diffusers 0.29.2__py3-none-any.whl → 0.30.0__py3-none-any.whl

diffusers/pipelines/stable_diffusion/convert_from_ckpt.py

@@ -1370,6 +1370,8 @@ def download_from_original_stable_diffusion_ckpt(
 
     if "unet_config" in original_config["model"]["params"]:
         original_config["model"]["params"]["unet_config"]["params"]["in_channels"] = num_in_channels
+    elif "network_config" in original_config["model"]["params"]:
+        original_config["model"]["params"]["network_config"]["params"]["in_channels"] = num_in_channels
 
     if (
         "parameterization" in original_config["model"]["params"]

diffusers/pipelines/stable_diffusion/pipeline_flax_stable_diffusion.py

@@ -55,7 +55,7 @@ EXAMPLE_DOC_STRING = """
         >>> from diffusers import FlaxStableDiffusionPipeline
 
         >>> pipeline, params = FlaxStableDiffusionPipeline.from_pretrained(
-        ...     "runwayml/stable-diffusion-v1-5", revision="bf16", dtype=jax.numpy.bfloat16
+        ...     "runwayml/stable-diffusion-v1-5", variant="bf16", dtype=jax.numpy.bfloat16
         ... )
 
         >>> prompt = "a photo of an astronaut riding a horse on mars"

diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py

@@ -21,7 +21,7 @@ from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPV
 from ...callbacks import MultiPipelineCallbacks, PipelineCallback
 from ...configuration_utils import FrozenDict
 from ...image_processor import PipelineImageInput, VaeImageProcessor
-from ...loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ...loaders import FromSingleFileMixin, IPAdapterMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin
 from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
 from ...models.lora import adjust_lora_scale_text_encoder
 from ...schedulers import KarrasDiffusionSchedulers
@@ -133,7 +133,7 @@ class StableDiffusionPipeline(
     DiffusionPipeline,
     StableDiffusionMixin,
     TextualInversionLoaderMixin,
-    LoraLoaderMixin,
+    StableDiffusionLoraLoaderMixin,
     IPAdapterMixin,
     FromSingleFileMixin,
 ):
@@ -145,8 +145,8 @@ class StableDiffusionPipeline(
 
     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.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+        - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights
         - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
         - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
 
@@ -342,7 +342,7 @@ class StableDiffusionPipeline(
         """
         # set lora scale so that monkey patched LoRA
         # function of text encoder can correctly access it
-        if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+        if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin):
             self._lora_scale = lora_scale
 
             # dynamically adjust the LoRA scale
@@ -475,7 +475,7 @@ class StableDiffusionPipeline(
             negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
 
         if self.text_encoder is not None:
-            if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+            if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND:
                 # Retrieve the original scale by scaling back the LoRA layers
                 unscale_lora_layers(self.text_encoder, lora_scale)
 
@@ -508,6 +508,9 @@ class StableDiffusionPipeline(
     def prepare_ip_adapter_image_embeds(
         self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
     ):
+        image_embeds = []
+        if do_classifier_free_guidance:
+            negative_image_embeds = []
         if ip_adapter_image_embeds is None:
             if not isinstance(ip_adapter_image, list):
                 ip_adapter_image = [ip_adapter_image]
@@ -517,7 +520,6 @@ class StableDiffusionPipeline(
                     f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
                 )
 
-            image_embeds = []
             for single_ip_adapter_image, image_proj_layer in zip(
                 ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
             ):
@@ -525,36 +527,28 @@ class StableDiffusionPipeline(
                 single_image_embeds, single_negative_image_embeds = self.encode_image(
                     single_ip_adapter_image, device, 1, output_hidden_state
                 )
-                single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
-                single_negative_image_embeds = torch.stack(
-                    [single_negative_image_embeds] * num_images_per_prompt, dim=0
-                )
 
+                image_embeds.append(single_image_embeds[None, :])
                 if do_classifier_free_guidance:
-                    single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
-                    single_image_embeds = single_image_embeds.to(device)
-
-                image_embeds.append(single_image_embeds)
+                    negative_image_embeds.append(single_negative_image_embeds[None, :])
         else:
-            repeat_dims = [1]
-            image_embeds = []
             for single_image_embeds in ip_adapter_image_embeds:
                 if do_classifier_free_guidance:
                     single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
-                    single_image_embeds = single_image_embeds.repeat(
-                        num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
-                    )
-                    single_negative_image_embeds = single_negative_image_embeds.repeat(
-                        num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
-                    )
-                    single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
-                else:
-                    single_image_embeds = single_image_embeds.repeat(
-                        num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
-                    )
+                    negative_image_embeds.append(single_negative_image_embeds)
                 image_embeds.append(single_image_embeds)
 
-        return image_embeds
+        ip_adapter_image_embeds = []
+        for i, single_image_embeds in enumerate(image_embeds):
+            single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0)
+            if do_classifier_free_guidance:
+                single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0)
+                single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0)
+
+            single_image_embeds = single_image_embeds.to(device=device)
+            ip_adapter_image_embeds.append(single_image_embeds)
+
+        return ip_adapter_image_embeds
 
     def run_safety_checker(self, image, device, dtype):
         if self.safety_checker is None:

diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_depth2img.py

@@ -20,11 +20,11 @@ import numpy as np
 import PIL.Image
 import torch
 from packaging import version
-from transformers import CLIPTextModel, CLIPTokenizer, DPTFeatureExtractor, DPTForDepthEstimation
+from transformers import CLIPTextModel, CLIPTokenizer, DPTForDepthEstimation, DPTImageProcessor
 
 from ...configuration_utils import FrozenDict
 from ...image_processor import PipelineImageInput, VaeImageProcessor
-from ...loaders import LoraLoaderMixin, TextualInversionLoaderMixin
+from ...loaders import StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin
 from ...models import AutoencoderKL, UNet2DConditionModel
 from ...models.lora import adjust_lora_scale_text_encoder
 from ...schedulers import KarrasDiffusionSchedulers
@@ -74,7 +74,7 @@ def preprocess(image):
     return image
 
 
-class StableDiffusionDepth2ImgPipeline(DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin):
+class StableDiffusionDepth2ImgPipeline(DiffusionPipeline, TextualInversionLoaderMixin, StableDiffusionLoraLoaderMixin):
     r"""
     Pipeline for text-guided depth-based image-to-image generation using Stable Diffusion.
 
@@ -83,8 +83,8 @@ class StableDiffusionDepth2ImgPipeline(DiffusionPipeline, TextualInversionLoader
 
     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.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+        - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights
 
     Args:
         vae ([`AutoencoderKL`]):
@@ -111,7 +111,7 @@ class StableDiffusionDepth2ImgPipeline(DiffusionPipeline, TextualInversionLoader
         unet: UNet2DConditionModel,
         scheduler: KarrasDiffusionSchedulers,
         depth_estimator: DPTForDepthEstimation,
-        feature_extractor: DPTFeatureExtractor,
+        feature_extractor: DPTImageProcessor,
     ):
         super().__init__()
 
@@ -225,7 +225,7 @@ class StableDiffusionDepth2ImgPipeline(DiffusionPipeline, TextualInversionLoader
         """
         # set lora scale so that monkey patched LoRA
         # function of text encoder can correctly access it
-        if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+        if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin):
             self._lora_scale = lora_scale
 
             # dynamically adjust the LoRA scale
@@ -358,7 +358,7 @@ class StableDiffusionDepth2ImgPipeline(DiffusionPipeline, TextualInversionLoader
             negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
 
         if self.text_encoder is not None:
-            if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+            if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND:
                 # Retrieve the original scale by scaling back the LoRA layers
                 unscale_lora_layers(self.text_encoder, lora_scale)
 
@@ -494,6 +494,13 @@ class StableDiffusionDepth2ImgPipeline(DiffusionPipeline, TextualInversionLoader
                 )
 
             elif isinstance(generator, list):
+                if image.shape[0] < batch_size and batch_size % image.shape[0] == 0:
+                    image = torch.cat([image] * (batch_size // image.shape[0]), dim=0)
+                elif image.shape[0] < batch_size and batch_size % image.shape[0] != 0:
+                    raise ValueError(
+                        f"Cannot duplicate `image` of batch size {image.shape[0]} to effective batch_size {batch_size} "
+                    )
+
                 init_latents = [
                     retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i])
                     for i in range(batch_size)

diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py

@@ -24,7 +24,7 @@ from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPV
 from ...callbacks import MultiPipelineCallbacks, PipelineCallback
 from ...configuration_utils import FrozenDict
 from ...image_processor import PipelineImageInput, VaeImageProcessor
-from ...loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ...loaders import FromSingleFileMixin, IPAdapterMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin
 from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
 from ...models.lora import adjust_lora_scale_text_encoder
 from ...schedulers import KarrasDiffusionSchedulers
@@ -175,7 +175,7 @@ class StableDiffusionImg2ImgPipeline(
     StableDiffusionMixin,
     TextualInversionLoaderMixin,
     IPAdapterMixin,
-    LoraLoaderMixin,
+    StableDiffusionLoraLoaderMixin,
     FromSingleFileMixin,
 ):
     r"""
@@ -186,8 +186,8 @@ class StableDiffusionImg2ImgPipeline(
 
     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.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+        - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights
         - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
         - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
 
@@ -385,7 +385,7 @@ class StableDiffusionImg2ImgPipeline(
         """
         # set lora scale so that monkey patched LoRA
         # function of text encoder can correctly access it
-        if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+        if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin):
             self._lora_scale = lora_scale
 
             # dynamically adjust the LoRA scale
@@ -518,7 +518,7 @@ class StableDiffusionImg2ImgPipeline(
             negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
 
         if self.text_encoder is not None:
-            if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+            if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND:
                 # Retrieve the original scale by scaling back the LoRA layers
                 unscale_lora_layers(self.text_encoder, lora_scale)
 
@@ -553,6 +553,9 @@ class StableDiffusionImg2ImgPipeline(
     def prepare_ip_adapter_image_embeds(
         self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
     ):
+        image_embeds = []
+        if do_classifier_free_guidance:
+            negative_image_embeds = []
         if ip_adapter_image_embeds is None:
             if not isinstance(ip_adapter_image, list):
                 ip_adapter_image = [ip_adapter_image]
@@ -562,7 +565,6 @@ class StableDiffusionImg2ImgPipeline(
                     f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
                 )
 
-            image_embeds = []
             for single_ip_adapter_image, image_proj_layer in zip(
                 ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
             ):
@@ -570,36 +572,28 @@ class StableDiffusionImg2ImgPipeline(
                 single_image_embeds, single_negative_image_embeds = self.encode_image(
                     single_ip_adapter_image, device, 1, output_hidden_state
                 )
-                single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
-                single_negative_image_embeds = torch.stack(
-                    [single_negative_image_embeds] * num_images_per_prompt, dim=0
-                )
 
+                image_embeds.append(single_image_embeds[None, :])
                 if do_classifier_free_guidance:
-                    single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
-                    single_image_embeds = single_image_embeds.to(device)
-
-                image_embeds.append(single_image_embeds)
+                    negative_image_embeds.append(single_negative_image_embeds[None, :])
         else:
-            repeat_dims = [1]
-            image_embeds = []
             for single_image_embeds in ip_adapter_image_embeds:
                 if do_classifier_free_guidance:
                     single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
-                    single_image_embeds = single_image_embeds.repeat(
-                        num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
-                    )
-                    single_negative_image_embeds = single_negative_image_embeds.repeat(
-                        num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
-                    )
-                    single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
-                else:
-                    single_image_embeds = single_image_embeds.repeat(
-                        num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
-                    )
+                    negative_image_embeds.append(single_negative_image_embeds)
                 image_embeds.append(single_image_embeds)
 
-        return image_embeds
+        ip_adapter_image_embeds = []
+        for i, single_image_embeds in enumerate(image_embeds):
+            single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0)
+            if do_classifier_free_guidance:
+                single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0)
+                single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0)
+
+            single_image_embeds = single_image_embeds.to(device=device)
+            ip_adapter_image_embeds.append(single_image_embeds)
+
+        return ip_adapter_image_embeds
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
     def run_safety_checker(self, image, device, dtype):
@@ -746,6 +740,13 @@ class StableDiffusionImg2ImgPipeline(
                 )
 
             elif isinstance(generator, list):
+                if image.shape[0] < batch_size and batch_size % image.shape[0] == 0:
+                    image = torch.cat([image] * (batch_size // image.shape[0]), dim=0)
+                elif image.shape[0] < batch_size and batch_size % image.shape[0] != 0:
+                    raise ValueError(
+                        f"Cannot duplicate `image` of batch size {image.shape[0]} to effective batch_size {batch_size} "
+                    )
+
                 init_latents = [
                     retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i])
                     for i in range(batch_size)

diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py

@@ -15,7 +15,6 @@
 import inspect
 from typing import Any, Callable, Dict, List, Optional, Union
 
-import numpy as np
 import PIL.Image
 import torch
 from packaging import version
@@ -24,7 +23,7 @@ from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPV
 from ...callbacks import MultiPipelineCallbacks, PipelineCallback
 from ...configuration_utils import FrozenDict
 from ...image_processor import PipelineImageInput, VaeImageProcessor
-from ...loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ...loaders import FromSingleFileMixin, IPAdapterMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin
 from ...models import AsymmetricAutoencoderKL, AutoencoderKL, ImageProjection, UNet2DConditionModel
 from ...models.lora import adjust_lora_scale_text_encoder
 from ...schedulers import KarrasDiffusionSchedulers
@@ -38,128 +37,6 @@ from .safety_checker import StableDiffusionSafetyChecker
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
-def prepare_mask_and_masked_image(image, mask, height, width, return_image: bool = False):
-    """
-    Prepares a pair (image, mask) to be consumed by the Stable Diffusion pipeline. This means that those inputs will be
-    converted to ``torch.Tensor`` with shapes ``batch x channels x height x width`` where ``channels`` is ``3`` for the
-    ``image`` and ``1`` for the ``mask``.
-
-    The ``image`` will be converted to ``torch.float32`` and normalized to be in ``[-1, 1]``. The ``mask`` will be
-    binarized (``mask > 0.5``) and cast to ``torch.float32`` too.
-
-    Args:
-        image (Union[np.array, PIL.Image, torch.Tensor]): The image to inpaint.
-            It can be a ``PIL.Image``, or a ``height x width x 3`` ``np.array`` or a ``channels x height x width``
-            ``torch.Tensor`` or a ``batch x channels x height x width`` ``torch.Tensor``.
-        mask (_type_): The mask to apply to the image, i.e. regions to inpaint.
-            It can be a ``PIL.Image``, or a ``height x width`` ``np.array`` or a ``1 x height x width``
-            ``torch.Tensor`` or a ``batch x 1 x height x width`` ``torch.Tensor``.
-
-
-    Raises:
-        ValueError: ``torch.Tensor`` images should be in the ``[-1, 1]`` range. ValueError: ``torch.Tensor`` mask
-        should be in the ``[0, 1]`` range. ValueError: ``mask`` and ``image`` should have the same spatial dimensions.
-        TypeError: ``mask`` is a ``torch.Tensor`` but ``image`` is not
-            (ot the other way around).
-
-    Returns:
-        tuple[torch.Tensor]: The pair (mask, masked_image) as ``torch.Tensor`` with 4
-            dimensions: ``batch x channels x height x width``.
-    """
-    deprecation_message = "The prepare_mask_and_masked_image method is deprecated and will be removed in a future version. Please use VaeImageProcessor.preprocess instead"
-    deprecate(
-        "prepare_mask_and_masked_image",
-        "0.30.0",
-        deprecation_message,
-    )
-    if image is None:
-        raise ValueError("`image` input cannot be undefined.")
-
-    if mask is None:
-        raise ValueError("`mask_image` input cannot be undefined.")
-
-    if isinstance(image, torch.Tensor):
-        if not isinstance(mask, torch.Tensor):
-            raise TypeError(f"`image` is a torch.Tensor but `mask` (type: {type(mask)} is not")
-
-        # Batch single image
-        if image.ndim == 3:
-            assert image.shape[0] == 3, "Image outside a batch should be of shape (3, H, W)"
-            image = image.unsqueeze(0)
-
-        # Batch and add channel dim for single mask
-        if mask.ndim == 2:
-            mask = mask.unsqueeze(0).unsqueeze(0)
-
-        # Batch single mask or add channel dim
-        if mask.ndim == 3:
-            # Single batched mask, no channel dim or single mask not batched but channel dim
-            if mask.shape[0] == 1:
-                mask = mask.unsqueeze(0)
-
-            # Batched masks no channel dim
-            else:
-                mask = mask.unsqueeze(1)
-
-        assert image.ndim == 4 and mask.ndim == 4, "Image and Mask must have 4 dimensions"
-        assert image.shape[-2:] == mask.shape[-2:], "Image and Mask must have the same spatial dimensions"
-        assert image.shape[0] == mask.shape[0], "Image and Mask must have the same batch size"
-
-        # Check image is in [-1, 1]
-        if image.min() < -1 or image.max() > 1:
-            raise ValueError("Image should be in [-1, 1] range")
-
-        # Check mask is in [0, 1]
-        if mask.min() < 0 or mask.max() > 1:
-            raise ValueError("Mask should be in [0, 1] range")
-
-        # Binarize mask
-        mask[mask < 0.5] = 0
-        mask[mask >= 0.5] = 1
-
-        # Image as float32
-        image = image.to(dtype=torch.float32)
-    elif isinstance(mask, torch.Tensor):
-        raise TypeError(f"`mask` is a torch.Tensor but `image` (type: {type(image)} is not")
-    else:
-        # preprocess image
-        if isinstance(image, (PIL.Image.Image, np.ndarray)):
-            image = [image]
-        if isinstance(image, list) and isinstance(image[0], PIL.Image.Image):
-            # resize all images w.r.t passed height an width
-            image = [i.resize((width, height), resample=PIL.Image.LANCZOS) for i in image]
-            image = [np.array(i.convert("RGB"))[None, :] for i in image]
-            image = np.concatenate(image, axis=0)
-        elif isinstance(image, list) and isinstance(image[0], np.ndarray):
-            image = np.concatenate([i[None, :] for i in image], axis=0)
-
-        image = image.transpose(0, 3, 1, 2)
-        image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0
-
-        # preprocess mask
-        if isinstance(mask, (PIL.Image.Image, np.ndarray)):
-            mask = [mask]
-
-        if isinstance(mask, list) and isinstance(mask[0], PIL.Image.Image):
-            mask = [i.resize((width, height), resample=PIL.Image.LANCZOS) for i in mask]
-            mask = np.concatenate([np.array(m.convert("L"))[None, None, :] for m in mask], axis=0)
-            mask = mask.astype(np.float32) / 255.0
-        elif isinstance(mask, list) and isinstance(mask[0], np.ndarray):
-            mask = np.concatenate([m[None, None, :] for m in mask], axis=0)
-
-        mask[mask < 0.5] = 0
-        mask[mask >= 0.5] = 1
-        mask = torch.from_numpy(mask)
-
-    masked_image = image * (mask < 0.5)
-
-    # n.b. ensure backwards compatibility as old function does not return image
-    if return_image:
-        return mask, masked_image, image
-
-    return mask, masked_image
-
-
 # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
 def retrieve_latents(
     encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
@@ -239,7 +116,7 @@ class StableDiffusionInpaintPipeline(
     StableDiffusionMixin,
     TextualInversionLoaderMixin,
     IPAdapterMixin,
-    LoraLoaderMixin,
+    StableDiffusionLoraLoaderMixin,
     FromSingleFileMixin,
 ):
     r"""
@@ -250,8 +127,8 @@ class StableDiffusionInpaintPipeline(
 
     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.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+        - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights
         - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
         - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
 
@@ -457,7 +334,7 @@ class StableDiffusionInpaintPipeline(
         """
         # set lora scale so that monkey patched LoRA
         # function of text encoder can correctly access it
-        if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+        if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin):
             self._lora_scale = lora_scale
 
             # dynamically adjust the LoRA scale
@@ -590,7 +467,7 @@ class StableDiffusionInpaintPipeline(
             negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
 
         if self.text_encoder is not None:
-            if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+            if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND:
                 # Retrieve the original scale by scaling back the LoRA layers
                 unscale_lora_layers(self.text_encoder, lora_scale)
 
@@ -625,6 +502,9 @@ class StableDiffusionInpaintPipeline(
     def prepare_ip_adapter_image_embeds(
         self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
     ):
+        image_embeds = []
+        if do_classifier_free_guidance:
+            negative_image_embeds = []
         if ip_adapter_image_embeds is None:
             if not isinstance(ip_adapter_image, list):
                 ip_adapter_image = [ip_adapter_image]
@@ -634,7 +514,6 @@ class StableDiffusionInpaintPipeline(
                     f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
                 )
 
-            image_embeds = []
             for single_ip_adapter_image, image_proj_layer in zip(
                 ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
             ):
@@ -642,36 +521,28 @@ class StableDiffusionInpaintPipeline(
                 single_image_embeds, single_negative_image_embeds = self.encode_image(
                     single_ip_adapter_image, device, 1, output_hidden_state
                 )
-                single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
-                single_negative_image_embeds = torch.stack(
-                    [single_negative_image_embeds] * num_images_per_prompt, dim=0
-                )
 
+                image_embeds.append(single_image_embeds[None, :])
                 if do_classifier_free_guidance:
-                    single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
-                    single_image_embeds = single_image_embeds.to(device)
-
-                image_embeds.append(single_image_embeds)
+                    negative_image_embeds.append(single_negative_image_embeds[None, :])
         else:
-            repeat_dims = [1]
-            image_embeds = []
             for single_image_embeds in ip_adapter_image_embeds:
                 if do_classifier_free_guidance:
                     single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
-                    single_image_embeds = single_image_embeds.repeat(
-                        num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
-                    )
-                    single_negative_image_embeds = single_negative_image_embeds.repeat(
-                        num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
-                    )
-                    single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
-                else:
-                    single_image_embeds = single_image_embeds.repeat(
-                        num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
-                    )
+                    negative_image_embeds.append(single_negative_image_embeds)
                 image_embeds.append(single_image_embeds)
 
-        return image_embeds
+        ip_adapter_image_embeds = []
+        for i, single_image_embeds in enumerate(image_embeds):
+            single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0)
+            if do_classifier_free_guidance:
+                single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0)
+                single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0)
+
+            single_image_embeds = single_image_embeds.to(device=device)
+            ip_adapter_image_embeds.append(single_image_embeds)
+
+        return ip_adapter_image_embeds
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
     def run_safety_checker(self, image, device, dtype):

diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_instruct_pix2pix.py

@@ -22,7 +22,7 @@ from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPV
 
 from ...callbacks import MultiPipelineCallbacks, PipelineCallback
 from ...image_processor import PipelineImageInput, VaeImageProcessor
-from ...loaders import IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ...loaders import IPAdapterMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin
 from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
 from ...schedulers import KarrasDiffusionSchedulers
 from ...utils import PIL_INTERPOLATION, deprecate, logging
@@ -74,7 +74,11 @@ def retrieve_latents(
 
 
 class StableDiffusionInstructPix2PixPipeline(
-    DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, LoraLoaderMixin, IPAdapterMixin
+    DiffusionPipeline,
+    StableDiffusionMixin,
+    TextualInversionLoaderMixin,
+    StableDiffusionLoraLoaderMixin,
+    IPAdapterMixin,
 ):
     r"""
     Pipeline for pixel-level image editing by following text instructions (based on Stable Diffusion).
@@ -84,8 +88,8 @@ class StableDiffusionInstructPix2PixPipeline(
 
     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.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+        - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights
         - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
 
     Args: