Update pipeline_stable_diffusion_xl_instantid_img2img.py

pipeline_stable_diffusion_xl_instantid_img2img.py

@@ -1,3 +1,18 @@
+# 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.
+# 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 math
 from typing import Any, Callable, Dict, List, Optional, Tuple, Union
 
@@ -6,69 +21,55 @@ import numpy as np
 import PIL.Image
 import torch
 import torch.nn as nn
+
 from diffusers import StableDiffusionXLControlNetImg2ImgPipeline
 from diffusers.image_processor import PipelineImageInput
 from diffusers.models import ControlNetModel
 from diffusers.pipelines.controlnet.multicontrolnet import MultiControlNetModel
 from diffusers.pipelines.stable_diffusion_xl import StableDiffusionXLPipelineOutput
-from diffusers.utils import deprecate, logging, replace_example_docstring
+from diffusers.utils import (
+    deprecate,
+    logging,
+    replace_example_docstring,
+)
 from diffusers.utils.import_utils import is_xformers_available
 from diffusers.utils.torch_utils import is_compiled_module, is_torch_version
 
-logger = logging.get_logger(__name__)  # Initialize logger
 
-# Check for xformers availability
 try:
     import xformers
     import xformers.ops
 
     xformers_available = True
-except ImportError:
+except Exception:
     xformers_available = False
 
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
-def reshape_tensor(x: torch.Tensor, heads: int) -> torch.Tensor:
-    """Reshapes tensor for multi-head attention processing."""
-    bs, length, width = x.shape
-    return x.view(bs, length, heads, -1).transpose(1, 2)
-
-def draw_kps(image_pil, kps, color_list=[(255, 0, 0), (0, 255, 0), (0, 0, 255), (255, 255, 0), (255, 0, 255)]):
-    import numpy as np
-    import cv2
-    from PIL import Image
-
-    stickwidth = 4
-    limbSeq = np.array([[0, 2], [1, 2], [3, 2], [4, 2]])
-    kps = np.array(kps)
-
-    w, h = image_pil.size
-    out_img = np.zeros([h, w, 3])
-
-    for i in range(len(limbSeq)):
-        index = limbSeq[i]
-        color = color_list[index[0]]
 
-        x = kps[index][:, 0]
-        y = kps[index][:, 1]
-        length = ((x[0] - x[1]) ** 2 + (y[0] - y[1]) ** 2) ** 0.5
-        angle = np.degrees(np.arctan2(y[0] - y[1], x[0] - x[1]))
-        polygon = cv2.ellipse2Poly(
-            (int(np.mean(x)), int(np.mean(y))), (int(length / 2), stickwidth), int(angle), 0, 360, 1
-        )
-        out_img = cv2.fillConvexPoly(out_img.copy(), polygon, color)
-    out_img = (out_img * 0.6).astype(np.uint8)
+def FeedForward(dim, mult=4):
+    inner_dim = int(dim * mult)
+    return nn.Sequential(
+        nn.LayerNorm(dim),
+        nn.Linear(dim, inner_dim, bias=False),
+        nn.GELU(),
+        nn.Linear(inner_dim, dim, bias=False),
+    )
 
-    for idx_kp, kp in enumerate(kps):
-        color = color_list[idx_kp]
-        x, y = kp
-        out_img = cv2.circle(out_img.copy(), (int(x), int(y)), 10, color, -1)
 
-    out_img_pil = Image.fromarray(out_img.astype(np.uint8))
-    return out_img_pil
+def reshape_tensor(x, heads):
+    bs, length, width = x.shape
+    # (bs, length, width) --> (bs, length, n_heads, dim_per_head)
+    x = x.view(bs, length, heads, -1)
+    # (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head)
+    x = x.transpose(1, 2)
+    # (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head)
+    x = x.reshape(bs, heads, length, -1)
+    return x
 
 
 class PerceiverAttention(nn.Module):
-    def __init__(self, dim: int, dim_head: int = 64, heads: int = 8):
+    def __init__(self, *, dim, dim_head=64, heads=8):
         super().__init__()
         self.scale = dim_head**-0.5
         self.dim_head = dim_head
@@ -77,96 +78,995 @@ class PerceiverAttention(nn.Module):
 
         self.norm1 = nn.LayerNorm(dim)
         self.norm2 = nn.LayerNorm(dim)
+
         self.to_q = nn.Linear(dim, inner_dim, bias=False)
         self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False)
         self.to_out = nn.Linear(inner_dim, dim, bias=False)
 
-    def forward(self, x: torch.Tensor, latents: torch.Tensor) -> torch.Tensor:
-        x, latents = self.norm1(x), self.norm2(latents)
-        q, kv = self.to_q(latents), self.to_kv(torch.cat((x, latents), dim=1))
-        k, v = kv.chunk(2, dim=-1)
+    def forward(self, x, latents):
+        """
+        Args:
+            x (torch.Tensor): image features
+                shape (b, n1, D)
+            latent (torch.Tensor): latent features
+                shape (b, n2, D)
+        """
+        x = self.norm1(x)
+        latents = self.norm2(latents)
+
+        b, l, _ = latents.shape
 
-        q, k, v = map(lambda t: reshape_tensor(t, self.heads), (q, k, v))
+        q = self.to_q(latents)
+        kv_input = torch.cat((x, latents), dim=-2)
+        k, v = self.to_kv(kv_input).chunk(2, dim=-1)
 
-        # Scaled dot-product attention
+        q = reshape_tensor(q, self.heads)
+        k = reshape_tensor(k, self.heads)
+        v = reshape_tensor(v, self.heads)
+
+        # attention
         scale = 1 / math.sqrt(math.sqrt(self.dim_head))
-        weight = (q * scale) @ (k * scale).transpose(-2, -1)
-        weight = torch.softmax(weight.float(), dim=-1).to(weight.dtype)
+        weight = (q * scale) @ (k * scale).transpose(-2, -1)  # More stable with f16 than dividing afterwards
+        weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
         out = weight @ v
 
-        return self.to_out(out.permute(0, 2, 1, 3).reshape(latents.shape[0], latents.shape[1], -1))
+        out = out.permute(0, 2, 1, 3).reshape(b, l, -1)
+
+        return self.to_out(out)
 
 
 class Resampler(nn.Module):
     def __init__(
         self,
-        dim: int = 1024,
-        depth: int = 8,
-        dim_head: int = 64,
-        heads: int = 16,
-        num_queries: int = 8,
-        embedding_dim: int = 768,
-        output_dim: int = 1024,
-        ff_mult: int = 4,
+        dim=1024,
+        depth=8,
+        dim_head=64,
+        heads=16,
+        num_queries=8,
+        embedding_dim=768,
+        output_dim=1024,
+        ff_mult=4,
     ):
         super().__init__()
-        self.latents = nn.Parameter(torch.randn(1, num_queries, dim) / math.sqrt(dim))
+
+        self.latents = nn.Parameter(torch.randn(1, num_queries, dim) / dim**0.5)
+
         self.proj_in = nn.Linear(embedding_dim, dim)
+
         self.proj_out = nn.Linear(dim, output_dim)
         self.norm_out = nn.LayerNorm(output_dim)
-        self.layers = nn.ModuleList([nn.ModuleList([PerceiverAttention(dim, dim_head, heads), nn.LayerNorm(dim)]) for _ in range(depth)])
 
-    def forward(self, x: torch.Tensor) -> torch.Tensor:
-        latents = self.latents.expand(x.size(0), -1, -1)
+        self.layers = nn.ModuleList([])
+        for _ in range(depth):
+            self.layers.append(
+                nn.ModuleList(
+                    [
+                        PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads),
+                        FeedForward(dim=dim, mult=ff_mult),
+                    ]
+                )
+            )
+
+    def forward(self, x):
+        latents = self.latents.repeat(x.size(0), 1, 1)
         x = self.proj_in(x)
 
-        for attn, norm in self.layers:
-            latents = norm(attn(x, latents) + latents)
+        for attn, ff in self.layers:
+            latents = attn(x, latents) + latents
+            latents = ff(latents) + latents
+
+        latents = self.proj_out(latents)
+        return self.norm_out(latents)
+
+
+class AttnProcessor(nn.Module):
+    r"""
+    Default processor for performing attention-related computations.
+    """
+
+    def __init__(
+        self,
+        hidden_size=None,
+        cross_attention_dim=None,
+    ):
+        super().__init__()
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        query = attn.head_to_batch_dim(query)
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        hidden_states = torch.bmm(attention_probs, value)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class IPAttnProcessor(nn.Module):
+    r"""
+    Attention processor for IP-Adapater.
+    Args:
+        hidden_size (`int`):
+            The hidden size of the attention layer.
+        cross_attention_dim (`int`):
+            The number of channels in the `encoder_hidden_states`.
+        scale (`float`, defaults to 1.0):
+            the weight scale of image prompt.
+        num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
+            The context length of the image features.
+    """
+
+    def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, num_tokens=4):
+        super().__init__()
+
+        self.hidden_size = hidden_size
+        self.cross_attention_dim = cross_attention_dim
+        self.scale = scale
+        self.num_tokens = num_tokens
+
+        self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+        self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        else:
+            # get encoder_hidden_states, ip_hidden_states
+            end_pos = encoder_hidden_states.shape[1] - self.num_tokens
+            encoder_hidden_states, ip_hidden_states = (
+                encoder_hidden_states[:, :end_pos, :],
+                encoder_hidden_states[:, end_pos:, :],
+            )
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        query = attn.head_to_batch_dim(query)
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        if xformers_available:
+            hidden_states = self._memory_efficient_attention_xformers(query, key, value, attention_mask)
+        else:
+            attention_probs = attn.get_attention_scores(query, key, attention_mask)
+            hidden_states = torch.bmm(attention_probs, value)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        # for ip-adapter
+        ip_key = self.to_k_ip(ip_hidden_states)
+        ip_value = self.to_v_ip(ip_hidden_states)
+
+        ip_key = attn.head_to_batch_dim(ip_key)
+        ip_value = attn.head_to_batch_dim(ip_value)
+
+        if xformers_available:
+            ip_hidden_states = self._memory_efficient_attention_xformers(query, ip_key, ip_value, None)
+        else:
+            ip_attention_probs = attn.get_attention_scores(query, ip_key, None)
+            ip_hidden_states = torch.bmm(ip_attention_probs, ip_value)
+        ip_hidden_states = attn.batch_to_head_dim(ip_hidden_states)
+
+        hidden_states = hidden_states + self.scale * ip_hidden_states
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+    def _memory_efficient_attention_xformers(self, query, key, value, attention_mask):
+        # TODO attention_mask
+        query = query.contiguous()
+        key = key.contiguous()
+        value = value.contiguous()
+        hidden_states = xformers.ops.memory_efficient_attention(query, key, value, attn_bias=attention_mask)
+        return hidden_states
+
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> # !pip install opencv-python transformers accelerate insightface
+        >>> import diffusers
+        >>> from diffusers.utils import load_image
+        >>> from diffusers.models import ControlNetModel
+
+        >>> import cv2
+        >>> import torch
+        >>> import numpy as np
+        >>> from PIL import Image
+
+        >>> from insightface.app import FaceAnalysis
+        >>> from pipeline_stable_diffusion_xl_instantid import StableDiffusionXLInstantIDPipeline, draw_kps
+
+        >>> # download 'antelopev2' under ./models
+        >>> app = FaceAnalysis(name='antelopev2', root='./', providers=['CUDAExecutionProvider', 'CPUExecutionProvider'])
+        >>> app.prepare(ctx_id=0, det_size=(640, 640))
+
+        >>> # download models under ./checkpoints
+        >>> face_adapter = f'./checkpoints/ip-adapter.bin'
+        >>> controlnet_path = f'./checkpoints/ControlNetModel'
+
+        >>> # load IdentityNet
+        >>> controlnet = ControlNetModel.from_pretrained(controlnet_path, torch_dtype=torch.float16)
+
+        >>> pipe = StableDiffusionXLInstantIDPipeline.from_pretrained(
+        ...     "stabilityai/stable-diffusion-xl-base-1.0", controlnet=controlnet, torch_dtype=torch.float16
+        ... )
+        >>> pipe.cuda()
 
-        return self.norm_out(self.proj_out(latents))
+        >>> # load adapter
+        >>> pipe.load_ip_adapter_instantid(face_adapter)
+
+        >>> prompt = "analog film photo of a man. faded film, desaturated, 35mm photo, grainy, vignette, vintage, Kodachrome, Lomography, stained, highly detailed, found footage, masterpiece, best quality"
+        >>> negative_prompt = "(lowres, low quality, worst quality:1.2), (text:1.2), watermark, painting, drawing, illustration, glitch, deformed, mutated, cross-eyed, ugly, disfigured (lowres, low quality, worst quality:1.2), (text:1.2), watermark, painting, drawing, illustration, glitch,deformed, mutated, cross-eyed, ugly, disfigured"
+
+        >>> # load an image
+        >>> image = load_image("your-example.jpg")
+
+        >>> face_info = app.get(cv2.cvtColor(np.array(face_image), cv2.COLOR_RGB2BGR))[-1]
+        >>> face_emb = face_info['embedding']
+        >>> face_kps = draw_kps(face_image, face_info['kps'])
+
+        >>> pipe.set_ip_adapter_scale(0.8)
+
+        >>> # generate image
+        >>> image = pipe(
+        ...     prompt, image_embeds=face_emb, image=face_kps, controlnet_conditioning_scale=0.8
+        ... ).images[0]
+        ```
+"""
+
+
+def draw_kps(image_pil, kps, color_list=[(255, 0, 0), (0, 255, 0), (0, 0, 255), (255, 255, 0), (255, 0, 255)]):
+    stickwidth = 4
+    limbSeq = np.array([[0, 2], [1, 2], [3, 2], [4, 2]])
+    kps = np.array(kps)
+
+    w, h = image_pil.size
+    out_img = np.zeros([h, w, 3])
+
+    for i in range(len(limbSeq)):
+        index = limbSeq[i]
+        color = color_list[index[0]]
+
+        x = kps[index][:, 0]
+        y = kps[index][:, 1]
+        length = ((x[0] - x[1]) ** 2 + (y[0] - y[1]) ** 2) ** 0.5
+        angle = math.degrees(math.atan2(y[0] - y[1], x[0] - x[1]))
+        polygon = cv2.ellipse2Poly(
+            (int(np.mean(x)), int(np.mean(y))), (int(length / 2), stickwidth), int(angle), 0, 360, 1
+        )
+        out_img = cv2.fillConvexPoly(out_img.copy(), polygon, color)
+    out_img = (out_img * 0.6).astype(np.uint8)
+
+    for idx_kp, kp in enumerate(kps):
+        color = color_list[idx_kp]
+        x, y = kp
+        out_img = cv2.circle(out_img.copy(), (int(x), int(y)), 10, color, -1)
+
+    out_img_pil = PIL.Image.fromarray(out_img.astype(np.uint8))
+    return out_img_pil
 
 
 class StableDiffusionXLInstantIDImg2ImgPipeline(StableDiffusionXLControlNetImg2ImgPipeline):
-    def cuda(self, dtype: torch.dtype = torch.float16, use_xformers: bool = False):
+    def cuda(self, dtype=torch.float16, use_xformers=False):
         self.to("cuda", dtype)
+
         if hasattr(self, "image_proj_model"):
             self.image_proj_model.to(self.unet.device).to(self.unet.dtype)
 
         if use_xformers:
             if is_xformers_available():
+                import xformers
+                from packaging import version
+
+                xformers_version = version.parse(xformers.__version__)
+                if xformers_version == version.parse("0.0.16"):
+                    logger.warning(
+                        "xFormers 0.0.16 cannot be used for training in some GPUs. If you observe problems during training, please update xFormers to at least 0.0.17. See https://huggingface.co/docs/diffusers/main/en/optimization/xformers for more details."
+                    )
                 self.enable_xformers_memory_efficient_attention()
             else:
-                raise ValueError("xFormers is not available. Ensure it is installed correctly.")
+                raise ValueError("xformers is not available. Make sure it is installed correctly")
 
-    def load_ip_adapter_instantid(self, model_ckpt: str, image_emb_dim: int = 512, num_tokens: int = 16, scale: float = 0.5):
+    def load_ip_adapter_instantid(self, model_ckpt, image_emb_dim=512, num_tokens=16, scale=0.5):
         self.set_image_proj_model(model_ckpt, image_emb_dim, num_tokens)
         self.set_ip_adapter(model_ckpt, num_tokens, scale)
 
-    def set_image_proj_model(self, model_ckpt: str, image_emb_dim: int = 512, num_tokens: int = 16):
-        self.image_proj_model = Resampler(
-            dim=1280, depth=4, dim_head=64, heads=20, num_queries=num_tokens, embedding_dim=image_emb_dim, output_dim=self.unet.config.cross_attention_dim
-        ).to(self.device, dtype=self.dtype).eval()
+    def set_image_proj_model(self, model_ckpt, image_emb_dim=512, num_tokens=16):
+        image_proj_model = Resampler(
+            dim=1280,
+            depth=4,
+            dim_head=64,
+            heads=20,
+            num_queries=num_tokens,
+            embedding_dim=image_emb_dim,
+            output_dim=self.unet.config.cross_attention_dim,
+            ff_mult=4,
+        )
+
+        image_proj_model.eval()
 
-        state_dict = torch.load(model_ckpt, map_location="cpu").get("image_proj", torch.load(model_ckpt, map_location="cpu"))
+        self.image_proj_model = image_proj_model.to(self.device, dtype=self.dtype)
+        state_dict = torch.load(model_ckpt, map_location="cpu")
+        if "image_proj" in state_dict:
+            state_dict = state_dict["image_proj"]
         self.image_proj_model.load_state_dict(state_dict)
+
         self.image_proj_model_in_features = image_emb_dim
 
-    def set_ip_adapter(self, model_ckpt: str, num_tokens: int, scale: float):
+    def set_ip_adapter(self, model_ckpt, num_tokens, scale):
+        unet = self.unet
         attn_procs = {}
-        for name, module in self.unet.attn_processors.items():
-            cross_attention_dim = None if name.endswith("attn1.processor") else self.unet.config.cross_attention_dim
-            hidden_size = self.unet.config.block_out_channels[{"mid_block": -1, "up_blocks": int(name[9]), "down_blocks": int(name[12])}[name.split(".")[0]]]
-            attn_procs[name] = (IPAttnProcessor(hidden_size, cross_attention_dim, scale, num_tokens)
-                                if cross_attention_dim else nn.Identity()).to(self.unet.device, dtype=self.unet.dtype)
+        for name in unet.attn_processors.keys():
+            cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim
+            if name.startswith("mid_block"):
+                hidden_size = unet.config.block_out_channels[-1]
+            elif name.startswith("up_blocks"):
+                block_id = int(name[len("up_blocks.")])
+                hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
+            elif name.startswith("down_blocks"):
+                block_id = int(name[len("down_blocks.")])
+                hidden_size = unet.config.block_out_channels[block_id]
+            if cross_attention_dim is None:
+                attn_procs[name] = AttnProcessor().to(unet.device, dtype=unet.dtype)
+            else:
+                attn_procs[name] = IPAttnProcessor(
+                    hidden_size=hidden_size,
+                    cross_attention_dim=cross_attention_dim,
+                    scale=scale,
+                    num_tokens=num_tokens,
+                ).to(unet.device, dtype=unet.dtype)
+        unet.set_attn_processor(attn_procs)
+
+        state_dict = torch.load(model_ckpt, map_location="cpu")
+        ip_layers = torch.nn.ModuleList(self.unet.attn_processors.values())
+        if "ip_adapter" in state_dict:
+            state_dict = state_dict["ip_adapter"]
+        ip_layers.load_state_dict(state_dict)
 
-        self.unet.set_attn_processor(attn_procs)
-        self.unet.attn_processors.load_state_dict(torch.load(model_ckpt, map_location="cpu").get("ip_adapter", torch.load(model_ckpt, map_location="cpu")))
+    def set_ip_adapter_scale(self, scale):
+        unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
+        for attn_processor in unet.attn_processors.values():
+            if isinstance(attn_processor, IPAttnProcessor):
+                attn_processor.scale = scale
 
     def _encode_prompt_image_emb(self, prompt_image_emb, device, dtype, do_classifier_free_guidance):
-        prompt_image_emb = torch.tensor(prompt_image_emb) if not isinstance(prompt_image_emb, torch.Tensor) else prompt_image_emb.clone().detach()
-        prompt_image_emb = prompt_image_emb.to(device, dtype=dtype).reshape([1, -1, self.image_proj_model_in_features])
+        if isinstance(prompt_image_emb, torch.Tensor):
+            prompt_image_emb = prompt_image_emb.clone().detach()
+        else:
+            prompt_image_emb = torch.tensor(prompt_image_emb)
+
+        prompt_image_emb = prompt_image_emb.to(device=device, dtype=dtype)
+        prompt_image_emb = prompt_image_emb.reshape([1, -1, self.image_proj_model_in_features])
 
         if do_classifier_free_guidance:
             prompt_image_emb = torch.cat([torch.zeros_like(prompt_image_emb), prompt_image_emb], dim=0)
+        else:
+            prompt_image_emb = torch.cat([prompt_image_emb], dim=0)
+        image_proj_model_device = self.image_proj_model.to(device)
+        prompt_image_emb = image_proj_model_device(prompt_image_emb)
+        return prompt_image_emb
+
+    @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,
+        image: PipelineImageInput = None,
+        control_image: PipelineImageInput = None,
+        strength: float = 0.8,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 50,
+        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,
+        image_embeds: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        controlnet_conditioning_scale: Union[float, List[float]] = 1.0,
+        guess_mode: bool = False,
+        control_guidance_start: Union[float, List[float]] = 0.0,
+        control_guidance_end: Union[float, List[float]] = 1.0,
+        original_size: Tuple[int, int] = None,
+        crops_coords_top_left: Tuple[int, int] = (0, 0),
+        target_size: 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,
+        aesthetic_score: float = 6.0,
+        negative_aesthetic_score: float = 2.5,
+        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"""
+        The call function to the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+            prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+                used in both text-encoders.
+            image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, `List[np.ndarray]`,:
+                    `List[List[torch.FloatTensor]]`, `List[List[np.ndarray]]` or `List[List[PIL.Image.Image]]`):
+                The ControlNet input condition to provide guidance to the `unet` for generation. If the type is
+                specified as `torch.FloatTensor`, it is passed to ControlNet as is. `PIL.Image.Image` can also be
+                accepted as an image. The dimensions of the output image defaults to `image`'s dimensions. If height
+                and/or width are passed, `image` is resized accordingly. If multiple ControlNets are specified in
+                `init`, images must be passed as a list such that each element of the list can be correctly batched for
+                input to a single ControlNet.
+            height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+                The height in pixels of the generated image. 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. 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.
+            guidance_scale (`float`, *optional*, defaults to 5.0):
+                A higher guidance scale value encourages the model to generate images closely linked to the text
+                `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+                pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+            negative_prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide what to not include in image generation. This is 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) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+                to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                A [`torch.Generator`](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 is 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 (prompt weighting). If not
+                provided, text embeddings are generated from the `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+                not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+                not provided, pooled text embeddings are 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 (prompt
+                weighting). If not provided, pooled `negative_prompt_embeds` are generated from `negative_prompt` input
+                argument.
+            image_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated image embeddings.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+                plain tuple.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+                [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            controlnet_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 1.0):
+                The outputs of the ControlNet are multiplied by `controlnet_conditioning_scale` before they are added
+                to the residual in the original `unet`. If multiple ControlNets are specified in `init`, you can set
+                the corresponding scale as a list.
+            guess_mode (`bool`, *optional*, defaults to `False`):
+                The ControlNet encoder tries to recognize the content of the input image even if you remove all
+                prompts. A `guidance_scale` value between 3.0 and 5.0 is recommended.
+            control_guidance_start (`float` or `List[float]`, *optional*, defaults to 0.0):
+                The percentage of total steps at which the ControlNet starts applying.
+            control_guidance_end (`float` or `List[float]`, *optional*, defaults to 1.0):
+                The percentage of total steps at which the ControlNet stops applying.
+            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.
+            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.
+            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.StableDiffusionPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+                otherwise a `tuple` is returned containing the output 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 using `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 using `callback_on_step_end`",
+            )
+
+        controlnet = self.controlnet._orig_mod if is_compiled_module(self.controlnet) else self.controlnet
+
+        # align format for control guidance
+        if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list):
+            control_guidance_start = len(control_guidance_end) * [control_guidance_start]
+        elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list):
+            control_guidance_end = len(control_guidance_start) * [control_guidance_end]
+        elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list):
+            mult = len(controlnet.nets) if isinstance(controlnet, MultiControlNetModel) else 1
+            control_guidance_start, control_guidance_end = (
+                mult * [control_guidance_start],
+                mult * [control_guidance_end],
+            )
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            prompt_2,
+            control_image,
+            strength,
+            num_inference_steps,
+            callback_steps,
+            negative_prompt,
+            negative_prompt_2,
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            negative_pooled_prompt_embeds,
+            None,
+            None,
+            controlnet_conditioning_scale,
+            control_guidance_start,
+            control_guidance_end,
+            callback_on_step_end_tensor_inputs,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._clip_skip = clip_skip
+        self._cross_attention_kwargs = cross_attention_kwargs
+
+        # 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
+
+        if isinstance(controlnet, MultiControlNetModel) and isinstance(controlnet_conditioning_scale, float):
+            controlnet_conditioning_scale = [controlnet_conditioning_scale] * len(controlnet.nets)
+
+        global_pool_conditions = (
+            controlnet.config.global_pool_conditions
+            if isinstance(controlnet, ControlNetModel)
+            else controlnet.nets[0].config.global_pool_conditions
+        )
+        guess_mode = guess_mode or global_pool_conditions
+
+        # 3.1 Encode input prompt
+        text_encoder_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_2,
+            device,
+            num_images_per_prompt,
+            self.do_classifier_free_guidance,
+            negative_prompt,
+            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=text_encoder_lora_scale,
+            clip_skip=self.clip_skip,
+        )
+
+        # 3.2 Encode image prompt
+        prompt_image_emb = self._encode_prompt_image_emb(
+            image_embeds, device, self.unet.dtype, self.do_classifier_free_guidance
+        )
+        bs_embed, seq_len, _ = prompt_image_emb.shape
+        prompt_image_emb = prompt_image_emb.repeat(1, num_images_per_prompt, 1)
+        prompt_image_emb = prompt_image_emb.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+        # 4. Prepare image and controlnet_conditioning_image
+        image = self.image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32)
+
+        if isinstance(controlnet, ControlNetModel):
+            control_image = self.prepare_control_image(
+                image=control_image,
+                width=width,
+                height=height,
+                batch_size=batch_size * num_images_per_prompt,
+                num_images_per_prompt=num_images_per_prompt,
+                device=device,
+                dtype=controlnet.dtype,
+                do_classifier_free_guidance=self.do_classifier_free_guidance,
+                guess_mode=guess_mode,
+            )
+            height, width = control_image.shape[-2:]
+        elif isinstance(controlnet, MultiControlNetModel):
+            control_images = []
+
+            for control_image_ in control_image:
+                control_image_ = self.prepare_control_image(
+                    image=control_image_,
+                    width=width,
+                    height=height,
+                    batch_size=batch_size * num_images_per_prompt,
+                    num_images_per_prompt=num_images_per_prompt,
+                    device=device,
+                    dtype=controlnet.dtype,
+                    do_classifier_free_guidance=self.do_classifier_free_guidance,
+                    guess_mode=guess_mode,
+                )
+
+                control_images.append(control_image_)
+
+            control_image = control_images
+            height, width = control_image[0].shape[-2:]
+        else:
+            assert False
+
+        # 5. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device)
+        latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
+        self._num_timesteps = len(timesteps)
+
+        # 6. Prepare latent variables
+        latents = self.prepare_latents(
+            image,
+            latent_timestep,
+            batch_size,
+            num_images_per_prompt,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            True,
+        )
+
+        # # 6.5 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)
+
+        # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 7.1 Create tensor stating which controlnets to keep
+        controlnet_keep = []
+        for i in range(len(timesteps)):
+            keeps = [
+                1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e)
+                for s, e in zip(control_guidance_start, control_guidance_end)
+            ]
+            controlnet_keep.append(keeps[0] if isinstance(controlnet, ControlNetModel) else keeps)
+
+        # 7.2 Prepare added time ids & embeddings
+        if isinstance(control_image, list):
+            original_size = original_size or control_image[0].shape[-2:]
+        else:
+            original_size = original_size or control_image.shape[-2:]
+        target_size = target_size or (height, width)
+
+        if negative_original_size is None:
+            negative_original_size = original_size
+        if negative_target_size is None:
+            negative_target_size = target_size
+        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, add_neg_time_ids = self._get_add_time_ids(
+            original_size,
+            crops_coords_top_left,
+            target_size,
+            aesthetic_score,
+            negative_aesthetic_score,
+            negative_original_size,
+            negative_crops_coords_top_left,
+            negative_target_size,
+            dtype=prompt_embeds.dtype,
+            text_encoder_projection_dim=text_encoder_projection_dim,
+        )
+        add_time_ids = add_time_ids.repeat(batch_size * num_images_per_prompt, 1)
+
+        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_neg_time_ids = add_neg_time_ids.repeat(batch_size * num_images_per_prompt, 1)
+            add_time_ids = torch.cat([add_neg_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)
+        encoder_hidden_states = torch.cat([prompt_embeds, prompt_image_emb], dim=1)
+
+        # 8. Denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        is_unet_compiled = is_compiled_module(self.unet)
+        is_controlnet_compiled = is_compiled_module(self.controlnet)
+        is_torch_higher_equal_2_1 = is_torch_version(">=", "2.1")
+
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                # Relevant thread:
+                # https://dev-discuss.pytorch.org/t/cudagraphs-in-pytorch-2-0/1428
+                if (is_unet_compiled and is_controlnet_compiled) and is_torch_higher_equal_2_1:
+                    torch._inductor.cudagraph_mark_step_begin()
+                # 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)
+
+                added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
+
+                # controlnet(s) inference
+                if guess_mode and self.do_classifier_free_guidance:
+                    # Infer ControlNet only for the conditional batch.
+                    control_model_input = latents
+                    control_model_input = self.scheduler.scale_model_input(control_model_input, t)
+                    controlnet_prompt_embeds = prompt_embeds.chunk(2)[1]
+                    controlnet_added_cond_kwargs = {
+                        "text_embeds": add_text_embeds.chunk(2)[1],
+                        "time_ids": add_time_ids.chunk(2)[1],
+                    }
+                else:
+                    control_model_input = latent_model_input
+                    controlnet_prompt_embeds = prompt_embeds
+                    controlnet_added_cond_kwargs = added_cond_kwargs
+
+                if isinstance(controlnet_keep[i], list):
+                    cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])]
+                else:
+                    controlnet_cond_scale = controlnet_conditioning_scale
+                    if isinstance(controlnet_cond_scale, list):
+                        controlnet_cond_scale = controlnet_cond_scale[0]
+                    cond_scale = controlnet_cond_scale * controlnet_keep[i]
+
+                down_block_res_samples, mid_block_res_sample = self.controlnet(
+                    control_model_input,
+                    t,
+                    encoder_hidden_states=prompt_image_emb,
+                    controlnet_cond=control_image,
+                    conditioning_scale=cond_scale,
+                    guess_mode=guess_mode,
+                    added_cond_kwargs=controlnet_added_cond_kwargs,
+                    return_dict=False,
+                )
+
+                if guess_mode and self.do_classifier_free_guidance:
+                    # Infered ControlNet only for the conditional batch.
+                    # To apply the output of ControlNet to both the unconditional and conditional batches,
+                    # add 0 to the unconditional batch to keep it unchanged.
+                    down_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples]
+                    mid_block_res_sample = torch.cat([torch.zeros_like(mid_block_res_sample), mid_block_res_sample])
+
+                # predict the noise residual
+                noise_pred = self.unet(
+                    latent_model_input,
+                    t,
+                    encoder_hidden_states=encoder_hidden_states,
+                    timestep_cond=timestep_cond,
+                    cross_attention_kwargs=self.cross_attention_kwargs,
+                    down_block_additional_residuals=down_block_res_samples,
+                    mid_block_additional_residual=mid_block_res_sample,
+                    added_cond_kwargs=added_cond_kwargs,
+                    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 + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+                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)
+
+                # 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 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)
+
+            image = self.vae.decode(latents / self.vae.config.scaling_factor, 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 self.image_proj_model.to(device)(prompt_image_emb)
+        return StableDiffusionXLPipelineOutput(images=image)