diffusers 0.23.1__py3-none-any.whl → 0.25.0__py3-none-any.whl

diffusers/pipelines/{stable_diffusion → stable_diffusion_panorama}/pipeline_stable_diffusion_panorama.py

@@ -16,11 +16,11 @@ import inspect
 from typing import Any, Callable, Dict, List, Optional, Union
 
 import torch
-from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
 
-from ...image_processor import VaeImageProcessor
-from ...loaders import LoraLoaderMixin, TextualInversionLoaderMixin
-from ...models import AutoencoderKL, UNet2DConditionModel
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
 from ...models.lora import adjust_lora_scale_text_encoder
 from ...schedulers import DDIMScheduler
 from ...utils import (
@@ -33,8 +33,8 @@ from ...utils import (
 )
 from ...utils.torch_utils import randn_tensor
 from ..pipeline_utils import DiffusionPipeline
-from . import StableDiffusionPipelineOutput
-from .safety_checker import StableDiffusionSafetyChecker
+from ..stable_diffusion import StableDiffusionPipelineOutput
+from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
@@ -59,13 +59,19 @@ EXAMPLE_DOC_STRING = """
 """
 
 
-class StableDiffusionPanoramaPipeline(DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin):
+class StableDiffusionPanoramaPipeline(DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, IPAdapterMixin):
     r"""
     Pipeline for text-to-image generation using MultiDiffusion.
 
     This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
     implemented for all pipelines (downloading, saving, running on a particular device, etc.).
 
+    The pipeline also inherits the following loading methods:
+        - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+        - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+        - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+        - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
     Args:
         vae ([`AutoencoderKL`]):
             Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
@@ -85,8 +91,9 @@ class StableDiffusionPanoramaPipeline(DiffusionPipeline, TextualInversionLoaderM
         feature_extractor ([`~transformers.CLIPImageProcessor`]):
             A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
     """
+
     model_cpu_offload_seq = "text_encoder->unet->vae"
-    _optional_components = ["safety_checker", "feature_extractor"]
+    _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
     _exclude_from_cpu_offload = ["safety_checker"]
 
     def __init__(
@@ -98,6 +105,7 @@ class StableDiffusionPanoramaPipeline(DiffusionPipeline, TextualInversionLoaderM
         scheduler: DDIMScheduler,
         safety_checker: StableDiffusionSafetyChecker,
         feature_extractor: CLIPImageProcessor,
+        image_encoder: Optional[CLIPVisionModelWithProjection] = None,
         requires_safety_checker: bool = True,
     ):
         super().__init__()
@@ -126,6 +134,7 @@ class StableDiffusionPanoramaPipeline(DiffusionPipeline, TextualInversionLoaderM
             scheduler=scheduler,
             safety_checker=safety_checker,
             feature_extractor=feature_extractor,
+            image_encoder=image_encoder,
         )
         self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
         self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
@@ -362,6 +371,31 @@ class StableDiffusionPanoramaPipeline(DiffusionPipeline, TextualInversionLoaderM
 
         return prompt_embeds, negative_prompt_embeds
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+    def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+        dtype = next(self.image_encoder.parameters()).dtype
+
+        if not isinstance(image, torch.Tensor):
+            image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+        image = image.to(device=device, dtype=dtype)
+        if output_hidden_states:
+            image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+            image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_enc_hidden_states = self.image_encoder(
+                torch.zeros_like(image), output_hidden_states=True
+            ).hidden_states[-2]
+            uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+                num_images_per_prompt, dim=0
+            )
+            return image_enc_hidden_states, uncond_image_enc_hidden_states
+        else:
+            image_embeds = self.image_encoder(image).image_embeds
+            image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_embeds = torch.zeros_like(image_embeds)
+
+            return image_embeds, uncond_image_embeds
+
     # 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:
@@ -528,6 +562,7 @@ class StableDiffusionPanoramaPipeline(DiffusionPipeline, TextualInversionLoaderM
         latents: Optional[torch.FloatTensor] = None,
         prompt_embeds: Optional[torch.FloatTensor] = None,
         negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        ip_adapter_image: Optional[PipelineImageInput] = None,
         output_type: Optional[str] = "pil",
         return_dict: bool = True,
         callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
@@ -577,6 +612,8 @@ class StableDiffusionPanoramaPipeline(DiffusionPipeline, TextualInversionLoaderM
             negative_prompt_embeds (`torch.FloatTensor`, *optional*):
                 Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
                 not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+            ip_adapter_image: (`PipelineImageInput`, *optional*):
+                Optional image input to work with IP Adapters.
             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`):
@@ -631,6 +668,14 @@ class StableDiffusionPanoramaPipeline(DiffusionPipeline, TextualInversionLoaderM
         # corresponds to doing no classifier free guidance.
         do_classifier_free_guidance = guidance_scale > 1.0
 
+        if ip_adapter_image is not None:
+            output_hidden_state = False if isinstance(self.unet.encoder_hid_proj, ImageProjection) else True
+            image_embeds, negative_image_embeds = self.encode_image(
+                ip_adapter_image, device, num_images_per_prompt, output_hidden_state
+            )
+            if do_classifier_free_guidance:
+                image_embeds = torch.cat([negative_image_embeds, image_embeds])
+
         # 3. Encode input prompt
         text_encoder_lora_scale = (
             cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
@@ -680,6 +725,9 @@ class StableDiffusionPanoramaPipeline(DiffusionPipeline, TextualInversionLoaderM
         # 7. 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)
 
+        # 7.1 Add image embeds for IP-Adapter
+        added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None
+
         # 8. Denoising loop
         # Each denoising step also includes refinement of the latents with respect to the
         # views.
@@ -742,6 +790,7 @@ class StableDiffusionPanoramaPipeline(DiffusionPipeline, TextualInversionLoaderM
                         t,
                         encoder_hidden_states=prompt_embeds_input,
                         cross_attention_kwargs=cross_attention_kwargs,
+                        added_cond_kwargs=added_cond_kwargs,
                     ).sample
 
                     # perform guidance

diffusers/pipelines/stable_diffusion_safe/pipeline_stable_diffusion_safe.py

@@ -5,10 +5,12 @@ from typing import Callable, List, Optional, Union
 import numpy as np
 import torch
 from packaging import version
-from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
 
 from ...configuration_utils import FrozenDict
-from ...models import AutoencoderKL, UNet2DConditionModel
+from ...image_processor import PipelineImageInput
+from ...loaders import IPAdapterMixin
+from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
 from ...schedulers import KarrasDiffusionSchedulers
 from ...utils import deprecate, logging
 from ...utils.torch_utils import randn_tensor
@@ -20,13 +22,16 @@ from .safety_checker import SafeStableDiffusionSafetyChecker
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
-class StableDiffusionPipelineSafe(DiffusionPipeline):
+class StableDiffusionPipelineSafe(DiffusionPipeline, IPAdapterMixin):
     r"""
     Pipeline based on the [`StableDiffusionPipeline`] for text-to-image generation using Safe Latent Diffusion.
 
     This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
     implemented for all pipelines (downloading, saving, running on a particular device, etc.).
 
+    The pipeline also inherits the following loading methods:
+        - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
     Args:
         vae ([`AutoencoderKL`]):
             Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
@@ -48,7 +53,7 @@ class StableDiffusionPipelineSafe(DiffusionPipeline):
     """
 
     model_cpu_offload_seq = "text_encoder->unet->vae"
-    _optional_components = ["safety_checker", "feature_extractor"]
+    _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
 
     def __init__(
         self,
@@ -59,6 +64,7 @@ class StableDiffusionPipelineSafe(DiffusionPipeline):
         scheduler: KarrasDiffusionSchedulers,
         safety_checker: SafeStableDiffusionSafetyChecker,
         feature_extractor: CLIPImageProcessor,
+        image_encoder: Optional[CLIPVisionModelWithProjection] = None,
         requires_safety_checker: bool = True,
     ):
         super().__init__()
@@ -140,6 +146,7 @@ class StableDiffusionPipelineSafe(DiffusionPipeline):
             scheduler=scheduler,
             safety_checker=safety_checker,
             feature_extractor=feature_extractor,
+            image_encoder=image_encoder,
         )
         self._safety_text_concept = safety_concept
         self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
@@ -467,6 +474,31 @@ class StableDiffusionPipelineSafe(DiffusionPipeline):
                 noise_guidance = noise_guidance - noise_guidance_safety
         return noise_guidance, safety_momentum
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+    def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+        dtype = next(self.image_encoder.parameters()).dtype
+
+        if not isinstance(image, torch.Tensor):
+            image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+        image = image.to(device=device, dtype=dtype)
+        if output_hidden_states:
+            image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+            image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_enc_hidden_states = self.image_encoder(
+                torch.zeros_like(image), output_hidden_states=True
+            ).hidden_states[-2]
+            uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+                num_images_per_prompt, dim=0
+            )
+            return image_enc_hidden_states, uncond_image_enc_hidden_states
+        else:
+            image_embeds = self.image_encoder(image).image_embeds
+            image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_embeds = torch.zeros_like(image_embeds)
+
+            return image_embeds, uncond_image_embeds
+
     @torch.no_grad()
     def __call__(
         self,
@@ -480,6 +512,7 @@ class StableDiffusionPipelineSafe(DiffusionPipeline):
         eta: float = 0.0,
         generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
         latents: Optional[torch.FloatTensor] = None,
+        ip_adapter_image: Optional[PipelineImageInput] = None,
         output_type: Optional[str] = "pil",
         return_dict: bool = True,
         callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
@@ -521,6 +554,8 @@ class StableDiffusionPipelineSafe(DiffusionPipeline):
                 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`.
+            ip_adapter_image: (`PipelineImageInput`, *optional*):
+                Optional image input to work with IP Adapters.
             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`):
@@ -560,10 +595,11 @@ class StableDiffusionPipelineSafe(DiffusionPipeline):
         ```py
         import torch
         from diffusers import StableDiffusionPipelineSafe
+        from diffusers.pipelines.stable_diffusion_safe import SafetyConfig
 
         pipeline = StableDiffusionPipelineSafe.from_pretrained(
             "AIML-TUDA/stable-diffusion-safe", torch_dtype=torch.float16
-        )
+        ).to("cuda")
         prompt = "the four horsewomen of the apocalypse, painting by tom of finland, gaston bussiere, craig mullins, j. c. leyendecker"
         image = pipeline(prompt=prompt, **SafetyConfig.MEDIUM).images[0]
         ```
@@ -588,6 +624,17 @@ class StableDiffusionPipelineSafe(DiffusionPipeline):
         if not enable_safety_guidance:
             warnings.warn("Safety checker disabled!")
 
+        if ip_adapter_image is not None:
+            output_hidden_state = False if isinstance(self.unet.encoder_hid_proj, ImageProjection) else True
+            image_embeds, negative_image_embeds = self.encode_image(
+                ip_adapter_image, device, num_images_per_prompt, output_hidden_state
+            )
+            if do_classifier_free_guidance:
+                if enable_safety_guidance:
+                    image_embeds = torch.cat([negative_image_embeds, image_embeds, image_embeds])
+                else:
+                    image_embeds = torch.cat([negative_image_embeds, image_embeds])
+
         # 3. Encode input prompt
         prompt_embeds = self._encode_prompt(
             prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt, enable_safety_guidance
@@ -613,6 +660,9 @@ class StableDiffusionPipelineSafe(DiffusionPipeline):
         # 6. Prepare extra step kwargs.
         extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
 
+        # 6.1 Add image embeds for IP-Adapter
+        added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None
+
         safety_momentum = None
 
         num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
@@ -627,7 +677,9 @@ class StableDiffusionPipelineSafe(DiffusionPipeline):
                 latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
 
                 # predict the noise residual
-                noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=prompt_embeds).sample
+                noise_pred = self.unet(
+                    latent_model_input, t, encoder_hidden_states=prompt_embeds, added_cond_kwargs=added_cond_kwargs
+                ).sample
 
                 # perform guidance
                 if do_classifier_free_guidance:

diffusers/pipelines/stable_diffusion_sag/__init__.py

@@ -0,0 +1,48 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+    DIFFUSERS_SLOW_IMPORT,
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    get_objects_from_module,
+    is_torch_available,
+    is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+
+try:
+    if not (is_transformers_available() and is_torch_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from ...utils import dummy_torch_and_transformers_objects  # noqa F403
+
+    _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+    _import_structure["pipeline_stable_diffusion_sag"] = ["StableDiffusionSAGPipeline"]
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+    try:
+        if not (is_transformers_available() and is_torch_available()):
+            raise OptionalDependencyNotAvailable()
+
+    except OptionalDependencyNotAvailable:
+        from ...utils.dummy_torch_and_transformers_objects import *
+    else:
+        from .pipeline_stable_diffusion_sag import StableDiffusionSAGPipeline
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(
+        __name__,
+        globals()["__file__"],
+        _import_structure,
+        module_spec=__spec__,
+    )
+
+    for name, value in _dummy_objects.items():
+        setattr(sys.modules[__name__], name, value)

diffusers/pipelines/{stable_diffusion → stable_diffusion_sag}/pipeline_stable_diffusion_sag.py

@@ -17,11 +17,11 @@ from typing import Any, Callable, Dict, List, Optional, Union
 
 import torch
 import torch.nn.functional as F
-from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
 
-from ...image_processor import VaeImageProcessor
-from ...loaders import LoraLoaderMixin, TextualInversionLoaderMixin
-from ...models import AutoencoderKL, UNet2DConditionModel
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
 from ...models.lora import adjust_lora_scale_text_encoder
 from ...schedulers import KarrasDiffusionSchedulers
 from ...utils import (
@@ -34,8 +34,8 @@ from ...utils import (
 )
 from ...utils.torch_utils import randn_tensor
 from ..pipeline_utils import DiffusionPipeline
-from . import StableDiffusionPipelineOutput
-from .safety_checker import StableDiffusionSafetyChecker
+from ..stable_diffusion import StableDiffusionPipelineOutput
+from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
@@ -98,13 +98,17 @@ class CrossAttnStoreProcessor:
 
 
 # Modified to get self-attention guidance scale in this paper (https://arxiv.org/pdf/2210.00939.pdf) as an input
-class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin):
+class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin, IPAdapterMixin):
     r"""
     Pipeline for text-to-image generation using Stable Diffusion.
 
     This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
     implemented for all pipelines (downloading, saving, running on a particular device, etc.).
 
+    The pipeline also inherits the following loading methods:
+        - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+        - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
     Args:
         vae ([`AutoencoderKL`]):
             Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
@@ -124,8 +128,9 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin)
         feature_extractor ([`~transformers.CLIPImageProcessor`]):
             A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
     """
+
     model_cpu_offload_seq = "text_encoder->unet->vae"
-    _optional_components = ["safety_checker", "feature_extractor"]
+    _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
     _exclude_from_cpu_offload = ["safety_checker"]
 
     def __init__(
@@ -137,6 +142,7 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin)
         scheduler: KarrasDiffusionSchedulers,
         safety_checker: StableDiffusionSafetyChecker,
         feature_extractor: CLIPImageProcessor,
+        image_encoder: Optional[CLIPVisionModelWithProjection] = None,
         requires_safety_checker: bool = True,
     ):
         super().__init__()
@@ -149,6 +155,7 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin)
             scheduler=scheduler,
             safety_checker=safety_checker,
             feature_extractor=feature_extractor,
+            image_encoder=image_encoder,
         )
         self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
         self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
@@ -385,6 +392,31 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin)
 
         return prompt_embeds, negative_prompt_embeds
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+    def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+        dtype = next(self.image_encoder.parameters()).dtype
+
+        if not isinstance(image, torch.Tensor):
+            image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+        image = image.to(device=device, dtype=dtype)
+        if output_hidden_states:
+            image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+            image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_enc_hidden_states = self.image_encoder(
+                torch.zeros_like(image), output_hidden_states=True
+            ).hidden_states[-2]
+            uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+                num_images_per_prompt, dim=0
+            )
+            return image_enc_hidden_states, uncond_image_enc_hidden_states
+        else:
+            image_embeds = self.image_encoder(image).image_embeds
+            image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_embeds = torch.zeros_like(image_embeds)
+
+            return image_embeds, uncond_image_embeds
+
     # 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:
@@ -518,6 +550,7 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin)
         latents: Optional[torch.FloatTensor] = None,
         prompt_embeds: Optional[torch.FloatTensor] = None,
         negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        ip_adapter_image: Optional[PipelineImageInput] = None,
         output_type: Optional[str] = "pil",
         return_dict: bool = True,
         callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
@@ -564,6 +597,8 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin)
             negative_prompt_embeds (`torch.FloatTensor`, *optional*):
                 Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
                 not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+            ip_adapter_image: (`PipelineImageInput`, *optional*):
+                Optional image input to work with IP Adapters.
             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`):
@@ -617,6 +652,14 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin)
         # `sag_scale = 0` means no self-attention guidance
         do_self_attention_guidance = sag_scale > 0.0
 
+        if ip_adapter_image is not None:
+            output_hidden_state = False if isinstance(self.unet.encoder_hid_proj, ImageProjection) else True
+            image_embeds, negative_image_embeds = self.encode_image(
+                ip_adapter_image, device, num_images_per_prompt, output_hidden_state
+            )
+            if do_classifier_free_guidance:
+                image_embeds = torch.cat([negative_image_embeds, image_embeds])
+
         # 3. Encode input prompt
         prompt_embeds, negative_prompt_embeds = self.encode_prompt(
             prompt,
@@ -654,6 +697,10 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin)
         # 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)
 
+        # 6.1 Add image embeds for IP-Adapter
+        added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None
+        added_uncond_kwargs = {"image_embeds": negative_image_embeds} if ip_adapter_image is not None else None
+
         # 7. Denoising loop
         store_processor = CrossAttnStoreProcessor()
         self.unet.mid_block.attentions[0].transformer_blocks[0].attn1.processor = store_processor
@@ -679,6 +726,7 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin)
                         t,
                         encoder_hidden_states=prompt_embeds,
                         cross_attention_kwargs=cross_attention_kwargs,
+                        added_cond_kwargs=added_cond_kwargs,
                     ).sample
 
                     # perform guidance
@@ -702,7 +750,12 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin)
                             )
                             uncond_emb, _ = prompt_embeds.chunk(2)
                             # forward and give guidance
-                            degraded_pred = self.unet(degraded_latents, t, encoder_hidden_states=uncond_emb).sample
+                            degraded_pred = self.unet(
+                                degraded_latents,
+                                t,
+                                encoder_hidden_states=uncond_emb,
+                                added_cond_kwargs=added_uncond_kwargs,
+                            ).sample
                             noise_pred += sag_scale * (noise_pred_uncond - degraded_pred)
                         else:
                             # DDIM-like prediction of x0
@@ -714,7 +767,12 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, TextualInversionLoaderMixin)
                                 pred_x0, cond_attn, map_size, t, self.pred_epsilon(latents, noise_pred, t)
                             )
                             # forward and give guidance
-                            degraded_pred = self.unet(degraded_latents, t, encoder_hidden_states=prompt_embeds).sample
+                            degraded_pred = self.unet(
+                                degraded_latents,
+                                t,
+                                encoder_hidden_states=prompt_embeds,
+                                added_cond_kwargs=added_cond_kwargs,
+                            ).sample
                             noise_pred += sag_scale * (noise_pred - degraded_pred)
 
                     # compute the previous noisy sample x_t -> x_t-1