update repo

module/aggregator.py

@@ -1,16 +1,3 @@
-# Copyright 2024 The HuggingFace Team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
 from dataclasses import dataclass
 from typing import Any, Dict, List, Optional, Tuple, Union
 
@@ -19,7 +6,7 @@ from torch import nn
 from torch.nn import functional as F
 
 from diffusers.configuration_utils import ConfigMixin, register_to_config
-from diffusers.loaders import FromOriginalControlNetMixin
+from diffusers.loaders.single_file_model import FromOriginalModelMixin
 from diffusers.utils import BaseOutput, logging
 from diffusers.models.attention_processor import (
     ADDED_KV_ATTENTION_PROCESSORS,
@@ -168,7 +155,7 @@ class ConditioningEmbedding(nn.Module):
         return embedding
 
 
-class Aggregator(ModelMixin, ConfigMixin, FromOriginalControlNetMixin):
+class Aggregator(ModelMixin, ConfigMixin, FromOriginalModelMixin):
     """
     Aggregator model.
 
@@ -781,7 +768,6 @@ class Aggregator(ModelMixin, ConfigMixin, FromOriginalControlNetMixin):
         attention_mask: Optional[torch.Tensor] = None,
         added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
         cross_attention_kwargs: Optional[Dict[str, Any]] = None,
-        guess_mode: bool = False,
         return_dict: bool = True,
     ) -> Union[AggregatorOutput, Tuple[Tuple[torch.FloatTensor, ...], torch.FloatTensor]]:
         """
@@ -812,9 +798,6 @@ class Aggregator(ModelMixin, ConfigMixin, FromOriginalControlNetMixin):
                 Additional conditions for the Stable Diffusion XL UNet.
             cross_attention_kwargs (`dict[str]`, *optional*, defaults to `None`):
                 A kwargs dictionary that if specified is passed along to the `AttnProcessor`.
-            guess_mode (`bool`, defaults to `False`):
-                In this mode, the ControlNet encoder tries its best to recognize the input content of the input even if
-                you remove all prompts. A `guidance_scale` between 3.0 and 5.0 is recommended.
             return_dict (`bool`, defaults to `True`):
                 Whether or not to return a [`~models.controlnet.ControlNetOutput`] instead of a plain tuple.
 
@@ -977,14 +960,8 @@ class Aggregator(ModelMixin, ConfigMixin, FromOriginalControlNetMixin):
         mid_block_res_sample = self.controlnet_mid_block((cond_latent, ref_latent), )
 
         # 6. scaling
-        if guess_mode and not self.config.global_pool_conditions:
-            scales = torch.logspace(-1, 0, len(down_block_res_samples) + 1, device=sample.device)  # 0.1 to 1.0
-            scales = scales * conditioning_scale
-            down_block_res_samples = [sample*scale for sample, scale in zip(down_block_res_samples, scales)]
-            mid_block_res_sample = mid_block_res_sample*scales[-1]  # last scale
-        else:
-            down_block_res_samples = [sample*conditioning_scale for sample in down_block_res_samples]
-            mid_block_res_sample = mid_block_res_sample*conditioning_scale
+        down_block_res_samples = [sample*conditioning_scale for sample in down_block_res_samples]
+        mid_block_res_sample = mid_block_res_sample*conditioning_scale
 
         if self.config.global_pool_conditions:
             down_block_res_samples = [

module/ip_adapter/attention_processor.py

@@ -1361,7 +1361,7 @@ class CNAttnProcessor2_0:
         return hidden_states
 
 
-def init_attn_proc(unet, ip_adapter_tokens=16, use_lcm=True, use_adaln=True, use_external_kv=False):
+def init_attn_proc(unet, ip_adapter_tokens=16, use_lcm=False, use_adaln=True, use_external_kv=False):
     attn_procs = {}
     unet_sd = unet.state_dict()
     for name in unet.attn_processors.keys():

module/ip_adapter/resampler.py

@@ -81,11 +81,11 @@ class PerceiverAttention(nn.Module):
 class Resampler(nn.Module):
     def __init__(
         self,
-        dim=1024,
-        depth=8,
+        dim=1280,
+        depth=4,
         dim_head=64,
-        heads=16,
-        num_queries=8,
+        heads=20,
+        num_queries=64,
         embedding_dim=768,
         output_dim=1024,
         ff_mult=4,

module/ip_adapter/utils.py

@@ -1,23 +1,32 @@
-import random
 import torch
 from collections import namedtuple, OrderedDict
 from safetensors import safe_open
 from .attention_processor import init_attn_proc
 from .ip_adapter import MultiIPAdapterImageProjection
+from .resampler import Resampler
 from transformers import (
     AutoModel, AutoImageProcessor,
     CLIPVisionModelWithProjection, CLIPImageProcessor)
 
 
-def init_ip_adapter_in_unet(
+def init_adapter_in_unet(
         unet,
-        image_proj_model,
+        image_proj_model=None,
         pretrained_model_path_or_dict=None,
-        adapter_tokens=16,
+        adapter_tokens=64,
+        embedding_dim=None,
         use_lcm=False,
         use_adaln=True,
-        use_external_kv=False,
     ):
+        device = unet.device
+        dtype = unet.dtype
+        if image_proj_model is None:
+            assert embedding_dim is not None, "embedding_dim must be provided if image_proj_model is None."
+            image_proj_model = Resampler(
+                embedding_dim=embedding_dim,
+                output_dim=unet.config.cross_attention_dim,
+                num_queries=adapter_tokens,
+            )
         if pretrained_model_path_or_dict is not None:
             if not isinstance(pretrained_model_path_or_dict, dict):
                 if pretrained_model_path_or_dict.endswith(".safetensors"):
@@ -37,7 +46,7 @@ def init_ip_adapter_in_unet(
                 state_dict = revise_state_dict(state_dict)
 
         # Creat IP cross-attention in unet.
-        attn_procs = init_attn_proc(unet, adapter_tokens, use_lcm, use_adaln, use_external_kv)
+        attn_procs = init_attn_proc(unet, adapter_tokens, use_lcm, use_adaln)
         unet.set_attn_processor(attn_procs)
 
         # Load pretrinaed model if needed.
@@ -58,24 +67,24 @@ def init_ip_adapter_in_unet(
 
         # Adjust unet config to handle addtional ip hidden states.
         unet.config.encoder_hid_dim_type = "ip_image_proj"
+        unet.to(dtype=dtype, device=device)
 
 
-def load_ip_adapter_to_pipe(
+def load_adapter_to_pipe(
         pipe,
         pretrained_model_path_or_dict,
-        image_encoder_path=None,
-        feature_extractor_path=None,
-        use_dino=False,
-        ip_adapter_tokens=16,
-        use_lcm=True,
+        image_encoder_or_path=None,
+        feature_extractor_or_path=None,
+        use_clip_encoder=False,
+        adapter_tokens=64,
+        use_lcm=False,
         use_adaln=True,
-        low_cpu_mem_usage=True,
     ):
 
         if not isinstance(pretrained_model_path_or_dict, dict):
             if pretrained_model_path_or_dict.endswith(".safetensors"):
                 state_dict = {"image_proj": {}, "ip_adapter": {}}
-                with safe_open(pretrained_model_path_or_dict, framework="pt", device="cpu") as f:
+                with safe_open(pretrained_model_path_or_dict, framework="pt", device=pipe.device) as f:
                     for key in f.keys():
                         if key.startswith("image_proj."):
                             state_dict["image_proj"][key.replace("image_proj.", "")] = f.get_tensor(key)
@@ -85,150 +94,72 @@ def load_ip_adapter_to_pipe(
                 state_dict = torch.load(pretrained_model_path_or_dict, map_location=pipe.device)
         else:
             state_dict = pretrained_model_path_or_dict
-
         keys = list(state_dict.keys())
-        if keys != ["image_proj", "ip_adapter"]:
+        if "image_proj" not in keys and "ip_adapter" not in keys:
             state_dict = revise_state_dict(state_dict)
 
         # load CLIP image encoder here if it has not been registered to the pipeline yet
-        if image_encoder_path is not None:
-            if isinstance(image_encoder_path, str):
-                feature_extractor_path = image_encoder_path if feature_extractor_path is None else feature_extractor_path
-
-                image_encoder_path = AutoModel.from_pretrained(
-                    image_encoder_path) if use_dino else \
-                        CLIPVisionModelWithProjection.from_pretrained(
-                            image_encoder_path)
-            image_encoder = image_encoder_path.to(pipe.device, dtype=pipe.dtype)
+        if image_encoder_or_path is not None:
+            if isinstance(image_encoder_or_path, str):
+                feature_extractor_or_path = image_encoder_or_path if feature_extractor_or_path is None else feature_extractor_or_path
+
+                image_encoder_or_path = (
+                    CLIPVisionModelWithProjection.from_pretrained(
+                        image_encoder_or_path
+                    ) if use_clip_encoder else
+                    AutoModel.from_pretrained(image_encoder_or_path)
+                )
 
-        if feature_extractor_path is not None:
-            if isinstance(feature_extractor_path, str):
-                feature_extractor_path = AutoImageProcessor.from_pretrained(feature_extractor_path) \
-                    if use_dino else CLIPImageProcessor()
-            feature_extractor = feature_extractor_path
+        if feature_extractor_or_path is not None:
+            if isinstance(feature_extractor_or_path, str):
+                feature_extractor_or_path = (
+                    CLIPImageProcessor() if use_clip_encoder else
+                    AutoImageProcessor.from_pretrained(feature_extractor_or_path)
+                )
 
         # create image encoder if it has not been registered to the pipeline yet
         if hasattr(pipe, "image_encoder") and getattr(pipe, "image_encoder", None) is None:
+            image_encoder = image_encoder_or_path.to(pipe.device, dtype=pipe.dtype)
             pipe.register_modules(image_encoder=image_encoder)
-
-        # create feature extractor if it has not been registered to the pipeline yet
-        if hasattr(pipe, "feature_extractor") and getattr(pipe, "feature_extractor", None) is None:
-            pipe.register_modules(feature_extractor=feature_extractor)
-
-        # load ip-adapter into unet
-        unet = getattr(pipe, pipe.unet_name) if not hasattr(pipe, "unet") else pipe.unet
-        attn_procs = init_attn_proc(unet, ip_adapter_tokens, use_lcm, use_adaln)
-        unet.set_attn_processor(attn_procs)
-        adapter_modules = torch.nn.ModuleList(unet.attn_processors.values())
-        missing, _ = adapter_modules.load_state_dict(state_dict["ip_adapter"], strict=False)
-        if len(missing) > 0:
-            raise ValueError(f"Missing keys in adapter_modules: {missing}")
-
-        # convert IP-Adapter Image Projection layers to diffusers
-        image_projection_layers = []
-        image_projection_layer = unet._convert_ip_adapter_image_proj_to_diffusers(
-            state_dict["image_proj"], low_cpu_mem_usage=low_cpu_mem_usage
-        )
-        image_projection_layers.append(image_projection_layer)
-
-        unet.encoder_hid_proj = MultiIPAdapterImageProjection(image_projection_layers)
-        unet.config.encoder_hid_dim_type = "ip_image_proj"
-
-        unet.to(dtype=pipe.dtype, device=pipe.device)
-
-
-def load_ip_adapter_to_controlnet_pipe(
-        pipe,
-        pretrained_model_path_or_dict,
-        image_encoder_path=None,
-        feature_extractor_path=None,
-        use_dino=False,
-        ip_adapter_tokens=16,
-        use_lcm=True,
-        use_adaln=True,
-        low_cpu_mem_usage=True,
-    ):
-
-        if not isinstance(pretrained_model_path_or_dict, dict):
-            if pretrained_model_path_or_dict.endswith(".safetensors"):
-                state_dict = {"image_proj": {}, "ip_adapter": {}}
-                with safe_open(pretrained_model_path_or_dict, framework="pt", device="cpu") as f:
-                    for key in f.keys():
-                        if key.startswith("image_proj."):
-                            state_dict["image_proj"][key.replace("image_proj.", "")] = f.get_tensor(key)
-                        elif key.startswith("ip_adapter."):
-                            state_dict["ip_adapter"][key.replace("ip_adapter.", "")] = f.get_tensor(key)
-            else:
-                state_dict = torch.load(pretrained_model_path_or_dict, map_location=pipe.device)
         else:
-            state_dict = pretrained_model_path_or_dict
-
-        keys = list(state_dict.keys())
-        if keys != ["image_proj", "ip_adapter"]:
-            state_dict = revise_state_dict(state_dict)
-
-        # load CLIP image encoder here if it has not been registered to the pipeline yet
-        if image_encoder_path is not None:
-            if isinstance(image_encoder_path, str):
-                feature_extractor_path = image_encoder_path if feature_extractor_path is None else feature_extractor_path
-
-                image_encoder_path = AutoModel.from_pretrained(
-                    image_encoder_path) if use_dino else \
-                        CLIPVisionModelWithProjection.from_pretrained(
-                            image_encoder_path)
-            image_encoder = image_encoder_path.to(pipe.device, dtype=pipe.dtype)
-
-        if feature_extractor_path is not None:
-            if isinstance(feature_extractor_path, str):
-                feature_extractor_path = AutoImageProcessor.from_pretrained(feature_extractor_path) \
-                    if use_dino else CLIPImageProcessor()
-            feature_extractor = feature_extractor_path
-
-        # create image encoder if it has not been registered to the pipeline yet
-        if hasattr(pipe, "image_encoder") and getattr(pipe, "image_encoder", None) is None:
-            pipe.register_modules(image_encoder=image_encoder)
+            image_encoder = pipe.image_encoder
 
         # create feature extractor if it has not been registered to the pipeline yet
         if hasattr(pipe, "feature_extractor") and getattr(pipe, "feature_extractor", None) is None:
+            feature_extractor = feature_extractor_or_path
             pipe.register_modules(feature_extractor=feature_extractor)
+        else:
+            feature_extractor = pipe.feature_extractor
 
-        # load ip-adapter into unet
+        # load adapter into unet
         unet = getattr(pipe, pipe.unet_name) if not hasattr(pipe, "unet") else pipe.unet
-        attn_procs = init_attn_proc(unet, ip_adapter_tokens, use_lcm, use_adaln)
+        attn_procs = init_attn_proc(unet, adapter_tokens, use_lcm, use_adaln)
         unet.set_attn_processor(attn_procs)
-        adapter_modules = torch.nn.ModuleList(unet.attn_processors.values())
-        missing, _ = adapter_modules.load_state_dict(state_dict["ip_adapter"], strict=False)
-        if len(missing) > 0:
-            raise ValueError(f"Missing keys in adapter_modules: {missing}")
+        image_proj_model = Resampler(
+            embedding_dim=image_encoder.config.hidden_size,
+            output_dim=unet.config.cross_attention_dim,
+            num_queries=adapter_tokens,
+        )
 
-        controlnet = getattr(pipe, pipe.controlnet_name) if not hasattr(pipe, "controlnet") else pipe.controlnet
-        controlnet_attn_procs = init_attn_proc(controlnet, ip_adapter_tokens, use_lcm, use_adaln)
-        controlnet.set_attn_processor(controlnet_attn_procs)
-        controlnet_adapter_modules = torch.nn.ModuleList(controlnet.attn_processors.values())
-        missing, unexpected = controlnet_adapter_modules.load_state_dict(state_dict["ip_adapter"], strict=False)
-        if len(missing) > 0:
-            raise ValueError(f"Missing keys in adapter_modules: {missing}")
-        if len(unexpected) > 0:
-            for mk in unexpected:
-                layer_id = int(mk.split(".")[0])
-                if layer_id < len(controlnet.attn_processors.keys()):
-                    raise ValueError(f"Failed to load {unexpected} in controlnet adapter_modules")
+        # Load pretrinaed model if needed.
+        if "ip_adapter" in state_dict.keys():
+            adapter_modules = torch.nn.ModuleList(unet.attn_processors.values())
+            missing, unexpected = adapter_modules.load_state_dict(state_dict["ip_adapter"], strict=False)
+            for mk in missing:
+                if "ln" not in mk:
+                    raise ValueError(f"Missing keys in adapter_modules: {missing}")
+        if "image_proj" in state_dict.keys():
+            image_proj_model.load_state_dict(state_dict["image_proj"])
 
         # convert IP-Adapter Image Projection layers to diffusers
         image_projection_layers = []
-        image_projection_layer = unet._convert_ip_adapter_image_proj_to_diffusers(
-            state_dict["image_proj"], low_cpu_mem_usage=low_cpu_mem_usage
-        )
-        image_projection_layers.append(image_projection_layer)
-
+        image_projection_layers.append(image_proj_model)
         unet.encoder_hid_proj = MultiIPAdapterImageProjection(image_projection_layers)
-        unet.config.encoder_hid_dim_type = "ip_image_proj"
-
-        controlnet.encoder_hid_proj = MultiIPAdapterImageProjection(image_projection_layers)
-        controlnet.config.encoder_hid_dim_type = "ip_image_proj"
 
+        # Adjust unet config to handle addtional ip hidden states.
+        unet.config.encoder_hid_dim_type = "ip_image_proj"
         unet.to(dtype=pipe.dtype, device=pipe.device)
-        controlnet.to(dtype=pipe.dtype, device=pipe.device)
+
 
 def revise_state_dict(old_state_dict_or_path, map_location="cpu"):
     new_state_dict = OrderedDict()

pipelines/sdxl_instantir.py

@@ -1,4 +1,4 @@
-# Copyright 2024 The HuggingFace Team. All rights reserved.
+# Copyright 2024 The InstantX Team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -53,6 +53,7 @@ from diffusers.utils import (
     replace_example_docstring,
     scale_lora_layers,
     unscale_lora_layers,
+    convert_unet_state_dict_to_peft
 )
 from diffusers.utils.torch_utils import is_compiled_module, is_torch_version, randn_tensor
 from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
@@ -62,6 +63,7 @@ from diffusers.pipelines.stable_diffusion_xl.pipeline_output import StableDiffus
 if is_invisible_watermark_available():
     from diffusers.pipelines.stable_diffusion_xl.watermark import StableDiffusionXLWatermarker
 
+from peft import LoraConfig, set_peft_model_state_dict
 from module.aggregator import Aggregator
 
 
@@ -71,44 +73,52 @@ logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 EXAMPLE_DOC_STRING = """
     Examples:
         ```py
-        >>> # !pip install opencv-python transformers accelerate
-        >>> from diffusers import StableDiffusionXLControlNetPipeline, ControlNetModel, AutoencoderKL
-        >>> from diffusers.utils import load_image
-        >>> import numpy as np
+        >>> # !pip install diffusers pillow transformers accelerate
         >>> import torch
-
-        >>> import cv2
         >>> from PIL import Image
 
-        >>> prompt = "aerial view, a futuristic research complex in a bright foggy jungle, hard lighting"
-        >>> negative_prompt = "low quality, bad quality, sketches"
+        >>> from diffusers import DDPMScheduler
+        >>> from schedulers.lcm_single_step_scheduler import LCMSingleStepScheduler
 
-        >>> # download an image
-        >>> image = load_image(
-        ...     "https://hf.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd_controlnet/hf-logo.png"
-        ... )
+        >>> from module.ip_adapter.utils import load_adapter_to_pipe
+        >>> from pipelines.sdxl_instantir import InstantIRPipeline
 
-        >>> # initialize the models and pipeline
-        >>> controlnet_conditioning_scale = 0.5  # recommended for good generalization
-        >>> controlnet = ControlNetModel.from_pretrained(
-        ...     "diffusers/controlnet-canny-sdxl-1.0", torch_dtype=torch.float16
-        ... )
-        >>> vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16)
-        >>> pipe = StableDiffusionXLControlNetPipeline.from_pretrained(
+        >>> # download models under ./models
+        >>> dcp_adapter = f'./models/adapter.pt'
+        >>> previewer_lora_path = f'./models'
+        >>> instantir_path = f'./models/aggregator.pt'
+
+        >>> # load pretrained models
+        >>> pipe = InstantIRPipeline.from_pretrained(
         ...     "stabilityai/stable-diffusion-xl-base-1.0", controlnet=controlnet, vae=vae, torch_dtype=torch.float16
         ... )
+        >>> # load adapter
+        >>> load_adapter_to_pipe(
+        ...     pipe,
+        ...     dcp_adapter,
+        ...     image_encoder_or_path = 'facebook/dinov2-large',
+        ... )
+        >>> # load previewer lora
+        >>> pipe.prepare_previewers(previewer_lora_path)
+        >>> pipe.scheduler = DDPMScheduler.from_pretrained('stabilityai/stable-diffusion-xl-base-1.0', subfolder="scheduler")
+        >>> lcm_scheduler = LCMSingleStepScheduler.from_config(pipe.scheduler.config)
+
+        >>> # load aggregator weights
+        >>> pretrained_state_dict = torch.load(instantir_path)
+        >>> pipe.aggregator.load_state_dict(pretrained_state_dict)
+
+        >>> # send to GPU and fp16
+        >>> pipe.to(device="cuda", dtype=torch.float16)
+        >>> pipe.aggregator.to(device="cuda", dtype=torch.float16)
         >>> pipe.enable_model_cpu_offload()
 
-        >>> # get canny image
-        >>> image = np.array(image)
-        >>> image = cv2.Canny(image, 100, 200)
-        >>> image = image[:, :, None]
-        >>> image = np.concatenate([image, image, image], axis=2)
-        >>> canny_image = Image.fromarray(image)
+        >>> # load a broken image
+        >>> low_quality_image = Image.open('path/to/your-image').convert("RGB")
 
-        >>> # generate image
+        >>> # restoration
         >>> image = pipe(
-        ...     prompt, controlnet_conditioning_scale=controlnet_conditioning_scale, image=canny_image
+        ...     image=low_quality_image,
+        ...     previewer_scheduler=lcm_scheduler,
         ... ).images[0]
         ```
 """
@@ -299,8 +309,8 @@ class InstantIRPipeline(
         tokenizer: CLIPTokenizer,
         tokenizer_2: CLIPTokenizer,
         unet: UNet2DConditionModel,
-        aggregator: Aggregator,
         scheduler: KarrasDiffusionSchedulers,
+        aggregator: Aggregator = None,
         force_zeros_for_empty_prompt: bool = True,
         add_watermarker: Optional[bool] = None,
         feature_extractor: CLIPImageProcessor = None,
@@ -308,6 +318,8 @@ class InstantIRPipeline(
     ):
         super().__init__()
 
+        if aggregator is None:
+            aggregator = Aggregator.from_unet(unet)
         remove_attn2(aggregator)
 
         self.register_modules(
@@ -336,6 +348,55 @@ class InstantIRPipeline(
 
         self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
 
+    def prepare_previewers(self, previewer_lora_path: str, use_lcm=False):
+        if use_lcm:
+            lora_state_dict, alpha_dict = self.lora_state_dict(
+                previewer_lora_path,
+            )
+        else:
+            lora_state_dict, alpha_dict = self.lora_state_dict(
+                previewer_lora_path,
+                weight_name="previewer_lora_weights.bin"
+            )
+        unet_state_dict = {
+            f'{k.replace("unet.", "")}': v for k, v in lora_state_dict.items() if k.startswith("unet.")
+        }
+        unet_state_dict = convert_unet_state_dict_to_peft(unet_state_dict)
+        lora_state_dict = dict()
+        for k, v in unet_state_dict.items():
+            if "ip" in k:
+                k = k.replace("attn2", "attn2.processor")
+                lora_state_dict[k] = v
+            else:
+                lora_state_dict[k] = v
+        if alpha_dict:
+            lora_alpha = next(iter(alpha_dict.values()))
+        else:
+            lora_alpha = 1
+        logger.info(f"use lora alpha {lora_alpha}")
+        lora_config = LoraConfig(
+            r=64,
+            target_modules=LCM_LORA_MODULES if use_lcm else PREVIEWER_LORA_MODULES,
+            lora_alpha=lora_alpha,
+            lora_dropout=0.0,
+        )
+
+        adapter_name = "lcm" if use_lcm else "previewer"
+        self.unet.add_adapter(lora_config, adapter_name)
+        incompatible_keys = set_peft_model_state_dict(self.unet, lora_state_dict, adapter_name=adapter_name)
+        if incompatible_keys is not None:
+            # check only for unexpected keys
+            unexpected_keys = getattr(incompatible_keys, "unexpected_keys", None)
+            missing_keys = getattr(incompatible_keys, "missing_keys", None)
+            if unexpected_keys:
+                raise ValueError(
+                    f"Loading adapter weights from state_dict led to unexpected keys not found in the model: "
+                    f" {unexpected_keys}. "
+                )
+        self.unet.disable_adapters()
+
+        return lora_alpha
+    
     # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt
     def encode_prompt(
         self,
@@ -1011,10 +1072,10 @@ class InstantIRPipeline(
         image: PipelineImageInput = None,
         height: Optional[int] = None,
         width: Optional[int] = None,
-        num_inference_steps: int = 50,
+        num_inference_steps: int = 30,
         timesteps: List[int] = None,
         denoising_end: Optional[float] = None,
-        guidance_scale: float = 5.0,
+        guidance_scale: float = 7.0,
         negative_prompt: Optional[Union[str, List[str]]] = None,
         negative_prompt_2: Optional[Union[str, List[str]]] = None,
         num_images_per_prompt: Optional[int] = 1,
@@ -1032,11 +1093,14 @@ class InstantIRPipeline(
         save_preview_row: bool = False,
         init_latents_with_lq: bool = True,
         multistep_restore: bool = False,
+        adastep_restore: bool = False,
         cross_attention_kwargs: Optional[Dict[str, Any]] = None,
         guidance_rescale: float = 0.0,
-        controlnet_conditioning_scale: Union[float, List[float]] = 1.0,
-        control_guidance_start: Union[float, List[float]] = 0.0,
-        control_guidance_end: Union[float, List[float]] = 1.0,
+        controlnet_conditioning_scale: float = 1.0,
+        control_guidance_start: float = 0.0,
+        control_guidance_end: float = 1.0,
+        preview_start: float = 0.0,
+        preview_end: float = 1.0,
         original_size: Tuple[int, int] = None,
         crops_coords_top_left: Tuple[int, int] = (0, 0),
         target_size: Tuple[int, int] = None,
@@ -1212,6 +1276,8 @@ class InstantIRPipeline(
             )
 
         aggregator = self.aggregator._orig_mod if is_compiled_module(self.aggregator) else self.aggregator
+        if not isinstance(ip_adapter_image, list):
+            ip_adapter_image = [ip_adapter_image] if ip_adapter_image is not None else [image]
 
         # 1. Check inputs. Raise error if not correct
         self.check_inputs(
@@ -1303,8 +1369,14 @@ class InstantIRPipeline(
         )
         height, width = image.shape[-2:]
         if image.shape[1] != 4:
+            needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+            if needs_upcasting:
+                image = image.float()
+                self.vae.to(dtype=torch.float32)
             image = self.vae.encode(image).latent_dist.sample()
             image = image * self.vae.config.scaling_factor
+            if needs_upcasting:
+                self.vae.to(dtype=torch.float16)
         else:
             height = int(height * self.vae_scale_factor)
             width = int(width * self.vae_scale_factor)
@@ -1341,9 +1413,12 @@ class InstantIRPipeline(
 
         # 7.1 Create tensor stating which controlnets to keep
         controlnet_keep = []
+        previewing = []
         for i in range(len(timesteps)):
             keeps = 1.0 - float(i / len(timesteps) < control_guidance_start or (i + 1) / len(timesteps) > control_guidance_end)
             controlnet_keep.append(keeps)
+            use_preview = 1.0 - float(i / len(timesteps) < preview_start or (i + 1) / len(timesteps) > preview_end)
+            previewing.append(use_preview)
         if isinstance(controlnet_conditioning_scale, list):
             assert len(controlnet_conditioning_scale) == len(timesteps), f"{len(controlnet_conditioning_scale)} controlnet scales do not match number of sampling steps {len(timesteps)}"
         else:
@@ -1427,83 +1502,105 @@ class InstantIRPipeline(
                 # expand the latents if we are doing classifier free guidance
                 latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
                 latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+                prev_t = t
+                unet_model_input = latent_model_input
 
-                added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
-
-                # preview with LCM
-                previewer_model_input = latent_model_input
-                previewer_prompt_embeds = prompt_embeds
-                previewer_added_cond_kwargs = {
+                added_cond_kwargs = {
                     "text_embeds": add_text_embeds,
                     "time_ids": add_time_ids,
                     "image_embeds": image_embeds
                 }
-                self.unet.enable_adapters()
-                preview_noise = self.unet(
-                    previewer_model_input,
-                    t,
-                    encoder_hidden_states=previewer_prompt_embeds,
-                    timestep_cond=timestep_cond,
-                    cross_attention_kwargs=self.cross_attention_kwargs,
-                    added_cond_kwargs=previewer_added_cond_kwargs,
-                    return_dict=False,
-                )[0]
-                preview_latent = previewer_scheduler.step(
-                    preview_noise,
-                    t.to(dtype=torch.int64),
-                    # torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents,
-                    latent_model_input,
-                    return_dict=False
-                )[0]
-                self.unet.disable_adapters()
-                if self.do_classifier_free_guidance:
-                    _, preview_latent_cond = preview_latent.chunk(2)
-                    _, noise_preview = preview_noise.chunk(2)
-                    preview_row.append(preview_latent_cond.to('cpu'))
-                else:
-                    noise_preview = preview_noise
-                    preview_row.append(preview_latent.to('cpu'))
-                # Prepare 2nd order step.
-                if multistep_restore and i+1 < len(timesteps):
-                    first_step = self.scheduler.step(noise_preview, t, latents, **extra_step_kwargs, return_dict=True, step_forward=False)
-                    prev_t = timesteps[i + 1]
-                    unet_model_input = torch.cat([first_step.prev_sample] * 2) if self.do_classifier_free_guidance else first_step.prev_sample
-                    unet_model_input = self.scheduler.scale_model_input(unet_model_input, prev_t, heun_step=True)
-                else:
-                    prev_t = t
-                    unet_model_input = latent_model_input
+                aggregator_added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
 
-                if reference_latents is not None:
-                    preview_latent = torch.cat([reference_latents] * 2) if self.do_classifier_free_guidance else reference_latents
+                # prepare time_embeds in advance as adapter input
+                cross_attention_t_emb = self.unet.get_time_embed(sample=latent_model_input, timestep=t)
+                cross_attention_emb = self.unet.time_embedding(cross_attention_t_emb, timestep_cond)
+                cross_attention_aug_emb = None
 
-                # Add fresh noise
-                # preview_noise = torch.randn_like(preview_latent)
-                # preview_latent = self.scheduler.add_noise(preview_latent, preview_noise, t)
+                cross_attention_aug_emb = self.unet.get_aug_embed(
+                    emb=cross_attention_emb,
+                    encoder_hidden_states=prompt_embeds,
+                    added_cond_kwargs=added_cond_kwargs
+                )
+
+                cross_attention_emb = cross_attention_emb + cross_attention_aug_emb if cross_attention_aug_emb is not None else cross_attention_emb
 
-                preview_latent=preview_latent.to(dtype=next(aggregator.parameters()).dtype)
+                if self.unet.time_embed_act is not None:
+                    cross_attention_emb = self.unet.time_embed_act(cross_attention_emb)
 
-                # Aggregator inference
-                generative_reference = preview_latent
+                current_cross_attention_kwargs = {"temb": cross_attention_emb}
+                if cross_attention_kwargs is not None:
+                    for k,v in cross_attention_kwargs.items():
+                        current_cross_attention_kwargs[k] = v
+                self._cross_attention_kwargs = current_cross_attention_kwargs
 
-                adaRes_scale = preview_factor.to(generative_reference.dtype).clamp(0.0, controlnet_conditioning_scale[i])
+                # adaptive restoration factors
+                adaRes_scale = preview_factor.to(latent_model_input.dtype).clamp(0.0, controlnet_conditioning_scale[i])
                 cond_scale = adaRes_scale * controlnet_keep[i]
                 cond_scale = torch.cat([cond_scale] * 2) if self.do_classifier_free_guidance else cond_scale
-                print(cond_scale.squeeze())
 
-                extra_kwargs = {"log_attn": 10} if i == 0 else None
-                down_block_res_samples, mid_block_res_sample = aggregator(
-                    image,
-                    prev_t,
-                    encoder_hidden_states=prompt_embeds,
-                    controlnet_cond=generative_reference,
-                    conditioning_scale=cond_scale,
-                    # cross_attention_kwargs=extra_kwargs,
-                    added_cond_kwargs=added_cond_kwargs,
-                    return_dict=False,
-                )
+                if (cond_scale>0.1).sum().item() > 0:
+                    if previewing[i] > 0:
+                        # preview with LCM
+                        self.unet.enable_adapters()
+                        preview_noise = self.unet(
+                            latent_model_input,
+                            t,
+                            encoder_hidden_states=prompt_embeds,
+                            timestep_cond=timestep_cond,
+                            cross_attention_kwargs=self.cross_attention_kwargs,
+                            added_cond_kwargs=added_cond_kwargs,
+                            return_dict=False,
+                        )[0]
+                        preview_latent = previewer_scheduler.step(
+                            preview_noise,
+                            t.to(dtype=torch.int64),
+                            # torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents,
+                            latent_model_input,     # scaled latents here for compatibility
+                            return_dict=False
+                        )[0]
+                        self.unet.disable_adapters()
+
+                        if self.do_classifier_free_guidance:
+                            preview_row.append(preview_latent.chunk(2)[1].to('cpu'))
+                        else:
+                            preview_row.append(preview_latent.to('cpu'))
+                        # Prepare 2nd order step.
+                        if multistep_restore and i+1 < len(timesteps):
+                            noise_preview = preview_noise.chunk(2)[1] if self.do_classifier_free_guidance else preview_noise
+                            first_step = self.scheduler.step(
+                                noise_preview, t, latents,
+                                **extra_step_kwargs, return_dict=True, step_forward=False
+                            )
+                            prev_t = timesteps[i + 1]
+                            unet_model_input = torch.cat([first_step.prev_sample] * 2) if self.do_classifier_free_guidance else first_step.prev_sample
+                            unet_model_input = self.scheduler.scale_model_input(unet_model_input, prev_t, heun_step=True)
+
+                    elif reference_latents is not None:
+                        preview_latent = torch.cat([reference_latents] * 2) if self.do_classifier_free_guidance else reference_latents
+                    else:
+                        preview_latent = image
+
+                    # Add fresh noise
+                    # preview_noise = torch.randn_like(preview_latent)
+                    # preview_latent = self.scheduler.add_noise(preview_latent, preview_noise, t)
+
+                    preview_latent=preview_latent.to(dtype=next(aggregator.parameters()).dtype)
+
+                    # Aggregator inference
+                    down_block_res_samples, mid_block_res_sample = aggregator(
+                        image,
+                        prev_t,
+                        encoder_hidden_states=prompt_embeds,
+                        controlnet_cond=preview_latent,
+                        # conditioning_scale=cond_scale,
+                        added_cond_kwargs=aggregator_added_cond_kwargs,
+                        return_dict=False,
+                    )
 
-                if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
-                    added_cond_kwargs["image_embeds"] = image_embeds
+                # aggregator features scaling
+                down_block_res_samples = [sample*cond_scale for sample in down_block_res_samples]
+                mid_block_res_sample = mid_block_res_sample*cond_scale
 
                 # predict the noise residual
                 noise_pred = self.unet(
@@ -1536,14 +1633,15 @@ class InstantIRPipeline(
                 unet_pred_latent = unet_step.pred_original_sample
 
                 # Adaptive restoration.
-                pred_x0_l2 = ((preview_latent[latents.shape[0]:].float()-unet_pred_latent.float())).pow(2).sum(dim=(1,2,3))
-                previewer_l2 = ((preview_latent[latents.shape[0]:].float()-previewer_mean.float())).pow(2).sum(dim=(1,2,3))
-                # unet_l2 = ((unet_pred_latent.float()-unet_mean.float())).pow(2).sum(dim=(1,2,3)).sqrt()
-                # l2_error = (((preview_latent[latents.shape[0]:]-previewer_mean) - (unet_pred_latent-unet_mean))).pow(2).mean(dim=(1,2,3))
-                # preview_error = torch.nn.functional.cosine_similarity(preview_latent[latents.shape[0]:].reshape(latents.shape[0], -1), unet_pred_latent.reshape(latents.shape[0],-1))
-                previewer_mean = preview_latent[latents.shape[0]:]
-                unet_mean = unet_pred_latent
-                preview_factor = (pred_x0_l2 / previewer_l2).reshape(-1, 1, 1, 1)
+                if adastep_restore:
+                    pred_x0_l2 = ((preview_latent[latents.shape[0]:].float()-unet_pred_latent.float())).pow(2).sum(dim=(1,2,3))
+                    previewer_l2 = ((preview_latent[latents.shape[0]:].float()-previewer_mean.float())).pow(2).sum(dim=(1,2,3))
+                    # unet_l2 = ((unet_pred_latent.float()-unet_mean.float())).pow(2).sum(dim=(1,2,3)).sqrt()
+                    # l2_error = (((preview_latent[latents.shape[0]:]-previewer_mean) - (unet_pred_latent-unet_mean))).pow(2).mean(dim=(1,2,3))
+                    # preview_error = torch.nn.functional.cosine_similarity(preview_latent[latents.shape[0]:].reshape(latents.shape[0], -1), unet_pred_latent.reshape(latents.shape[0],-1))
+                    previewer_mean = preview_latent[latents.shape[0]:]
+                    unet_mean = unet_pred_latent
+                    preview_factor = (pred_x0_l2 / previewer_l2).reshape(-1, 1, 1, 1)
 
                 if latents.dtype != latents_dtype:
                     if torch.backends.mps.is_available():
@@ -1610,7 +1708,7 @@ class InstantIRPipeline(
             if needs_upcasting:
                 self.upcast_vae()
             for preview_latents in preview_row:
-                preview_latents = preview_latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
+                preview_latents = preview_latents.to(device=self.device, dtype=next(iter(self.vae.post_quant_conv.parameters())).dtype)
                 if has_latents_mean and has_latents_std:
                     latents_mean = (
                         torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(preview_latents.device, preview_latents.dtype)

pipelines/stage1_sdxl_pipeline.py

@@ -0,0 +1,1283 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import torch
+from transformers import (
+    CLIPImageProcessor,
+    CLIPTextModel,
+    CLIPTextModelWithProjection,
+    CLIPTokenizer,
+    CLIPVisionModelWithProjection,
+)
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import (
+    FromSingleFileMixin,
+    IPAdapterMixin,
+    StableDiffusionXLLoraLoaderMixin,
+    TextualInversionLoaderMixin,
+)
+from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
+from ...models.attention_processor import (
+    AttnProcessor2_0,
+    FusedAttnProcessor2_0,
+    LoRAAttnProcessor2_0,
+    LoRAXFormersAttnProcessor,
+    XFormersAttnProcessor,
+)
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import (
+    USE_PEFT_BACKEND,
+    deprecate,
+    is_invisible_watermark_available,
+    is_torch_xla_available,
+    logging,
+    replace_example_docstring,
+    scale_lora_layers,
+    unscale_lora_layers,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from .pipeline_output import StableDiffusionXLPipelineOutput
+
+
+if is_invisible_watermark_available():
+    from .watermark import StableDiffusionXLWatermarker
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers import StableDiffusionXLPipeline
+
+        >>> pipe = StableDiffusionXLPipeline.from_pretrained(
+        ...     "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
+        ... )
+        >>> pipe = pipe.to("cuda")
+
+        >>> prompt = "a photo of an astronaut riding a horse on mars"
+        >>> image = pipe(prompt).images[0]
+        ```
+"""
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg
+def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
+    """
+    Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
+    Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
+    """
+    std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
+    std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
+    # rescale the results from guidance (fixes overexposure)
+    noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
+    # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
+    noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
+    return noise_cfg
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    **kwargs,
+):
+    """
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+                Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
+                timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps`
+                must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+class StableDiffusionXLPipeline(
+    DiffusionPipeline,
+    StableDiffusionMixin,
+    FromSingleFileMixin,
+    StableDiffusionXLLoraLoaderMixin,
+    TextualInversionLoaderMixin,
+    IPAdapterMixin,
+):
+    r"""
+    Pipeline for text-to-image generation using Stable Diffusion XL.
+
+    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.)
+
+    The pipeline also inherits the following loading methods:
+        - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+        - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+        - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+        - [`~loaders.StableDiffusionXLLoraLoaderMixin.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.
+        text_encoder ([`CLIPTextModel`]):
+            Frozen text-encoder. Stable Diffusion XL uses the text portion of
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+        text_encoder_2 ([` CLIPTextModelWithProjection`]):
+            Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection),
+            specifically the
+            [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)
+            variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        tokenizer_2 (`CLIPTokenizer`):
+            Second Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+        force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`):
+            Whether the negative prompt embeddings shall be forced to always be set to 0. Also see the config of
+            `stabilityai/stable-diffusion-xl-base-1-0`.
+        add_watermarker (`bool`, *optional*):
+            Whether to use the [invisible_watermark library](https://github.com/ShieldMnt/invisible-watermark/) to
+            watermark output images. If not defined, it will default to True if the package is installed, otherwise no
+            watermarker will be used.
+    """
+
+    model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae"
+    _optional_components = [
+        "tokenizer",
+        "tokenizer_2",
+        "text_encoder",
+        "text_encoder_2",
+        "image_encoder",
+        "feature_extractor",
+    ]
+    _callback_tensor_inputs = [
+        "latents",
+        "prompt_embeds",
+        "negative_prompt_embeds",
+        "add_text_embeds",
+        "add_time_ids",
+        "negative_pooled_prompt_embeds",
+        "negative_add_time_ids",
+    ]
+
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModel,
+        text_encoder_2: CLIPTextModelWithProjection,
+        tokenizer: CLIPTokenizer,
+        tokenizer_2: CLIPTokenizer,
+        unet: UNet2DConditionModel,
+        scheduler: KarrasDiffusionSchedulers,
+        image_encoder: CLIPVisionModelWithProjection = None,
+        feature_extractor: CLIPImageProcessor = None,
+        force_zeros_for_empty_prompt: bool = True,
+        add_watermarker: Optional[bool] = None,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            text_encoder_2=text_encoder_2,
+            tokenizer=tokenizer,
+            tokenizer_2=tokenizer_2,
+            unet=unet,
+            scheduler=scheduler,
+            image_encoder=image_encoder,
+            feature_extractor=feature_extractor,
+        )
+        self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+        self.default_sample_size = self.unet.config.sample_size
+
+        add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available()
+
+        if add_watermarker:
+            self.watermark = StableDiffusionXLWatermarker()
+        else:
+            self.watermark = None
+
+    def encode_prompt(
+        self,
+        prompt: str,
+        prompt_2: Optional[str] = None,
+        device: Optional[torch.device] = None,
+        num_images_per_prompt: int = 1,
+        do_classifier_free_guidance: bool = True,
+        negative_prompt: Optional[str] = None,
+        negative_prompt_2: Optional[str] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        lora_scale: Optional[float] = None,
+        clip_skip: Optional[int] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+                used in both text-encoders
+            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]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            negative_prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+                `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+                If not provided, pooled text embeddings will be generated from `prompt` input argument.
+            negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+                input argument.
+            lora_scale (`float`, *optional*):
+                A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+            clip_skip (`int`, *optional*):
+                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+                the output of the pre-final layer will be used for computing the prompt embeddings.
+        """
+        device = device or self._execution_device
+
+        # 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, StableDiffusionXLLoraLoaderMixin):
+            self._lora_scale = lora_scale
+
+            # dynamically adjust the LoRA scale
+            if self.text_encoder is not None:
+                if not USE_PEFT_BACKEND:
+                    adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+                else:
+                    scale_lora_layers(self.text_encoder, lora_scale)
+
+            if self.text_encoder_2 is not None:
+                if not USE_PEFT_BACKEND:
+                    adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale)
+                else:
+                    scale_lora_layers(self.text_encoder_2, lora_scale)
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        # Define tokenizers and text encoders
+        tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2]
+        text_encoders = (
+            [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
+        )
+
+        if prompt_embeds is None:
+            prompt_2 = prompt_2 or prompt
+            prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2
+
+            # textual inversion: process multi-vector tokens if necessary
+            prompt_embeds_list = []
+            prompts = [prompt, prompt_2]
+            for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders):
+                if isinstance(self, TextualInversionLoaderMixin):
+                    prompt = self.maybe_convert_prompt(prompt, tokenizer)
+
+                text_inputs = tokenizer(
+                    prompt,
+                    padding="max_length",
+                    max_length=tokenizer.model_max_length,
+                    truncation=True,
+                    return_tensors="pt",
+                )
+
+                text_input_ids = text_inputs.input_ids
+                untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+                if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+                    text_input_ids, untruncated_ids
+                ):
+                    removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1])
+                    logger.warning(
+                        "The following part of your input was truncated because CLIP can only handle sequences up to"
+                        f" {tokenizer.model_max_length} tokens: {removed_text}"
+                    )
+
+                prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
+
+                # We are only ALWAYS interested in the pooled output of the final text encoder
+                pooled_prompt_embeds = prompt_embeds[0]
+                if clip_skip is None:
+                    prompt_embeds = prompt_embeds.hidden_states[-2]
+                else:
+                    # "2" because SDXL always indexes from the penultimate layer.
+                    prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)]
+
+                prompt_embeds_list.append(prompt_embeds)
+
+            prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
+
+        # get unconditional embeddings for classifier free guidance
+        zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt
+        if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt:
+            negative_prompt_embeds = torch.zeros_like(prompt_embeds)
+            negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
+        elif do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt_2 = negative_prompt_2 or negative_prompt
+
+            # normalize str to list
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+            negative_prompt_2 = (
+                batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2
+            )
+
+            uncond_tokens: List[str]
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            else:
+                uncond_tokens = [negative_prompt, negative_prompt_2]
+
+            negative_prompt_embeds_list = []
+            for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders):
+                if isinstance(self, TextualInversionLoaderMixin):
+                    negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer)
+
+                max_length = prompt_embeds.shape[1]
+                uncond_input = tokenizer(
+                    negative_prompt,
+                    padding="max_length",
+                    max_length=max_length,
+                    truncation=True,
+                    return_tensors="pt",
+                )
+
+                negative_prompt_embeds = text_encoder(
+                    uncond_input.input_ids.to(device),
+                    output_hidden_states=True,
+                )
+                # We are only ALWAYS interested in the pooled output of the final text encoder
+                negative_pooled_prompt_embeds = negative_prompt_embeds[0]
+                negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
+
+                negative_prompt_embeds_list.append(negative_prompt_embeds)
+
+            negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
+
+        if self.text_encoder_2 is not None:
+            prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
+        else:
+            prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device)
+
+        bs_embed, seq_len, _ = prompt_embeds.shape
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+        if do_classifier_free_guidance:
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = negative_prompt_embeds.shape[1]
+
+            if self.text_encoder_2 is not None:
+                negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
+            else:
+                negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device)
+
+            negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+            negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+        pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+            bs_embed * num_images_per_prompt, -1
+        )
+        if do_classifier_free_guidance:
+            negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+                bs_embed * num_images_per_prompt, -1
+            )
+
+        if self.text_encoder is not None:
+            if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
+                # Retrieve the original scale by scaling back the LoRA layers
+                unscale_lora_layers(self.text_encoder, lora_scale)
+
+        if self.text_encoder_2 is not None:
+            if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
+                # Retrieve the original scale by scaling back the LoRA layers
+                unscale_lora_layers(self.text_encoder_2, lora_scale)
+
+        return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_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.prepare_ip_adapter_image_embeds
+    def prepare_ip_adapter_image_embeds(
+        self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
+    ):
+        if ip_adapter_image_embeds is None:
+            if not isinstance(ip_adapter_image, list):
+                ip_adapter_image = [ip_adapter_image]
+
+            if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
+                raise ValueError(
+                    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
+            ):
+                output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
+                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
+                )
+
+                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)
+        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:]))
+                    )
+                image_embeds.append(single_image_embeds)
+
+        return image_embeds
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    def check_inputs(
+        self,
+        prompt,
+        prompt_2,
+        height,
+        width,
+        callback_steps,
+        negative_prompt=None,
+        negative_prompt_2=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        pooled_prompt_embeds=None,
+        negative_pooled_prompt_embeds=None,
+        ip_adapter_image=None,
+        ip_adapter_image_embeds=None,
+        callback_on_step_end_tensor_inputs=None,
+    ):
+        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 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)}."
+            )
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt_2 is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+        elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)):
+            raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}")
+
+        if negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+        elif negative_prompt_2 is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+
+        if prompt_embeds is not None and pooled_prompt_embeds is None:
+            raise ValueError(
+                "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`."
+            )
+
+        if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None:
+            raise ValueError(
+                "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`."
+            )
+
+        if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
+            raise ValueError(
+                "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
+            )
+
+        if ip_adapter_image_embeds is not None:
+            if not isinstance(ip_adapter_image_embeds, list):
+                raise ValueError(
+                    f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
+                )
+            elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
+                raise ValueError(
+                    f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
+                )
+
+    # 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 _get_add_time_ids(
+        self, original_size, crops_coords_top_left, target_size, dtype, text_encoder_projection_dim=None
+    ):
+        add_time_ids = list(original_size + crops_coords_top_left + target_size)
+
+        passed_add_embed_dim = (
+            self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim
+        )
+        expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features
+
+        if expected_add_embed_dim != passed_add_embed_dim:
+            raise ValueError(
+                f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."
+            )
+
+        add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
+        return add_time_ids
+
+    def upcast_vae(self):
+        dtype = self.vae.dtype
+        self.vae.to(dtype=torch.float32)
+        use_torch_2_0_or_xformers = isinstance(
+            self.vae.decoder.mid_block.attentions[0].processor,
+            (
+                AttnProcessor2_0,
+                XFormersAttnProcessor,
+                LoRAXFormersAttnProcessor,
+                LoRAAttnProcessor2_0,
+                FusedAttnProcessor2_0,
+            ),
+        )
+        # if xformers or torch_2_0 is used attention block does not need
+        # to be in float32 which can save lots of memory
+        if use_torch_2_0_or_xformers:
+            self.vae.post_quant_conv.to(dtype)
+            self.vae.decoder.conv_in.to(dtype)
+            self.vae.decoder.mid_block.to(dtype)
+
+    # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
+    def get_guidance_scale_embedding(
+        self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32
+    ) -> torch.FloatTensor:
+        """
+        See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
+
+        Args:
+            w (`torch.Tensor`):
+                Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings.
+            embedding_dim (`int`, *optional*, defaults to 512):
+                Dimension of the embeddings to generate.
+            dtype (`torch.dtype`, *optional*, defaults to `torch.float32`):
+                Data type of the generated embeddings.
+
+        Returns:
+            `torch.FloatTensor`: Embedding vectors with shape `(len(w), embedding_dim)`.
+        """
+        assert len(w.shape) == 1
+        w = w * 1000.0
+
+        half_dim = embedding_dim // 2
+        emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
+        emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
+        emb = w.to(dtype)[:, None] * emb[None, :]
+        emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+        if embedding_dim % 2 == 1:  # zero pad
+            emb = torch.nn.functional.pad(emb, (0, 1))
+        assert emb.shape == (w.shape[0], embedding_dim)
+        return emb
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def guidance_rescale(self):
+        return self._guidance_rescale
+
+    @property
+    def clip_skip(self):
+        return self._clip_skip
+
+    # 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.
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
+
+    @property
+    def cross_attention_kwargs(self):
+        return self._cross_attention_kwargs
+
+    @property
+    def denoising_end(self):
+        return self._denoising_end
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        prompt_2: Optional[Union[str, List[str]]] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 50,
+        timesteps: List[int] = None,
+        denoising_end: Optional[float] = None,
+        guidance_scale: float = 5.0,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        negative_prompt_2: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        ip_adapter_image: Optional[PipelineImageInput] = None,
+        ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        guidance_rescale: float = 0.0,
+        original_size: Optional[Tuple[int, int]] = None,
+        crops_coords_top_left: Tuple[int, int] = (0, 0),
+        target_size: Optional[Tuple[int, int]] = None,
+        negative_original_size: Optional[Tuple[int, int]] = None,
+        negative_crops_coords_top_left: Tuple[int, int] = (0, 0),
+        negative_target_size: Optional[Tuple[int, int]] = None,
+        clip_skip: Optional[int] = None,
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        **kwargs,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+                used in both text-encoders
+            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The height in pixels of the generated image. This is set to 1024 by default for the best results.
+                Anything below 512 pixels won't work well for
+                [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
+                and checkpoints that are not specifically fine-tuned on low resolutions.
+            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The width in pixels of the generated image. This is set to 1024 by default for the best results.
+                Anything below 512 pixels won't work well for
+                [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
+                and checkpoints that are not specifically fine-tuned on low resolutions.
+            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.
+            timesteps (`List[int]`, *optional*):
+                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+                passed will be used. Must be in descending order.
+            denoising_end (`float`, *optional*):
+                When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be
+                completed before it is intentionally prematurely terminated. As a result, the returned sample will
+                still retain a substantial amount of noise as determined by the discrete timesteps selected by the
+                scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a
+                "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image
+                Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output)
+            guidance_scale (`float`, *optional*, defaults to 5.0):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            negative_prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+                `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+                If not provided, pooled text embeddings will be generated from `prompt` input argument.
+            negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+                input argument.
+            ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
+            ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
+                Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of
+                IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should
+                contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not
+                provided, embeddings are computed from the `ip_adapter_image` input argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead
+                of a plain tuple.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            guidance_rescale (`float`, *optional*, defaults to 0.0):
+                Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are
+                Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of
+                [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf).
+                Guidance rescale factor should fix overexposure when using zero terminal SNR.
+            original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled.
+                `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as
+                explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+            crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+                `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
+                `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
+                `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+            target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                For most cases, `target_size` should be set to the desired height and width of the generated image. If
+                not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in
+                section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+            negative_original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                To negatively condition the generation process based on a specific image resolution. Part of SDXL's
+                micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+            negative_crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+                To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's
+                micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+            negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                To negatively condition the generation process based on a target image resolution. It should be as same
+                as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] or `tuple`:
+            [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a
+            `tuple`. When returning a tuple, the first element is a list with the generated images.
+        """
+
+        callback = kwargs.pop("callback", None)
+        callback_steps = kwargs.pop("callback_steps", None)
+
+        if callback is not None:
+            deprecate(
+                "callback",
+                "1.0.0",
+                "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+            )
+        if callback_steps is not None:
+            deprecate(
+                "callback_steps",
+                "1.0.0",
+                "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+            )
+
+        # 0. Default height and width to unet
+        height = height or self.default_sample_size * self.vae_scale_factor
+        width = width or self.default_sample_size * self.vae_scale_factor
+
+        original_size = original_size or (height, width)
+        target_size = target_size or (height, width)
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            prompt_2,
+            height,
+            width,
+            callback_steps,
+            negative_prompt,
+            negative_prompt_2,
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            negative_pooled_prompt_embeds,
+            ip_adapter_image,
+            ip_adapter_image_embeds,
+            callback_on_step_end_tensor_inputs,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._guidance_rescale = guidance_rescale
+        self._clip_skip = clip_skip
+        self._cross_attention_kwargs = cross_attention_kwargs
+        self._denoising_end = denoising_end
+        self._interrupt = False
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+
+        # 3. Encode input prompt
+        lora_scale = (
+            self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
+        )
+
+        (
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            negative_pooled_prompt_embeds,
+        ) = self.encode_prompt(
+            prompt=prompt,
+            prompt_2=prompt_2,
+            device=device,
+            num_images_per_prompt=num_images_per_prompt,
+            do_classifier_free_guidance=self.do_classifier_free_guidance,
+            negative_prompt=negative_prompt,
+            negative_prompt_2=negative_prompt_2,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            lora_scale=lora_scale,
+            clip_skip=self.clip_skip,
+        )
+
+        # 4. Prepare timesteps
+        timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+
+        # 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,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 7. Prepare added time ids & embeddings
+        add_text_embeds = pooled_prompt_embeds
+        if self.text_encoder_2 is None:
+            text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
+        else:
+            text_encoder_projection_dim = self.text_encoder_2.config.projection_dim
+
+        add_time_ids = self._get_add_time_ids(
+            original_size,
+            crops_coords_top_left,
+            target_size,
+            dtype=prompt_embeds.dtype,
+            text_encoder_projection_dim=text_encoder_projection_dim,
+        )
+        if negative_original_size is not None and negative_target_size is not None:
+            negative_add_time_ids = self._get_add_time_ids(
+                negative_original_size,
+                negative_crops_coords_top_left,
+                negative_target_size,
+                dtype=prompt_embeds.dtype,
+                text_encoder_projection_dim=text_encoder_projection_dim,
+            )
+        else:
+            negative_add_time_ids = add_time_ids
+
+        if self.do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+            add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
+            add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0)
+
+        prompt_embeds = prompt_embeds.to(device)
+        add_text_embeds = add_text_embeds.to(device)
+        add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
+
+        if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+            image_embeds = self.prepare_ip_adapter_image_embeds(
+                ip_adapter_image,
+                ip_adapter_image_embeds,
+                device,
+                batch_size * num_images_per_prompt,
+                self.do_classifier_free_guidance,
+            )
+
+        # 8. Denoising loop
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+
+        # 8.1 Apply denoising_end
+        if (
+            self.denoising_end is not None
+            and isinstance(self.denoising_end, float)
+            and self.denoising_end > 0
+            and self.denoising_end < 1
+        ):
+            discrete_timestep_cutoff = int(
+                round(
+                    self.scheduler.config.num_train_timesteps
+                    - (self.denoising_end * self.scheduler.config.num_train_timesteps)
+                )
+            )
+            num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps)))
+            timesteps = timesteps[:num_inference_steps]
+
+        # 9. Optionally get Guidance Scale Embedding
+        timestep_cond = None
+        if self.unet.config.time_cond_proj_dim is not None:
+            guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
+            timestep_cond = self.get_guidance_scale_embedding(
+                guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
+            ).to(device=device, dtype=latents.dtype)
+
+        self._num_timesteps = len(timesteps)
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                # expand the latents if we are doing classifier free guidance
+                latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+
+                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                # predict the noise residual
+                added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
+                if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+                    added_cond_kwargs["image_embeds"] = image_embeds
+
+                noise_pred = self.unet(
+                    latent_model_input,
+                    t,
+                    encoder_hidden_states=prompt_embeds,                # [B, 77, 2048]
+                    timestep_cond=timestep_cond,                        # None
+                    cross_attention_kwargs=self.cross_attention_kwargs, # None
+                    added_cond_kwargs=added_cond_kwargs,                # {[B, 1280], [B, 6]}
+                    return_dict=False,
+                )[0]
+
+                # perform guidance
+                if self.do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                if self.do_classifier_free_guidance and self.guidance_rescale > 0.0:
+                    # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+                    noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents_dtype = latents.dtype
+                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+                if latents.dtype != latents_dtype:
+                    if torch.backends.mps.is_available():
+                        # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
+                        latents = latents.to(latents_dtype)
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+                    add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds)
+                    negative_pooled_prompt_embeds = callback_outputs.pop(
+                        "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds
+                    )
+                    add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids)
+                    negative_add_time_ids = callback_outputs.pop("negative_add_time_ids", negative_add_time_ids)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+                    if callback is not None and i % callback_steps == 0:
+                        step_idx = i // getattr(self.scheduler, "order", 1)
+                        callback(step_idx, t, latents)
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        if not output_type == "latent":
+            # make sure the VAE is in float32 mode, as it overflows in float16
+            needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+
+            if needs_upcasting:
+                self.upcast_vae()
+                latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
+            elif latents.dtype != self.vae.dtype:
+                if torch.backends.mps.is_available():
+                    # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
+                    self.vae = self.vae.to(latents.dtype)
+
+            # unscale/denormalize the latents
+            # denormalize with the mean and std if available and not None
+            has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None
+            has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None
+            if has_latents_mean and has_latents_std:
+                latents_mean = (
+                    torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype)
+                )
+                latents_std = (
+                    torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype)
+                )
+                latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean
+            else:
+                latents = latents / self.vae.config.scaling_factor
+
+            image = self.vae.decode(latents, return_dict=False)[0]
+
+            # cast back to fp16 if needed
+            if needs_upcasting:
+                self.vae.to(dtype=torch.float16)
+        else:
+            image = latents
+
+        if not output_type == "latent":
+            # apply watermark if available
+            if self.watermark is not None:
+                image = self.watermark.apply_watermark(image)
+
+            image = self.image_processor.postprocess(image, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (image,)
+
+        return StableDiffusionXLPipelineOutput(images=image)