range(type_size):
+            all_embd_list.append(class_embd[:, i:i+1])
+            all_embd_list.append(canvas_embd[:, i])
+        image_embeddings = th.cat(all_embd_list, 1)
+        return image_embeddings
+
+
+class HighLowMixer(th.nn.Module):
+    def __init__(self, in_size=IN_SIZE, out_size=OUT_SIZE):
+        super().__init__()
+        mid_size = (in_size + out_size) // 2
+        
+        self.lower_projector = th.nn.Sequential(
+            th.nn.LayerNorm(IN_SIZE//2),
+            th.nn.SiLU()
+        )
+        self.upper_projector = th.nn.Sequential(
+            th.nn.LayerNorm(IN_SIZE//2),
+            th.nn.SiLU()
+        )
+        self.projectors = th.nn.ModuleList([
+            # add layer norm
+            th.nn.Linear(in_size, mid_size),
+            th.nn.SiLU(),
+            th.nn.Linear(mid_size, out_size)
+        ])
+        # initialize
+        for proj in self.projectors:
+            if isinstance(proj, th.nn.Linear):
+                th.nn.init.xavier_uniform_(proj.weight)
+                th.nn.init.zeros_(proj.bias)
+
+    def forward(self, lower_in, upper_in, ):
+        # ALso format lower_in
+        lower_in = self.lower_projector(lower_in)
+        upper_in = self.upper_projector(upper_in)
+        x = th.cat([lower_in, upper_in], -1)
+        for proj in self.projectors:
+            x = proj(x)
+        return x
+
+class DinoSiglipMixer(th.nn.Module):
+    def __init__(self, in_size=OUT_SIZE * 2, out_size=OUT_SIZE):
+        super().__init__()
+        self.dino_projector = HighLowMixer()
+        self.siglip_projector = HighLowMixer()
+        self.projectors = th.nn.Sequential(
+            th.nn.SiLU(),
+            th.nn.Linear(in_size, out_size),
+        )
+        # initialize
+        for proj in self.projectors:
+            if isinstance(proj, th.nn.Linear):
+                th.nn.init.xavier_uniform_(proj.weight)
+                th.nn.init.zeros_(proj.bias)
+
+    
+    def forward(self, dino_in, siglip_in):
+        # ALso format lower_in
+        lower, upper = th.chunk(dino_in, 2, -1)
+        dino_out = self.dino_projector(lower, upper)
+        lower, upper = th.chunk(siglip_in, 2, -1)
+        siglip_out = self.siglip_projector(lower, upper)
+        x = th.cat([dino_out, siglip_out], -1)
+        for proj in self.projectors:
+            x = proj(x)
+        return x
+
+class AnalogyEncoder(ModelMixin, ConfigMixin):
+    @register_to_config
+    def __init__(self, load_pretrained=False, 
+                 dino_config_dict=None, siglip_config_dict=None):
+        super().__init__()
+        if load_pretrained:
+            image_encoder_dino = AutoModel.from_pretrained('facebook/dinov2-large', torch_dtype=th.float16)
+            image_encoder_siglip = SiglipVisionModel.from_pretrained("google/siglip-large-patch16-256", torch_dtype=th.float16, attn_implementation="sdpa")
+        else:
+            image_encoder_dino = AutoModel.from_config(Dinov2Config.from_dict(dino_config_dict))
+            image_encoder_siglip = AutoModel.from_config(SiglipVisionConfig.from_dict(siglip_config_dict))
+            
+        image_encoder_dino.requires_grad_(False)
+        image_encoder_dino = image_encoder_dino.to(memory_format=th.channels_last)
+
+        image_encoder_siglip.requires_grad_(False)
+        image_encoder_siglip = image_encoder_siglip.to(memory_format=th.channels_last)
+        self.image_encoder_dino = image_encoder_dino
+        self.image_encoder_siglip = image_encoder_siglip
+
+
+    def dino_normalization(self, encoder_output):
+        embeds = encoder_output.last_hidden_state
+        embeds_pooled = embeds[:, 0:1]
+        embeds = embeds / th.norm(embeds_pooled, dim=-1, keepdim=True)
+        return embeds
+    
+    def siglip_normalization(self, encoder_output):
+        embeds = th.cat ([encoder_output.pooler_output[:, None, :], encoder_output.last_hidden_state], dim=1)
+        embeds_pooled = embeds[:, 0:1]
+        embeds = embeds / th.norm(embeds_pooled, dim=-1, keepdim=True)
+        return embeds
+    
+    def forward(self, dino_in, siglip_in):
+
+        x_1 = self.image_encoder_dino(dino_in, output_hidden_states=True)
+        x_1_first = x_1.hidden_states[0]
+        x_1 = self.dino_normalization(x_1)
+        x_2 = self.image_encoder_siglip(siglip_in, output_hidden_states=True)
+        x_2_first = x_2.hidden_states[0]
+        x_2_first_pool = th.mean(x_2_first, dim=1, keepdim=True)
+        x_2_first = th.cat([x_2_first_pool, x_2_first], 1)
+        x_2 = self.siglip_normalization(x_2)
+        dino_embd = th.cat([x_1, x_1_first], -1)
+        siglip_embd = th.cat([x_2, x_2_first], -1)
+        return dino_embd, siglip_embd
+    
+
+class PatternAnalogyTrifuser(DiffusionPipeline):
+    r"""
+    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.)
+    """
+
+    model_cpu_offload_seq = "bert->unet->vqvae"
+
+    analogy_input_processor: AnalogyInputProcessor
+    analogy_encoder: AnalogyEncoder
+    analogy_projector: AnalogyProjector
+    unet: UNet2DConditionModel
+    vae: AutoencoderKL
+    scheduler: KarrasDiffusionSchedulers
+    
+    def __init__(self, 
+        analogy_input_processor: AnalogyInputProcessor,
+        analogy_projector: AnalogyProjector,
+        analogy_encoder: AnalogyEncoder,
+        unet: UNet2DConditionModel,
+        vae: AutoencoderKL,
+        scheduler: KarrasDiffusionSchedulers,):
+        
+        
+        super().__init__()
+        self.register_modules(
+            analogy_input_processor=analogy_input_processor,
+            analogy_encoder=analogy_encoder,
+            analogy_projector=analogy_projector,
+            unet=unet,
+            vae=vae,
+            scheduler=scheduler,
+        )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_image_variation.StableDiffusionImageVariationPipeline.check_inputs
+    def check_inputs(self, analogy_prompt, negative_analogy_prompt, height, width, callback_steps):
+        if (
+            not isinstance(analogy_prompt, th.Tensor)
+            and not isinstance(analogy_prompt, PIL.Image.Image)
+            and not isinstance(analogy_prompt, list)
+        ):
+            raise ValueError(
+                "`analogy_prompt` contents have to be of type `torch.Tensor` or `PIL.Image.Image` or `List[PIL.Image.Image]` but is"
+                f" {type(analogy_prompt)}"
+            )
+        if not negative_analogy_prompt is None:
+            if (
+                not isinstance(negative_analogy_prompt, th.Tensor)
+                and not isinstance(negative_analogy_prompt, PIL.Image.Image)
+                and not isinstance(negative_analogy_prompt, list)
+            ):
+                raise ValueError(
+                    "`negative_analogy_prompt` contents have to be of type `torch.Tensor` or `PIL.Image.Image` or `List[PIL.Image.Image]` but is"
+                    f" {type(negative_analogy_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)}."
+            )
+
+    # 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,
+            int(height) // self.vae_scale_factor,
+            int(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
+    
+    # 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
+
+    # 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,
+            int(height) // self.vae_scale_factor,
+            int(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 _encode_prompt(self, analogy_prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`):
+                prompt to be encoded
+            device: (`torch.device`):
+                torch device
+            num_images_per_prompt (`int`):
+                number of images that should be generated per prompt
+            do_classifier_free_guidance (`bool`):
+                whether to use classifier free guidance or not
+            negative_prompt (`str` or `List[str]`):
+                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`).
+        """
+        weight_dtype = self.unet.dtype
+        dino_input, siglip_input = self.analogy_input_processor(analogy_prompt)
+        dino_input = dino_input.to(device=device).to(dtype=weight_dtype)
+        siglip_input = siglip_input.to(device=device).to(dtype=weight_dtype)
+        batch_size = dino_input.shape[1]
+        dino_input_reshaped = einops.rearrange(dino_input, "k b c h w -> (k b) c h w")
+        siglip_input_reshaped = einops.rearrange(siglip_input, "k b c h w -> (k b) c h w")
+        dino_enc, siglip_enc = self.analogy_encoder(dino_input_reshaped, siglip_input_reshaped)
+        image_embeddings = self.analogy_projector(dino_enc, siglip_enc, batch_size)
+        # Check size here.
+        
+        bs_embed, seq_len, _ = image_embeddings.shape
+        image_embeddings = image_embeddings.repeat(num_images_per_prompt, 1, 1)
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance:
+            uncond_images: List[str]
+            if negative_prompt is None:
+                uncond_images = [np.zeros((512, 512, 3)) + 0.5] * batch_size
+            elif type(negative_prompt) is not type(analogy_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(analogy_prompt)} !="
+                    f" {type(negative_prompt)}."
+                )
+            elif isinstance(negative_prompt, PIL.Image.Image):
+                uncond_images = [negative_prompt]
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {analogy_prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            else:
+                uncond_images = negative_prompt
+            dino_neg, siglip_neg = self.analogy_input_processor.get_negative(dino_input, siglip_input)
+            
+            dino_neg = dino_neg.to(device=device).to(dtype=weight_dtype)
+            siglip_neg = siglip_neg.to(device=device).to(dtype=weight_dtype)
+            dino_neg_reshaped = einops.rearrange(dino_neg, "k b c h w -> (k b) c h w")
+            siglip_neg_reshaped = einops.rearrange(siglip_neg, "k b c h w -> (k b) c h w")
+            dino_neg_enc, siglip_neg_enc = self.analogy_encoder(dino_neg_reshaped, siglip_neg_reshaped)
+            negative_prompt_embeds = self.analogy_projector(dino_neg_enc, siglip_neg_enc, batch_size)
+                
+            negative_prompt_embeds = negative_prompt_embeds.repeat(num_images_per_prompt, 1, 1)
+            image_embeddings = th.cat([negative_prompt_embeds, image_embeddings])
+
+
+        return image_embeddings
+    
+    @th.no_grad()
+    def __call__(
+        self,
+        analogy_prompt: Union[str, List[str]] = None,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 7.5,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        negative_analogy_prompt: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[Union[th.Generator, List[th.Generator]]] = None,
+        latents: Optional[th.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, th.Tensor], None]] = None,
+        callback_steps: int = 1,
+        start_step: int = 0,
+    ):
+        r"""
+        The call function to the pipeline for generation.
+
+        Args:
+            image (`PIL.Image.Image`, `List[PIL.Image.Image]` or `torch.Tensor`):
+                The image prompt or prompts to guide the image generation.
+            height (`int`, *optional*, defaults to `self.image_unet.config.sample_size * self.vae_scale_factor`):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to `self.image_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):
+                A higher guidance scale value encourages the model to generate images closely linked to the text
+                `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+                pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 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) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+                to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+            generator (`torch.Generator`, *optional*):
+                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+                generation deterministic.
+            latents (`torch.Tensor`, *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 is generated by sampling using the supplied random `generator`.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generated image. Choose between `PIL.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 calls every `callback_steps` steps during inference. The function is called with the
+                following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`.
+            callback_steps (`int`, *optional*, defaults to 1):
+                The frequency at which the `callback` function is called. If not specified, the callback is called at
+                every step.
+
+        Examples:
+
+        ```py
+        >>> from diffusers import VersatileDiffusionImageVariationPipeline
+        >>> import torch
+        >>> import requests
+        >>> from io import BytesIO
+        >>> from PIL import Image
+
+        >>> # let's download an initial image
+        >>> url = "https://huggingface.co/datasets/diffusers/images/resolve/main/benz.jpg"
+
+        >>> response = requests.get(url)
+        >>> image = Image.open(BytesIO(response.content)).convert("RGB")
+
+        >>> pipe = VersatileDiffusionImageVariationPipeline.from_pretrained(
+        ...     "shi-labs/versatile-diffusion", torch_dtype=torch.float16
+        ... )
+        >>> pipe = pipe.to("cuda")
+
+        >>> generator = torch.Generator(device="cuda").manual_seed(0)
+        >>> image = pipe(image, generator=generator).images[0]
+        >>> image.save("./car_variation.png")
+        ```
+
+        Returns:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+                otherwise a `tuple` is returned where the first element is a list with the generated images.
+        """
+
+        # 1. Check inputs. Raise error if not correct
+        height = height or self.unet.config.sample_size * self.vae_scale_factor
+        width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(analogy_prompt, negative_analogy_prompt, height, width, callback_steps)
+        
+        # 2. Define call parameters
+        if isinstance(analogy_prompt, list):
+            batch_size = len(analogy_prompt)
+        elif isinstance(analogy_prompt, tuple):
+            batch_size = 1
+        else:
+            raise ValueError(
+                f"`analogy_prompt` has to be a list of images or a tuple of images but is of type {type(analogy_prompt)}"
+            )
+        device = self._execution_device
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        # 3. Encode input prompt
+        analogy_embeddings = self._encode_prompt(
+            analogy_prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_analogy_prompt
+        )
+
+        # 4. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        
+        timesteps = self.scheduler.timesteps
+        # Now this should be from start step onwards
+        timesteps = timesteps[start_step:]
+        # 5. Prepare latent variables
+        num_channels_latents = self.unet.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            analogy_embeddings.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. Prepare extra step kwargs.
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 7. Denoising loop
+        for i, t in enumerate(self.progress_bar(timesteps)):
+            # expand the latents if we are doing classifier free guidance
+            latent_model_input = th.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=analogy_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:
+                step_idx = i // getattr(self.scheduler, "order", 1)
+                callback(step_idx, t, latents)
+
+        if not output_type == "latent":
+            image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+        else:
+            image = latents
+
+        image = self.image_processor.postprocess(image, output_type=output_type)
+
+        if not return_dict:
+            return (image,)
+
+        return ImagePipelineOutput(images=image)