add turbo

app.py

@@ -5,11 +5,10 @@ import os
 import torch
 from PIL import Image
 
-from pipeline_freescale import StableDiffusionXLPipeline
 from free_lunch_utils import register_free_upblock2d, register_free_crossattn_upblock2d
 
 @spaces.GPU(duration=120)
-def infer_gpu_part(pipe, seed, prompt, negative_prompt, ddim_steps, guidance_scale, resolutions_list, fast_mode, cosine_scale, disable_freeu):
+def infer_gpu_normal(pipe, seed, prompt, negative_prompt, ddim_steps, guidance_scale, resolutions_list, fast_mode, cosine_scale, disable_freeu, restart_steps):
     pipe = pipe.to("cuda") 
     generator = torch.Generator(device='cuda')
     generator = generator.manual_seed(seed)
@@ -19,28 +18,69 @@ def infer_gpu_part(pipe, seed, prompt, negative_prompt, ddim_steps, guidance_sca
     result = pipe(prompt, negative_prompt=negative_prompt, generator=generator,
                 num_inference_steps=ddim_steps, guidance_scale=guidance_scale,
                 resolutions_list=resolutions_list, fast_mode=fast_mode, cosine_scale=cosine_scale,
+                restart_steps=restart_steps,
+                ).images[0]
+    return result
+
+@spaces.GPU(duration=30)
+def infer_gpu_turbo(pipe, seed, prompt, negative_prompt, ddim_steps, guidance_scale, resolutions_list, fast_mode, cosine_scale, disable_freeu, restart_steps):
+    pipe = pipe.to("cuda") 
+    generator = torch.Generator(device='cuda')
+    generator = generator.manual_seed(seed)
+    if not disable_freeu:
+        register_free_upblock2d(pipe, b1=1.1, b2=1.2, s1=0.6, s2=0.4)
+        register_free_crossattn_upblock2d(pipe, b1=1.1, b2=1.2, s1=0.6, s2=0.4)
+    result = pipe(prompt, negative_prompt=negative_prompt, generator=generator,
+                num_inference_steps=ddim_steps, guidance_scale=guidance_scale,
+                resolutions_list=resolutions_list, fast_mode=fast_mode, cosine_scale=cosine_scale,
+                restart_steps=restart_steps,
                 ).images[0]
     return result
 
 def infer(prompt, output_size, ddim_steps, guidance_scale, cosine_scale, seed, options, negative_prompt):
 
+    disable_turbo = 'Disable Turbo' in options
     disable_freeu = 'Disable FreeU' in options
-    fast_mode = True
-    if output_size == "2048 x 2048":
-        resolutions_list = [[1024, 1024],
-                            [2048, 2048]]
-    elif output_size == "1024 x 2048":
-        resolutions_list = [[512, 1024],
-                            [1024, 2048]]
-    elif output_size == "2048 x 1024":
-        resolutions_list = [[1024, 512],
-                            [2048, 1024]]
-
-    model_ckpt = "stabilityai/stable-diffusion-xl-base-1.0"
+
+    if disable_turbo:
+        from pipeline_freescale import StableDiffusionXLPipeline
+        model_ckpt = "stabilityai/stable-diffusion-xl-base-1.0"
+        fast_mode = True
+        if output_size == "2048 x 2048":
+            resolutions_list = [[1024, 1024],
+                                [2048, 2048]]
+        elif output_size == "1024 x 2048":
+            resolutions_list = [[512, 1024],
+                                [1024, 2048]]
+        elif output_size == "2048 x 1024":
+            resolutions_list = [[1024, 512],
+                                [2048, 1024]]
+        infer_gpu_part = infer_gpu_normal
+        restart_steps = [int(ddim_steps * 0.3)] * len(resolutions_list)
+
+    else:
+        from pipeline_freescale_turbo import StableDiffusionXLPipeline
+        model_ckpt = "stabilityai/sdxl-turbo"
+        fast_mode = False
+        if output_size == "2048 x 2048":
+            resolutions_list = [[512, 512],
+                                [1024, 1024],
+                                [2048, 2048]]
+        elif output_size == "1024 x 2048":
+            resolutions_list = [[256, 512]
+                                [512, 1024],
+                                [1024, 2048]]
+        elif output_size == "2048 x 1024":
+            resolutions_list = [[512, 256]
+                                [1024, 512],
+                                [2048, 1024]]
+        infer_gpu_part = infer_gpu_turbo
+        restart_steps = [int(ddim_steps * 0.5)] * len(resolutions_list)
+
     pipe = StableDiffusionXLPipeline.from_pretrained(model_ckpt, torch_dtype=torch.float16)
 
     print('GPU starts')
-    result = infer_gpu_part(pipe, seed, prompt, negative_prompt, ddim_steps, guidance_scale, resolutions_list, fast_mode, cosine_scale, disable_freeu)
+    result = infer_gpu_part(pipe, seed, prompt, negative_prompt, ddim_steps, guidance_scale, resolutions_list, fast_mode, cosine_scale, disable_freeu, restart_steps)
     print('GPU ends')
 
     save_path = 'output.png'
@@ -138,6 +178,16 @@ img[src*='#center'] {
     }
 """
 
+def step_update(options):
+    if 'Disable Turbo' in options:
+        return gr.Slider(minimum=5,
+                        maximum=200,
+                        value=50)
+    else:
+        return gr.Slider(minimum=2,
+                        maximum=8,
+                        value=4)
+
 with gr.Blocks(css=css) as demo:
     with gr.Column(elem_id="col-container"):
         gr.Markdown(
@@ -160,13 +210,13 @@ with gr.Blocks(css=css) as demo:
             with gr.Accordion('FreeScale Parameters (feel free to adjust these parameters based on your prompt): ', open=False):
                 with gr.Row():
                     output_size = gr.Dropdown(["2048 x 2048", "1024 x 2048", "2048 x 1024"], value="2048 x 2048", label="Output Size (H x W)", info="Due to GPU constraints, run the demo locally for higher resolutions.", scale=2)
-                    options = gr.CheckboxGroup(['Disable FreeU'], label='Options (NOT recommended to change)', scale=1)
+                    options = gr.CheckboxGroup(['Disable Turbo', 'Disable FreeU'], label='Options (NOT recommended to change)', scale=1)
                 with gr.Row():
                     ddim_steps = gr.Slider(label='DDIM Steps',
-                             minimum=5,
-                             maximum=200,
+                             minimum=2,
+                             maximum=8,
                              step=1,
-                             value=50)
+                             value=4)
                     guidance_scale = gr.Slider(label='Guidance Scale',
                              minimum=1.0,
                              maximum=20.0,
@@ -186,6 +236,8 @@ with gr.Blocks(css=css) as demo:
                 with gr.Row():
                     negative_prompt = gr.Textbox(label='Negative Prompt', value='blurry, ugly, duplicate, poorly drawn, deformed, mosaic')
 
+        options.change(step_update, options, ddim_steps)
+
         submit_btn = gr.Button("Generate", variant='primary')
         image_result = gr.Image(label="Image Output")
 

pipeline_freescale_turbo.py

@@ -0,0 +1,1204 @@
+import inspect
+import os
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import torch
+from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer
+
+from diffusers.image_processor import VaeImageProcessor
+from diffusers.loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from diffusers.models import AutoencoderKL, UNet2DConditionModel
+from diffusers.models.attention_processor import (
+    AttnProcessor2_0,
+    LoRAAttnProcessor2_0,
+    LoRAXFormersAttnProcessor,
+    XFormersAttnProcessor,
+)
+from diffusers.schedulers import KarrasDiffusionSchedulers
+from diffusers.utils import (
+    is_accelerate_available,
+    is_accelerate_version,
+    is_invisible_watermark_available,
+    logging,
+    randn_tensor,
+    replace_example_docstring,
+)
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.pipelines.stable_diffusion_xl import StableDiffusionXLPipelineOutput
+
+if is_invisible_watermark_available():
+    from .watermark import StableDiffusionXLWatermarker
+
+from inspect import isfunction
+from functools import partial
+import numpy as np
+
+from diffusers.models.attention import BasicTransformerBlock
+from scale_attention_turbo import ori_forward, scale_forward
+
+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]
+        ```
+"""
+
+def default(val, d):
+    if exists(val):
+        return val
+    return d() if isfunction(d) else d
+
+def exists(val):
+    return val is not None
+
+def extract_into_tensor(a, t, x_shape):
+    b, *_ = t.shape
+    out = a.gather(-1, t)
+    return out.reshape(b, *((1,) * (len(x_shape) - 1)))
+
+def make_beta_schedule(schedule, n_timestep, linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
+    if schedule == "linear":
+        betas = (
+                torch.linspace(linear_start ** 0.5, linear_end ** 0.5, n_timestep, dtype=torch.float64) ** 2
+        )
+    elif schedule == "cosine":
+        timesteps = (
+                torch.arange(n_timestep + 1, dtype=torch.float64) / n_timestep + cosine_s
+        )
+        alphas = timesteps / (1 + cosine_s) * np.pi / 2
+        alphas = torch.cos(alphas).pow(2)
+        alphas = alphas / alphas[0]
+        betas = 1 - alphas[1:] / alphas[:-1]
+        betas = np.clip(betas, a_min=0, a_max=0.999)
+    elif schedule == "sqrt_linear":
+        betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64)
+    elif schedule == "sqrt":
+        betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64) ** 0.5
+    else:
+        raise ValueError(f"schedule '{schedule}' unknown.")
+    return betas.numpy()
+
+to_torch = partial(torch.tensor, dtype=torch.float16)
+betas = make_beta_schedule("linear", 1000, linear_start=0.00085, linear_end=0.012)
+alphas = 1. - betas
+alphas_cumprod = np.cumprod(alphas, axis=0)
+sqrt_alphas_cumprod = to_torch(np.sqrt(alphas_cumprod))
+sqrt_one_minus_alphas_cumprod = to_torch(np.sqrt(1. - alphas_cumprod))
+
+def q_sample(x_start, t, init_noise_sigma = 1.0, noise=None, device=None):
+    noise = default(noise, lambda: torch.randn_like(x_start)).to(device) * init_noise_sigma
+    return (extract_into_tensor(sqrt_alphas_cumprod.to(device), t, x_start.shape) * x_start +
+            extract_into_tensor(sqrt_one_minus_alphas_cumprod.to(device), t, x_start.shape) * noise)
+
+def get_views(height, width, h_window_size=128, w_window_size=128, h_window_stride=64, w_window_stride=64, vae_scale_factor=8):
+    height //= vae_scale_factor
+    width //= vae_scale_factor
+    num_blocks_height = int((height - h_window_size) / h_window_stride - 1e-6) + 2 if height > h_window_size else 1
+    num_blocks_width = int((width - w_window_size) / w_window_stride - 1e-6) + 2 if width > w_window_size else 1
+    total_num_blocks = int(num_blocks_height * num_blocks_width)
+    views = []
+    for i in range(total_num_blocks):
+        h_start = int((i // num_blocks_width) * h_window_stride)
+        h_end = h_start + h_window_size
+        w_start = int((i % num_blocks_width) * w_window_stride)
+        w_end = w_start + w_window_size
+
+        if h_end > height:
+            h_start = int(h_start + height - h_end)
+            h_end = int(height)
+        if w_end > width:
+            w_start = int(w_start + width - w_end)
+            w_end = int(width)
+        if h_start < 0:
+            h_end = int(h_end - h_start)
+            h_start = 0
+        if w_start < 0:
+            w_end = int(w_end - w_start)
+            w_start = 0
+
+        random_jitter = True
+        if random_jitter:
+            h_jitter_range = (h_window_size - h_window_stride) // 4
+            w_jitter_range = (w_window_size - w_window_stride) // 4
+            h_jitter = 0
+            w_jitter = 0
+            
+            if (w_start != 0) and (w_end != width):
+                w_jitter = random.randint(-w_jitter_range, w_jitter_range)
+            elif (w_start == 0) and (w_end != width):
+                w_jitter = random.randint(-w_jitter_range, 0)
+            elif (w_start != 0) and (w_end == width):
+                w_jitter = random.randint(0, w_jitter_range)
+            if (h_start != 0) and (h_end != height):
+                h_jitter = random.randint(-h_jitter_range, h_jitter_range)
+            elif (h_start == 0) and (h_end != height):
+                h_jitter = random.randint(-h_jitter_range, 0)
+            elif (h_start != 0) and (h_end == height):
+                h_jitter = random.randint(0, h_jitter_range)
+            h_start += (h_jitter + h_jitter_range)
+            h_end += (h_jitter + h_jitter_range)
+            w_start += (w_jitter + w_jitter_range)
+            w_end += (w_jitter + w_jitter_range)
+
+        views.append((h_start, h_end, w_start, w_end))
+    return views
+
+def gaussian_kernel(kernel_size=3, sigma=1.0, channels=3):
+    x_coord = torch.arange(kernel_size)
+    gaussian_1d = torch.exp(-(x_coord - (kernel_size - 1) / 2) ** 2 / (2 * sigma ** 2))
+    gaussian_1d = gaussian_1d / gaussian_1d.sum()
+    gaussian_2d = gaussian_1d[:, None] * gaussian_1d[None, :]
+    kernel = gaussian_2d[None, None, :, :].repeat(channels, 1, 1, 1)
+    
+    return kernel
+
+def gaussian_filter(latents, kernel_size=3, sigma=1.0):
+    channels = latents.shape[1]
+    kernel = gaussian_kernel(kernel_size, sigma, channels).to(latents.device, latents.dtype)
+    if len(latents.shape) == 5:
+        b = latents.shape[0]
+        latents = rearrange(latents, 'b c t i j -> (b t) c i j')
+        blurred_latents = F.conv2d(latents, kernel, padding=kernel_size//2, groups=channels)
+        blurred_latents = rearrange(blurred_latents, '(b t) c i j -> b c t i j', b=b)
+    else:
+        blurred_latents = F.conv2d(latents, kernel, padding=kernel_size//2, groups=channels)
+    
+    return blurred_latents
+
+# 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
+
+
+class StableDiffusionXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoaderMixin):
+    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.)
+
+    In addition the pipeline inherits the following loading methods:
+        - *LoRA*: [`StableDiffusionXLPipeline.load_lora_weights`]
+        - *Ckpt*: [`loaders.FromSingleFileMixin.from_single_file`]
+
+    as well as the following saving methods:
+        - *LoRA*: [`loaders.StableDiffusionXLPipeline.save_lora_weights`]
+
+    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`].
+    """
+
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModel,
+        text_encoder_2: CLIPTextModelWithProjection,
+        tokenizer: CLIPTokenizer,
+        tokenizer_2: CLIPTokenizer,
+        unet: UNet2DConditionModel,
+        scheduler: KarrasDiffusionSchedulers,
+        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,
+        )
+        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
+
+        self.vae.enable_tiling()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
+    def enable_vae_tiling(self):
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+        """
+        self.vae.enable_tiling()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
+    def enable_model_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
+        to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
+        method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
+        `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
+        """
+        if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
+            from accelerate import cpu_offload_with_hook
+        else:
+            raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        if self.device.type != "cpu":
+            self.to("cpu", silence_dtype_warnings=True)
+            torch.cuda.empty_cache()  # otherwise we don't see the memory savings (but they probably exist)
+
+        model_sequence = (
+            [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
+        )
+        model_sequence.extend([self.unet, self.vae])
+
+        hook = None
+        for cpu_offloaded_model in model_sequence:
+            _, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook)
+
+        # We'll offload the last model manually.
+        self.final_offload_hook = hook
+
+    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,
+    ):
+        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.
+        """
+        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, LoraLoaderMixin):
+            self._lora_scale = lora_scale
+
+        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]
+
+        # 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
+            # textual inversion: procecss 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]
+                prompt_embeds = prompt_embeds.hidden_states[-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
+
+            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 isinstance(negative_prompt, str):
+                uncond_tokens = [negative_prompt, negative_prompt_2]
+            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)
+
+        prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.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]
+            negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.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
+            )
+
+        return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_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,
+    ):
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+        if (callback_steps is None) or (
+            callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+        ):
+            raise ValueError(
+                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+                f" {type(callback_steps)}."
+            )
+
+        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`."
+            )
+
+    # 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, height // self.vae_scale_factor, 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):
+        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) + self.text_encoder_2.config.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
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
+    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,
+            ),
+        )
+        # 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)
+
+    @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,
+        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,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: int = 1,
+        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,
+        resolutions_list: Optional[Union[int, List[int]]] = None,
+        restart_steps: Optional[Union[int, List[int]]] = None,
+        cosine_scale: float = 2.0,
+        dilate_tau: int = 35,
+        fast_mode: bool = False,
+    ):
+        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.
+            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The width in pixels of the generated image.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            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.
+            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.
+            callback (`Callable`, *optional*):
+                A function that will be called every `callback_steps` steps during inference. The function will be
+                called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+            callback_steps (`int`, *optional*, defaults to 1):
+                The frequency at which the `callback` function will be called. If not specified, the callback will be
+                called at every step.
+            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 0.7):
+                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 `(width, height)` 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 `(width, height)`. 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).
+
+        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.
+        """
+        
+
+        # 0. Default height and width to unet
+        if resolutions_list:
+            height, width = resolutions_list[0]
+            target_sizes = resolutions_list[1:]
+            if not restart_steps:
+                restart_steps = [1] * len(target_sizes)
+        else:
+            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,
+        )
+
+        # 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
+
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        # 3. Encode input prompt
+        text_encoder_lora_scale = (
+            cross_attention_kwargs.get("scale", None) if 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=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=text_encoder_lora_scale,
+        )
+
+        # 4. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+
+        timesteps = self.scheduler.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
+        add_time_ids = self._get_add_time_ids(
+            original_size, crops_coords_top_left, target_size, dtype=prompt_embeds.dtype
+        )
+
+        if 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([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)
+
+        # 8. Denoising loop
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+
+        # 9.1 Apply denoising_end
+        if denoising_end is not None and type(denoising_end) == float and denoising_end > 0 and denoising_end < 1:
+            discrete_timestep_cutoff = int(
+                round(
+                    self.scheduler.config.num_train_timesteps
+                    - (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]
+
+        for block in self.unet.down_blocks + [self.unet.mid_block] + self.unet.up_blocks:
+            for module in block.modules():
+                if isinstance(module, BasicTransformerBlock):
+                    module.forward = ori_forward.__get__(module, BasicTransformerBlock)
+
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                # expand the latents if we are doing classifier free guidance
+                latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+
+                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                # predict the noise residual
+                added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
+                noise_pred = self.unet(
+                    latent_model_input,
+                    t,
+                    encoder_hidden_states=prompt_embeds,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    added_cond_kwargs=added_cond_kwargs,
+                    return_dict=False,
+                )[0]
+
+                # perform guidance
+                if do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                if do_classifier_free_guidance and 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=guidance_rescale)
+
+                # 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]
+
+                # 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:
+                        callback(i, t, latents)
+
+        for restart_index, target_size in enumerate(target_sizes):
+            restart_step = restart_steps[restart_index]
+            target_size_ = [target_size[0]//8, target_size[1]//8]
+
+            for block in self.unet.down_blocks + [self.unet.mid_block] + self.unet.up_blocks:
+                for module in block.modules():
+                    if isinstance(module, BasicTransformerBlock):
+                        module.forward = scale_forward.__get__(module, BasicTransformerBlock)
+                        module.current_hw = target_size
+                        module.fast_mode = fast_mode
+
+            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)
+
+            latents = latents / self.vae.config.scaling_factor
+            image = self.vae.decode(latents, return_dict=False)[0]
+            image = torch.nn.functional.interpolate(
+                image,
+                size=target_size,
+                mode='bicubic',
+                )
+            latents = self.vae.encode(image).latent_dist.sample().half()
+            latents = latents * self.vae.config.scaling_factor
+
+            noise_latents = []
+            noise = torch.randn_like(latents)
+            for timestep in self.scheduler.timesteps:
+                noise_latent = self.scheduler.add_noise(latents, noise, timestep.unsqueeze(0))
+                noise_latents.append(noise_latent)
+            latents = noise_latents[restart_step]
+
+            self.scheduler._step_index = 0
+            with self.progress_bar(total=num_inference_steps) as progress_bar:
+                for i, t in enumerate(timesteps):
+
+                    if i < restart_step:
+                        self.scheduler._step_index += 1
+                        progress_bar.update()
+                        continue
+
+                    cosine_factor = 0.5 * (1 + torch.cos(torch.pi * (self.scheduler.config.num_train_timesteps - t) / self.scheduler.config.num_train_timesteps)).cpu()
+                    c1 = cosine_factor ** cosine_scale
+                    latents = latents * (1 - c1) + noise_latents[i] * c1
+
+                    dilate_coef=target_size[1]//1024
+
+                    dilate_layers = [
+                        # "down_blocks.1.resnets.0.conv1",
+                        # "down_blocks.1.resnets.0.conv2",
+                        # "down_blocks.1.resnets.1.conv1",
+                        # "down_blocks.1.resnets.1.conv2",
+                        "down_blocks.1.downsamplers.0.conv",
+                        "down_blocks.2.resnets.0.conv1",
+                        "down_blocks.2.resnets.0.conv2",
+                        "down_blocks.2.resnets.1.conv1",
+                        "down_blocks.2.resnets.1.conv2",
+                        # "up_blocks.0.resnets.0.conv1",
+                        # "up_blocks.0.resnets.0.conv2",
+                        # "up_blocks.0.resnets.1.conv1",
+                        # "up_blocks.0.resnets.1.conv2",
+                        # "up_blocks.0.resnets.2.conv1",
+                        # "up_blocks.0.resnets.2.conv2",
+                        # "up_blocks.0.upsamplers.0.conv",
+                        # "up_blocks.1.resnets.0.conv1",
+                        # "up_blocks.1.resnets.0.conv2",
+                        # "up_blocks.1.resnets.1.conv1",
+                        # "up_blocks.1.resnets.1.conv2",
+                        # "up_blocks.1.resnets.2.conv1",
+                        # "up_blocks.1.resnets.2.conv2",
+                        # "up_blocks.1.upsamplers.0.conv",
+                        # "up_blocks.2.resnets.0.conv1",
+                        # "up_blocks.2.resnets.0.conv2",
+                        # "up_blocks.2.resnets.1.conv1",
+                        # "up_blocks.2.resnets.1.conv2",
+                        # "up_blocks.2.resnets.2.conv1",
+                        # "up_blocks.2.resnets.2.conv2",
+                        "mid_block.resnets.0.conv1",
+                        "mid_block.resnets.0.conv2",
+                        "mid_block.resnets.1.conv1",
+                        "mid_block.resnets.1.conv2"
+                        ]
+
+                    for name, module in self.unet.named_modules():
+                        if name in dilate_layers:
+                            if i < dilate_tau:
+                                module.dilation = (dilate_coef, dilate_coef)
+                                module.padding = (dilate_coef, dilate_coef)
+                            else:
+                                module.dilation = (1, 1)
+                                module.padding = (1, 1)
+
+                    # expand the latents if we are doing classifier free guidance
+                    latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+
+                    latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+
+                    # predict the noise residual
+                    added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
+                    noise_pred = self.unet(
+                        latent_model_input,
+                        t,
+                        encoder_hidden_states=prompt_embeds,
+                        cross_attention_kwargs=cross_attention_kwargs,
+                        added_cond_kwargs=added_cond_kwargs,
+                        return_dict=False,
+                    )[0]
+
+                    # perform guidance
+                    if do_classifier_free_guidance:
+                        noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                        noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                    if do_classifier_free_guidance and 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=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)
+
+                    # 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:
+                            callback(i, t, latents)
+
+                    for name, module in self.unet.named_modules():
+                        # if ('.conv' in name) and ('.conv_' not in name):
+                        if name in dilate_layers:
+                            module.dilation = (1, 1)
+                            module.padding = (1, 1)
+
+        # make sure the VAE is in float32 mode, as it overflows in float16
+        if self.vae.dtype == torch.float16 and self.vae.config.force_upcast:
+            self.upcast_vae()
+            latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
+
+        if not output_type == "latent":
+            image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+        else:
+            image = latents
+            return StableDiffusionXLPipelineOutput(images=image)
+
+        # 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 last model to CPU
+        if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+            self.final_offload_hook.offload()
+
+        if not return_dict:
+            return (image,)
+
+        return StableDiffusionXLPipelineOutput(images=image)
+
+    # Overrride to properly handle the loading and unloading of the additional text encoder.
+    def load_lora_weights(self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]], **kwargs):
+        # We could have accessed the unet config from `lora_state_dict()` too. We pass
+        # it here explicitly to be able to tell that it's coming from an SDXL
+        # pipeline.
+        state_dict, network_alphas = self.lora_state_dict(
+            pretrained_model_name_or_path_or_dict,
+            unet_config=self.unet.config,
+            **kwargs,
+        )
+        self.load_lora_into_unet(state_dict, network_alphas=network_alphas, unet=self.unet)
+
+        text_encoder_state_dict = {k: v for k, v in state_dict.items() if "text_encoder." in k}
+        if len(text_encoder_state_dict) > 0:
+            self.load_lora_into_text_encoder(
+                text_encoder_state_dict,
+                network_alphas=network_alphas,
+                text_encoder=self.text_encoder,
+                prefix="text_encoder",
+                lora_scale=self.lora_scale,
+            )
+
+        text_encoder_2_state_dict = {k: v for k, v in state_dict.items() if "text_encoder_2." in k}
+        if len(text_encoder_2_state_dict) > 0:
+            self.load_lora_into_text_encoder(
+                text_encoder_2_state_dict,
+                network_alphas=network_alphas,
+                text_encoder=self.text_encoder_2,
+                prefix="text_encoder_2",
+                lora_scale=self.lora_scale,
+            )
+
+    @classmethod
+    def save_lora_weights(
+        self,
+        save_directory: Union[str, os.PathLike],
+        unet_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
+        text_encoder_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
+        text_encoder_2_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
+        is_main_process: bool = True,
+        weight_name: str = None,
+        save_function: Callable = None,
+        safe_serialization: bool = True,
+    ):
+        state_dict = {}
+
+        def pack_weights(layers, prefix):
+            layers_weights = layers.state_dict() if isinstance(layers, torch.nn.Module) else layers
+            layers_state_dict = {f"{prefix}.{module_name}": param for module_name, param in layers_weights.items()}
+            return layers_state_dict
+
+        state_dict.update(pack_weights(unet_lora_layers, "unet"))
+
+        if text_encoder_lora_layers and text_encoder_2_lora_layers:
+            state_dict.update(pack_weights(text_encoder_lora_layers, "text_encoder"))
+            state_dict.update(pack_weights(text_encoder_2_lora_layers, "text_encoder_2"))
+
+        self.write_lora_layers(
+            state_dict=state_dict,
+            save_directory=save_directory,
+            is_main_process=is_main_process,
+            weight_name=weight_name,
+            save_function=save_function,
+            safe_serialization=safe_serialization,
+        )
+
+    def _remove_text_encoder_monkey_patch(self):
+        self._remove_text_encoder_monkey_patch_classmethod(self.text_encoder)
+        self._remove_text_encoder_monkey_patch_classmethod(self.text_encoder_2)

scale_attention_turbo.py

@@ -0,0 +1,372 @@
+from typing import Any, Dict, Optional
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+from einops import rearrange, repeat
+import random
+
+def gaussian_kernel(kernel_size=3, sigma=1.0, channels=3):
+    x_coord = torch.arange(kernel_size)
+    gaussian_1d = torch.exp(-(x_coord - (kernel_size - 1) / 2) ** 2 / (2 * sigma ** 2))
+    gaussian_1d = gaussian_1d / gaussian_1d.sum()
+    gaussian_2d = gaussian_1d[:, None] * gaussian_1d[None, :]
+    kernel = gaussian_2d[None, None, :, :].repeat(channels, 1, 1, 1)
+    
+    return kernel
+
+def gaussian_filter(latents, kernel_size=3, sigma=1.0):
+    channels = latents.shape[1]
+    kernel = gaussian_kernel(kernel_size, sigma, channels).to(latents.device, latents.dtype)
+    blurred_latents = F.conv2d(latents, kernel, padding=kernel_size//2, groups=channels)
+    
+    return blurred_latents
+    
+def get_views(height, width, h_window_size=64, w_window_size=64, scale_factor=8):
+    height = int(height)
+    width = int(width)
+    h_window_stride = h_window_size // 2
+    w_window_stride = w_window_size // 2
+    h_window_size = int(h_window_size / scale_factor)
+    w_window_size = int(w_window_size / scale_factor)
+    h_window_stride = int(h_window_stride / scale_factor)
+    w_window_stride = int(w_window_stride / scale_factor)
+    num_blocks_height = int((height - h_window_size) / h_window_stride - 1e-6) + 2 if height > h_window_size else 1
+    num_blocks_width = int((width - w_window_size) / w_window_stride - 1e-6) + 2 if width > w_window_size else 1
+    total_num_blocks = int(num_blocks_height * num_blocks_width)
+    views = []
+    for i in range(total_num_blocks):
+        h_start = int((i // num_blocks_width) * h_window_stride)
+        h_end = h_start + h_window_size
+        w_start = int((i % num_blocks_width) * w_window_stride)
+        w_end = w_start + w_window_size
+
+        if h_end > height:
+            h_start = int(h_start + height - h_end)
+            h_end = int(height)
+        if w_end > width:
+            w_start = int(w_start + width - w_end)
+            w_end = int(width)
+        if h_start < 0:
+            h_end = int(h_end - h_start)
+            h_start = 0
+        if w_start < 0:
+            w_end = int(w_end - w_start)
+            w_start = 0
+
+        random_jitter = True
+        if random_jitter:
+            h_jitter_range = h_window_size // 8
+            w_jitter_range = w_window_size // 8
+            h_jitter = 0
+            w_jitter = 0
+            
+            if (w_start != 0) and (w_end != width):
+                w_jitter = random.randint(-w_jitter_range, w_jitter_range)
+            elif (w_start == 0) and (w_end != width):
+                w_jitter = random.randint(-w_jitter_range, 0)
+            elif (w_start != 0) and (w_end == width):
+                w_jitter = random.randint(0, w_jitter_range)
+            if (h_start != 0) and (h_end != height):
+                h_jitter = random.randint(-h_jitter_range, h_jitter_range)
+            elif (h_start == 0) and (h_end != height):
+                h_jitter = random.randint(-h_jitter_range, 0)
+            elif (h_start != 0) and (h_end == height):
+                h_jitter = random.randint(0, h_jitter_range)
+            h_start += (h_jitter + h_jitter_range)
+            h_end += (h_jitter + h_jitter_range)
+            w_start += (w_jitter + w_jitter_range)
+            w_end += (w_jitter + w_jitter_range)
+
+        views.append((h_start, h_end, w_start, w_end))
+    return views
+
+def scale_forward(
+    self,
+    hidden_states: torch.FloatTensor,
+    attention_mask: Optional[torch.FloatTensor] = None,
+    encoder_hidden_states: Optional[torch.FloatTensor] = None,
+    encoder_attention_mask: Optional[torch.FloatTensor] = None,
+    timestep: Optional[torch.LongTensor] = None,
+    cross_attention_kwargs: Dict[str, Any] = None,
+    class_labels: Optional[torch.LongTensor] = None,
+):
+    # Notice that normalization is always applied before the real computation in the following blocks.
+    if self.current_hw:
+        current_scale_num_h, current_scale_num_w = self.current_hw[0] // 512, self.current_hw[1] // 512
+    else:
+        current_scale_num_h, current_scale_num_w = 1, 1
+
+    # 0. Self-Attention
+    if self.use_ada_layer_norm:
+        norm_hidden_states = self.norm1(hidden_states, timestep)
+    elif self.use_ada_layer_norm_zero:
+        norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
+            hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
+        )
+    else:
+        norm_hidden_states = self.norm1(hidden_states)
+
+    # 2. Prepare GLIGEN inputs
+    cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
+    gligen_kwargs = cross_attention_kwargs.pop("gligen", None)
+
+    ratio_hw = current_scale_num_h / current_scale_num_w
+    latent_h = int((norm_hidden_states.shape[1] * ratio_hw) ** 0.5)
+    latent_w = int(latent_h / ratio_hw)
+    scale_factor = 64 * current_scale_num_h / latent_h
+    if ratio_hw > 1:
+        sub_h = 64
+        sub_w = int(64 / ratio_hw)
+    else:
+        sub_h = int(64 * ratio_hw)
+        sub_w = 64 
+
+    h_jitter_range = int(sub_h / scale_factor // 8)
+    w_jitter_range = int(sub_w / scale_factor // 8)
+    views = get_views(latent_h, latent_w, sub_h, sub_w, scale_factor = scale_factor)
+
+    current_scale_num = max(current_scale_num_h, current_scale_num_w)
+    global_views = [[h, w] for h in range(current_scale_num_h) for w in range(current_scale_num_w)]
+
+    if self.fast_mode:
+        four_window = False
+        fourg_window = True
+    else:
+        four_window = True
+        fourg_window = False
+
+    if four_window:
+        norm_hidden_states_ = rearrange(norm_hidden_states, 'bh (h w) d -> bh h w d', h = latent_h)
+        norm_hidden_states_ = F.pad(norm_hidden_states_, (0, 0, w_jitter_range, w_jitter_range, h_jitter_range, h_jitter_range), 'constant', 0)
+        value = torch.zeros_like(norm_hidden_states_)
+        count = torch.zeros_like(norm_hidden_states_)
+        for index, view in enumerate(views):
+            h_start, h_end, w_start, w_end = view
+            local_states = norm_hidden_states_[:, h_start:h_end, w_start:w_end, :]
+            local_states = rearrange(local_states, 'bh h w d -> bh (h w) d')
+            local_output = self.attn1(
+                local_states,
+                encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
+                attention_mask=attention_mask,
+                **cross_attention_kwargs,
+            )
+            local_output = rearrange(local_output, 'bh (h w) d -> bh h w d', h = int(sub_h / scale_factor))
+
+            value[:, h_start:h_end, w_start:w_end, :] += local_output * 1
+            count[:, h_start:h_end, w_start:w_end, :] += 1
+
+        value = value[:, h_jitter_range:-h_jitter_range, w_jitter_range:-w_jitter_range, :]
+        count = count[:, h_jitter_range:-h_jitter_range, w_jitter_range:-w_jitter_range, :]
+        attn_output = torch.where(count>0, value/count, value)
+        
+        gaussian_local = gaussian_filter(attn_output, kernel_size=(2*current_scale_num-1), sigma=1.0)
+
+        attn_output_global = self.attn1(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
+            attention_mask=attention_mask,
+            **cross_attention_kwargs,
+        )
+        attn_output_global = rearrange(attn_output_global, 'bh (h w) d -> bh h w d', h = latent_h)
+
+        gaussian_global = gaussian_filter(attn_output_global, kernel_size=(2*current_scale_num-1), sigma=1.0)
+
+        attn_output = gaussian_local + (attn_output_global - gaussian_global)
+        attn_output = rearrange(attn_output, 'bh h w d -> bh (h w) d')
+
+    elif fourg_window:
+        norm_hidden_states = rearrange(norm_hidden_states, 'bh (h w) d -> bh h w d', h = latent_h)
+        norm_hidden_states_ = F.pad(norm_hidden_states, (0, 0, w_jitter_range, w_jitter_range, h_jitter_range, h_jitter_range), 'constant', 0)
+        value = torch.zeros_like(norm_hidden_states_)
+        count = torch.zeros_like(norm_hidden_states_)
+        for index, view in enumerate(views):
+            h_start, h_end, w_start, w_end = view
+            local_states = norm_hidden_states_[:, h_start:h_end, w_start:w_end, :]
+            local_states = rearrange(local_states, 'bh h w d -> bh (h w) d')
+            local_output = self.attn1(
+                local_states,
+                encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
+                attention_mask=attention_mask,
+                **cross_attention_kwargs,
+            )
+            local_output = rearrange(local_output, 'bh (h w) d -> bh h w d', h = int(sub_h / scale_factor))
+
+            value[:, h_start:h_end, w_start:w_end, :] += local_output * 1
+            count[:, h_start:h_end, w_start:w_end, :] += 1
+
+        value = value[:, h_jitter_range:-h_jitter_range, w_jitter_range:-w_jitter_range, :]
+        count = count[:, h_jitter_range:-h_jitter_range, w_jitter_range:-w_jitter_range, :]
+        attn_output = torch.where(count>0, value/count, value)
+        
+        gaussian_local = gaussian_filter(attn_output, kernel_size=(2*current_scale_num-1), sigma=1.0)
+
+        value = torch.zeros_like(norm_hidden_states)
+        count = torch.zeros_like(norm_hidden_states)
+        for index, global_view in enumerate(global_views):
+            h, w = global_view
+            global_states = norm_hidden_states[:, h::current_scale_num_h, w::current_scale_num_w, :]
+            global_states = rearrange(global_states, 'bh h w d -> bh (h w) d')
+            global_output = self.attn1(
+                global_states,
+                encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
+                attention_mask=attention_mask,
+                **cross_attention_kwargs,
+            )
+            global_output = rearrange(global_output, 'bh (h w) d -> bh h w d', h = int(global_output.shape[1] ** 0.5))
+
+            value[:, h::current_scale_num_h, w::current_scale_num_w, :] += global_output * 1
+            count[:, h::current_scale_num_h, w::current_scale_num_w, :] += 1
+
+        attn_output_global = torch.where(count>0, value/count, value)
+
+        gaussian_global = gaussian_filter(attn_output_global, kernel_size=(2*current_scale_num-1), sigma=1.0)
+
+        attn_output = gaussian_local + (attn_output_global - gaussian_global)
+        attn_output = rearrange(attn_output, 'bh h w d -> bh (h w) d')
+        
+    else:
+        attn_output = self.attn1(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
+            attention_mask=attention_mask,
+            **cross_attention_kwargs,
+        )
+
+    if self.use_ada_layer_norm_zero:
+        attn_output = gate_msa.unsqueeze(1) * attn_output
+    hidden_states = attn_output + hidden_states
+
+    # 2.5 GLIGEN Control
+    if gligen_kwargs is not None:
+        hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"])
+    # 2.5 ends
+
+    # 3. Cross-Attention
+    if self.attn2 is not None:
+        norm_hidden_states = (
+            self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(hidden_states)
+        )
+        attn_output = self.attn2(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states,
+            attention_mask=encoder_attention_mask,
+            **cross_attention_kwargs,
+        )
+        hidden_states = attn_output + hidden_states
+
+    # 4. Feed-forward
+    norm_hidden_states = self.norm3(hidden_states)
+
+    if self.use_ada_layer_norm_zero:
+        norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+
+    if self._chunk_size is not None:
+        # "feed_forward_chunk_size" can be used to save memory
+        if norm_hidden_states.shape[self._chunk_dim] % self._chunk_size != 0:
+            raise ValueError(
+                f"`hidden_states` dimension to be chunked: {norm_hidden_states.shape[self._chunk_dim]} has to be divisible by chunk size: {self._chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`."
+            )
+
+        num_chunks = norm_hidden_states.shape[self._chunk_dim] // self._chunk_size
+        ff_output = torch.cat(
+            [
+                self.ff(hid_slice)
+                for hid_slice in norm_hidden_states.chunk(num_chunks, dim=self._chunk_dim)
+            ],
+            dim=self._chunk_dim,
+        )
+    else:
+        ff_output = self.ff(norm_hidden_states)
+
+    if self.use_ada_layer_norm_zero:
+        ff_output = gate_mlp.unsqueeze(1) * ff_output
+
+    hidden_states = ff_output + hidden_states
+
+    return hidden_states
+
+def ori_forward(
+    self,
+    hidden_states: torch.FloatTensor,
+    attention_mask: Optional[torch.FloatTensor] = None,
+    encoder_hidden_states: Optional[torch.FloatTensor] = None,
+    encoder_attention_mask: Optional[torch.FloatTensor] = None,
+    timestep: Optional[torch.LongTensor] = None,
+    cross_attention_kwargs: Dict[str, Any] = None,
+    class_labels: Optional[torch.LongTensor] = None,
+):
+    # Notice that normalization is always applied before the real computation in the following blocks.
+    # 0. Self-Attention
+    if self.use_ada_layer_norm:
+        norm_hidden_states = self.norm1(hidden_states, timestep)
+    elif self.use_ada_layer_norm_zero:
+        norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
+            hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
+        )
+    else:
+        norm_hidden_states = self.norm1(hidden_states)
+
+    # 2. Prepare GLIGEN inputs
+    cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
+    gligen_kwargs = cross_attention_kwargs.pop("gligen", None)
+
+    attn_output = self.attn1(
+        norm_hidden_states,
+        encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
+        attention_mask=attention_mask,
+        **cross_attention_kwargs,
+    )
+
+    if self.use_ada_layer_norm_zero:
+        attn_output = gate_msa.unsqueeze(1) * attn_output
+    hidden_states = attn_output + hidden_states
+
+    # 2.5 GLIGEN Control
+    if gligen_kwargs is not None:
+        hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"])
+    # 2.5 ends
+
+    # 3. Cross-Attention
+    if self.attn2 is not None:
+        norm_hidden_states = (
+            self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(hidden_states)
+        )
+        attn_output = self.attn2(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states,
+            attention_mask=encoder_attention_mask,
+            **cross_attention_kwargs,
+        )
+        hidden_states = attn_output + hidden_states
+
+    # 4. Feed-forward
+    norm_hidden_states = self.norm3(hidden_states)
+
+    if self.use_ada_layer_norm_zero:
+        norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+
+    if self._chunk_size is not None:
+        # "feed_forward_chunk_size" can be used to save memory
+        if norm_hidden_states.shape[self._chunk_dim] % self._chunk_size != 0:
+            raise ValueError(
+                f"`hidden_states` dimension to be chunked: {norm_hidden_states.shape[self._chunk_dim]} has to be divisible by chunk size: {self._chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`."
+            )
+
+        num_chunks = norm_hidden_states.shape[self._chunk_dim] // self._chunk_size
+        ff_output = torch.cat(
+            [
+                self.ff(hid_slice)
+                for hid_slice in norm_hidden_states.chunk(num_chunks, dim=self._chunk_dim)
+            ],
+            dim=self._chunk_dim,
+        )
+    else:
+        ff_output = self.ff(norm_hidden_states)
+
+    if self.use_ada_layer_norm_zero:
+        ff_output = gate_mlp.unsqueeze(1) * ff_output
+
+    hidden_states = ff_output + hidden_states
+
+    return hidden_states