init demo

.gitignore

@@ -0,0 +1,160 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+#poetry.lock
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#pdm.lock
+#   pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
+#   in version control.
+#   https://pdm.fming.dev/#use-with-ide
+.pdm.toml
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# PyCharm
+#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#  and can be added to the global gitignore or merged into this file.  For a more nuclear
+#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
+#.idea/

README.md

@@ -1,13 +1,23 @@
----
-title: Oms Sdxl Lcm
-emoji: 🚀
-colorFrom: gray
-colorTo: indigo
-sdk: gradio
-sdk_version: 4.7.1
-app_file: app.py
-pinned: false
-license: openrail++
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# One More Step: A Versatile Plug-and-Play Module for Rectifying Diffusion Schedule Flaws and Enhancing Low-Frequency Controls
+
+One More Step (OMS) module was proposed in [One More Step: A Versatile Plug-and-Play Module for Rectifying Diffusion Schedule Flaws and Enhancing Low-Frequency Controls](https://github.com/mhh0318/OneMoreStep)
+by *Minghui Hu, Jianbin Zheng, Chuanxia Zheng, Tat-Jen Cham et al.*
+
+By incorporating **one minor, additional step** atop the existing sampling process, it can address inherent limitations in the diffusion schedule of current diffusion models.  Crucially, this augmentation does not necessitate alterations to the original parameters of the model. Furthermore, the OMS module enhances control over low-frequency elements, such as color, within the generated images.
+
+Our model is **versatile** and allowing for **seamless integration** with a broad spectrum of prevalent Stable Diffusion frameworks.  It demonstrates compatibility with community-favored tools and techniques, including LoRA, ControlNet, Adapter, and other foundational models, underscoring its utility and adaptability in diverse applications.
+
+## Usage
+
+OMS now is supported diffusers with a customized pipeline, as detailed in [github](https://github.com/mhh0318/OneMoreStep). To run the model, first install the latest version of `diffusers` (especially for `LCM` feature) as well as `accelerate` and `transformers`.
+
+```bash
+pip install --upgrade pip
+pip install --upgrade diffusers transformers accelerate
+```
+
+And then we clone the repo
+```bash
+git clone https://github.com/mhh0318/OneMoreStep.git
+cd OneMoreStep
+```

app.py

@@ -0,0 +1,103 @@
+import torch
+import gradio as gr
+from functools import partial
+
+from diffusers_patch import OMSPipeline
+
+
+def create_sdxl_lcm_lora_pipe(sd_pipe_name_or_path, oms_name_or_path, lora_name_or_path):
+    from diffusers import StableDiffusionXLPipeline, LCMScheduler
+    sd_pipe = StableDiffusionXLPipeline.from_pretrained(sd_pipe_name_or_path, torch_dtype=torch.float16, variant="fp16", add_watermarker=False).to('cuda')
+    print('successfully load pipe')
+    sd_scheduler = LCMScheduler.from_config(sd_pipe.scheduler.config)
+    sd_pipe.load_lora_weights(lora_name_or_path, variant="fp16")
+
+    pipe = OMSPipeline.from_pretrained(oms_name_or_path, sd_pipeline = sd_pipe, torch_dtype=torch.float16, variant="fp16", trust_remote_code=True, sd_scheduler=sd_scheduler)
+    pipe.to('cuda')
+    
+    return pipe
+        
+
+class GradioDemo:
+    def __init__(
+        self, 
+        sd_pipe_name_or_path = "stabilityai/stable-diffusion-xl-base-1.0", 
+        oms_name_or_path = 'h1t/oms_b_openclip_xl', 
+        lora_name_or_path = 'latent-consistency/lcm-lora-sdxl'
+    ):
+        self.pipe = create_sdxl_lcm_lora_pipe(sd_pipe_name_or_path, oms_name_or_path, lora_name_or_path)
+        
+    def _inference(
+            self,
+            prompt = None,
+            oms_prompt = None,
+            oms_guidance_scale = 1.0,
+            num_inference_steps = 4,
+            sd_pipe_guidance_scale = 1.0,
+            seed = 1024,
+        ):
+        pipe_kwargs = dict(
+            prompt = prompt,
+            num_inference_steps = num_inference_steps,
+            guidance_scale = sd_pipe_guidance_scale,
+        )
+
+        generator = torch.Generator(device=self.pipe.device).manual_seed(seed)
+        pipe_kwargs.update(oms_flag=False)
+        print(f'raw kwargs: {pipe_kwargs}')
+        image_raw = self.pipe(
+            **pipe_kwargs,
+            generator=generator
+        )['images'][0]
+        
+        generator = torch.Generator(device=self.pipe.device).manual_seed(seed)
+        pipe_kwargs.update(oms_flag=True, oms_prompt=oms_prompt, oms_guidance_scale=1.0)
+        print(f'w/ oms wo/ cfg kwargs: {pipe_kwargs}')
+        image_oms = self.pipe(
+            **pipe_kwargs,
+            generator=generator
+        )['images'][0]
+        
+        oms_guidance_flag = oms_guidance_scale != 1.0
+        if oms_guidance_flag:
+            generator = torch.Generator(device=self.pipe.device).manual_seed(seed)
+            pipe_kwargs.update(oms_guidance_scale=oms_guidance_scale)
+            print(f'w/ oms +cfg kwargs: {pipe_kwargs}')
+            image_oms_cfg = self.pipe(
+                **pipe_kwargs,
+                generator=generator
+            )['images'][0]
+        else:
+            image_oms_cfg = None
+        
+        return image_raw, image_oms, image_oms_cfg, gr.update(visible=oms_guidance_flag)
+    
+    def mainloop(self):
+        with gr.Blocks() as demo:
+            gr.Markdown("# One More Step Demo")
+            with gr.Row():
+                with gr.Column():
+                    prompt = gr.Textbox(label="Prompt", value="a cat")
+                    oms_prompt = gr.Textbox(label="OMS Prompt", value="orange car")
+                    oms_guidance_scale = gr.Slider(label="OMS Guidance Scale", minimum=1.0, maximum=5.0, value=1.5, step=0.1)
+                    run_button = gr.Button(value="Generate images")
+                    with gr.Accordion("Advanced options", open=False):
+                        num_steps = gr.Slider(label="Steps", minimum=1, maximum=100, value=4, step=1)
+                        sd_guidance_scale = gr.Slider(label="SD Pipe Guidance Scale", minimum=0.1, maximum=30.0, value=1.0, step=0.1)
+                        seed = gr.Slider(label="Seed", minimum=-1, maximum=2147483647, step=1, randomize=False, value=1024)
+                with gr.Column():
+                    output_raw = gr.Image(label="SDXL w/ LCM-LoRA w/o OMS ")
+                    output_oms = gr.Image(label="w/ OMS w/o OMS CFG")
+                    with gr.Column(visible=False) as oms_cfg_wd:
+                        output_oms_cfg = gr.Image(label=f"w/ OMS w/ OMS CFG")
+
+            ips = [prompt, oms_prompt, oms_guidance_scale, num_steps, sd_guidance_scale, seed]
+            run_button.click(fn=self._inference, inputs=ips, outputs=[output_raw, output_oms, output_oms_cfg, oms_cfg_wd])
+
+            demo.queue(max_size=20)
+            demo.launch()
+
+
+if __name__ == "__main__":
+    gradio_demo = GradioDemo()
+    gradio_demo.mainloop()

diffusers_patch/__init__.py

@@ -0,0 +1 @@
+from .pipelines.oms import OMSPipeline

diffusers_patch/models/unet_2d_condition_woct.py

@@ -0,0 +1,756 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.loaders import UNet2DConditionLoadersMixin
+from diffusers.utils import BaseOutput, logging
+from diffusers.models.activations import get_activation
+from diffusers.models.attention_processor import AttentionProcessor, AttnProcessor
+from diffusers.models.embeddings import (
+    GaussianFourierProjection,
+    ImageHintTimeEmbedding,
+    ImageProjection,
+    ImageTimeEmbedding,
+    TextImageProjection,
+    TextImageTimeEmbedding,
+    TextTimeEmbedding,
+    TimestepEmbedding,
+    Timesteps,
+)
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.unet_2d_blocks import (
+    CrossAttnDownBlock2D,
+    CrossAttnUpBlock2D,
+    DownBlock2D,
+    UNetMidBlock2DCrossAttn,
+    UNetMidBlock2DSimpleCrossAttn,
+    UpBlock2D,
+    get_down_block,
+    get_up_block,
+)
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class UNet2DConditionOutput(BaseOutput):
+    """
+    The output of [`UNet2DConditionModel`].
+
+    Args:
+        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model.
+    """
+
+    sample: torch.FloatTensor = None
+
+
+class UNet2DConditionWoCTModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
+    r"""
+    A conditional 2D UNet model that takes a noisy sample, conditional state, but w/o a timestep and returns a sample
+    shaped output.
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+    for all models (such as downloading or saving).
+
+    Parameters:
+        sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
+            Height and width of input/output sample.
+        in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample.
+        out_channels (`int`, *optional*, defaults to 4): Number of channels in the output.
+        center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
+        down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
+            The tuple of downsample blocks to use.
+        mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`):
+            Block type for middle of UNet, it can be either `UNetMidBlock2DCrossAttn` or
+            `UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped.
+        up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`):
+            The tuple of upsample blocks to use.
+        only_cross_attention(`bool` or `Tuple[bool]`, *optional*, default to `False`):
+            Whether to include self-attention in the basic transformer blocks, see
+            [`~models.attention.BasicTransformerBlock`].
+        block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
+            The tuple of output channels for each block.
+        layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
+        downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution.
+        mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block.
+        act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
+        norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
+            If `None`, normalization and activation layers is skipped in post-processing.
+        norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
+        cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
+            The dimension of the cross attention features.
+        transformer_layers_per_block (`int` or `Tuple[int]`, *optional*, defaults to 1):
+            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
+            [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+            [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+        encoder_hid_dim (`int`, *optional*, defaults to None):
+            If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
+            dimension to `cross_attention_dim`.
+        encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
+            If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text
+            embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
+        attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
+        num_attention_heads (`int`, *optional*):
+            The number of attention heads. If not defined, defaults to `attention_head_dim`
+        conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer.
+        conv_out_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_out` layer.
+        mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`):
+            Whether to use cross attention with the mid block when using the `UNetMidBlock2DSimpleCrossAttn`. If
+            `only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is `None`, the
+            `only_cross_attention` value is used as the value for `mid_block_only_cross_attention`. Default to `False`
+            otherwise.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        sample_size: Optional[int] = None,
+        in_channels: int = 4,
+        out_channels: int = 4,
+        center_input_sample: bool = False,
+        down_block_types: Tuple[str] = (
+            "CrossAttnDownBlock2D",
+            "CrossAttnDownBlock2D",
+            "CrossAttnDownBlock2D",
+            "DownBlock2D",
+        ),
+        mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
+        up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"),
+        only_cross_attention: Union[bool, Tuple[bool]] = False,
+        block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
+        layers_per_block: Union[int, Tuple[int]] = 2,
+        downsample_padding: int = 1,
+        mid_block_scale_factor: float = 1,
+        act_fn: str = "silu",
+        norm_num_groups: Optional[int] = 32,
+        norm_eps: float = 1e-5,
+        cross_attention_dim: Union[int, Tuple[int]] = 1280,
+        transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+        encoder_hid_dim: Optional[int] = None,
+        encoder_hid_dim_type: Optional[str] = None,
+        attention_head_dim: Union[int, Tuple[int]] = 8,
+        num_attention_heads: Optional[Union[int, Tuple[int]]] = None,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        upcast_attention: bool = False,
+        resnet_out_scale_factor: int = 1.0,
+        conv_in_kernel: int = 3,
+        conv_out_kernel: int = 3,
+        mid_block_only_cross_attention: Optional[bool] = None,
+        cross_attention_norm: Optional[str] = None,
+    ):
+        super().__init__()
+
+        self.sample_size = sample_size
+
+        if num_attention_heads is not None:
+            raise ValueError(
+                "At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19."
+            )
+
+        # If `num_attention_heads` is not defined (which is the case for most models)
+        # it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
+        # The reason for this behavior is to correct for incorrectly named variables that were introduced
+        # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
+        # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
+        # which is why we correct for the naming here.
+        num_attention_heads = num_attention_heads or attention_head_dim
+
+        # Check inputs
+        if len(down_block_types) != len(up_block_types):
+            raise ValueError(
+                f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
+            )
+
+        if len(block_out_channels) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}."
+            )
+
+        if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(layers_per_block, int) and len(layers_per_block) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}."
+            )
+
+        # input
+        conv_in_padding = (conv_in_kernel - 1) // 2
+        self.conv_in = nn.Conv2d(
+            in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding
+        )
+
+        if encoder_hid_dim_type is None and encoder_hid_dim is not None:
+            encoder_hid_dim_type = "text_proj"
+            self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
+            logger.info("encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined.")
+
+        if encoder_hid_dim is None and encoder_hid_dim_type is not None:
+            raise ValueError(
+                f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}."
+            )
+
+        if encoder_hid_dim_type == "text_proj":
+            self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim)
+        elif encoder_hid_dim_type == "text_image_proj":
+            # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+            # case when `addition_embed_type == "text_image_proj"` (Kadinsky 2.1)`
+            self.encoder_hid_proj = TextImageProjection(
+                text_embed_dim=encoder_hid_dim,
+                image_embed_dim=cross_attention_dim,
+                cross_attention_dim=cross_attention_dim,
+            )
+        elif encoder_hid_dim_type == "image_proj":
+            # Kandinsky 2.2
+            self.encoder_hid_proj = ImageProjection(
+                image_embed_dim=encoder_hid_dim,
+                cross_attention_dim=cross_attention_dim,
+            )
+        elif encoder_hid_dim_type is not None:
+            raise ValueError(
+                f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'."
+            )
+        else:
+            self.encoder_hid_proj = None
+
+        self.down_blocks = nn.ModuleList([])
+        self.up_blocks = nn.ModuleList([])
+
+        if isinstance(only_cross_attention, bool):
+            if mid_block_only_cross_attention is None:
+                mid_block_only_cross_attention = only_cross_attention
+
+            only_cross_attention = [only_cross_attention] * len(down_block_types)
+
+        if mid_block_only_cross_attention is None:
+            mid_block_only_cross_attention = False
+
+        if isinstance(num_attention_heads, int):
+            num_attention_heads = (num_attention_heads,) * len(down_block_types)
+
+        if isinstance(attention_head_dim, int):
+            attention_head_dim = (attention_head_dim,) * len(down_block_types)
+
+        if isinstance(cross_attention_dim, int):
+            cross_attention_dim = (cross_attention_dim,) * len(down_block_types)
+
+        if isinstance(layers_per_block, int):
+            layers_per_block = [layers_per_block] * len(down_block_types)
+
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types)
+
+        # disable time cond
+        time_embed_dim = None
+        blocks_time_embed_dim = time_embed_dim
+        resnet_time_scale_shift = None
+        resnet_skip_time_act = False
+
+        # down
+        output_channel = block_out_channels[0]
+        for i, down_block_type in enumerate(down_block_types):
+            input_channel = output_channel
+            output_channel = block_out_channels[i]
+            is_final_block = i == len(block_out_channels) - 1
+
+            down_block = get_down_block(
+                down_block_type,
+                num_layers=layers_per_block[i],
+                transformer_layers_per_block=transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                temb_channels=blocks_time_embed_dim,
+                add_downsample=not is_final_block,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=cross_attention_dim[i],
+                num_attention_heads=num_attention_heads[i],
+                downsample_padding=downsample_padding,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                resnet_skip_time_act=resnet_skip_time_act,
+                resnet_out_scale_factor=resnet_out_scale_factor,
+                cross_attention_norm=cross_attention_norm,
+                attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
+            )
+            self.down_blocks.append(down_block)
+
+        # mid
+        if mid_block_type == "UNetMidBlock2DCrossAttn":
+            self.mid_block = UNetMidBlock2DCrossAttn(
+                transformer_layers_per_block=transformer_layers_per_block[-1],
+                in_channels=block_out_channels[-1],
+                temb_channels=blocks_time_embed_dim,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                output_scale_factor=mid_block_scale_factor,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                cross_attention_dim=cross_attention_dim[-1],
+                num_attention_heads=num_attention_heads[-1],
+                resnet_groups=norm_num_groups,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                upcast_attention=upcast_attention,
+            )
+        elif mid_block_type == "UNetMidBlock2DSimpleCrossAttn":
+            self.mid_block = UNetMidBlock2DSimpleCrossAttn(
+                in_channels=block_out_channels[-1],
+                temb_channels=blocks_time_embed_dim,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                output_scale_factor=mid_block_scale_factor,
+                cross_attention_dim=cross_attention_dim[-1],
+                attention_head_dim=attention_head_dim[-1],
+                resnet_groups=norm_num_groups,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                skip_time_act=resnet_skip_time_act,
+                only_cross_attention=mid_block_only_cross_attention,
+                cross_attention_norm=cross_attention_norm,
+            )
+        elif mid_block_type is None:
+            self.mid_block = None
+        else:
+            raise ValueError(f"unknown mid_block_type : {mid_block_type}")
+
+        # count how many layers upsample the images
+        self.num_upsamplers = 0
+
+        # up
+        reversed_block_out_channels = list(reversed(block_out_channels))
+        reversed_num_attention_heads = list(reversed(num_attention_heads))
+        reversed_layers_per_block = list(reversed(layers_per_block))
+        reversed_cross_attention_dim = list(reversed(cross_attention_dim))
+        reversed_transformer_layers_per_block = list(reversed(transformer_layers_per_block))
+        only_cross_attention = list(reversed(only_cross_attention))
+
+        output_channel = reversed_block_out_channels[0]
+        for i, up_block_type in enumerate(up_block_types):
+            is_final_block = i == len(block_out_channels) - 1
+
+            prev_output_channel = output_channel
+            output_channel = reversed_block_out_channels[i]
+            input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
+
+            # add upsample block for all BUT final layer
+            if not is_final_block:
+                add_upsample = True
+                self.num_upsamplers += 1
+            else:
+                add_upsample = False
+
+            up_block = get_up_block(
+                up_block_type,
+                num_layers=reversed_layers_per_block[i] + 1,
+                transformer_layers_per_block=reversed_transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                prev_output_channel=prev_output_channel,
+                temb_channels=blocks_time_embed_dim,
+                add_upsample=add_upsample,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=reversed_cross_attention_dim[i],
+                num_attention_heads=reversed_num_attention_heads[i],
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                resnet_skip_time_act=resnet_skip_time_act,
+                resnet_out_scale_factor=resnet_out_scale_factor,
+                cross_attention_norm=cross_attention_norm,
+                attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
+            )
+            self.up_blocks.append(up_block)
+            prev_output_channel = output_channel
+
+        # out
+        if norm_num_groups is not None:
+            self.conv_norm_out = nn.GroupNorm(
+                num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps
+            )
+
+            self.conv_act = get_activation(act_fn)
+
+        else:
+            self.conv_norm_out = None
+            self.conv_act = None
+
+        conv_out_padding = (conv_out_kernel - 1) // 2
+        self.conv_out = nn.Conv2d(
+            block_out_channels[0], out_channels, kernel_size=conv_out_kernel, padding=conv_out_padding
+        )
+
+    @property
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+            if hasattr(module, "set_processor"):
+                processors[f"{name}.processor"] = module.processor
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"))
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    def set_default_attn_processor(self):
+        """
+        Disables custom attention processors and sets the default attention implementation.
+        """
+        self.set_attn_processor(AttnProcessor())
+
+    def set_attention_slice(self, slice_size):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module splits the input tensor in slices to compute attention in
+        several steps. This is useful for saving some memory in exchange for a small decrease in speed.
+
+        Args:
+            slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
+                When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
+                `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
+                provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+                must be a multiple of `slice_size`.
+        """
+        sliceable_head_dims = []
+
+        def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
+            if hasattr(module, "set_attention_slice"):
+                sliceable_head_dims.append(module.sliceable_head_dim)
+
+            for child in module.children():
+                fn_recursive_retrieve_sliceable_dims(child)
+
+        # retrieve number of attention layers
+        for module in self.children():
+            fn_recursive_retrieve_sliceable_dims(module)
+
+        num_sliceable_layers = len(sliceable_head_dims)
+
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = [dim // 2 for dim in sliceable_head_dims]
+        elif slice_size == "max":
+            # make smallest slice possible
+            slice_size = num_sliceable_layers * [1]
+
+        slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size
+
+        if len(slice_size) != len(sliceable_head_dims):
+            raise ValueError(
+                f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
+                f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
+            )
+
+        for i in range(len(slice_size)):
+            size = slice_size[i]
+            dim = sliceable_head_dims[i]
+            if size is not None and size > dim:
+                raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
+
+        # Recursively walk through all the children.
+        # Any children which exposes the set_attention_slice method
+        # gets the message
+        def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]):
+            if hasattr(module, "set_attention_slice"):
+                module.set_attention_slice(slice_size.pop())
+
+            for child in module.children():
+                fn_recursive_set_attention_slice(child, slice_size)
+
+        reversed_slice_size = list(reversed(slice_size))
+        for module in self.children():
+            fn_recursive_set_attention_slice(module, reversed_slice_size)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, (CrossAttnDownBlock2D, DownBlock2D, CrossAttnUpBlock2D, UpBlock2D)):
+            module.gradient_checkpointing = value
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        encoder_hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+        down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+        mid_block_additional_residual: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ) -> Union[UNet2DConditionOutput, Tuple]:
+        r"""
+        The [`UNet2DConditionModel`] forward method.
+
+        Args:
+            sample (`torch.FloatTensor`):
+                The noisy input tensor with the following shape `(batch, channel, height, width)`.
+            encoder_hidden_states (`torch.FloatTensor`):
+                The encoder hidden states with shape `(batch, sequence_length, feature_dim)`.
+            encoder_attention_mask (`torch.Tensor`):
+                A cross-attention mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If
+                `True` the mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias,
+                which adds large negative values to the attention scores corresponding to "discard" tokens.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+                tuple.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the [`AttnProcessor`].
+            added_cond_kwargs: (`dict`, *optional*):
+                A kwargs dictionary containin additional embeddings that if specified are added to the embeddings that
+                are passed along to the UNet blocks.
+
+        Returns:
+            [`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
+                If `return_dict` is True, an [`~models.unet_2d_condition.UNet2DConditionOutput`] is returned, otherwise
+                a `tuple` is returned where the first element is the sample tensor.
+        """
+        # By default samples have to be AT least a multiple of the overall upsampling factor.
+        # The overall upsampling factor is equal to 2 ** (# num of upsampling layers).
+        # However, the upsampling interpolation output size can be forced to fit any upsampling size
+        # on the fly if necessary.
+        default_overall_up_factor = 2**self.num_upsamplers
+
+        # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
+        forward_upsample_size = False
+        upsample_size = None
+
+        if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]):
+            logger.info("Forward upsample size to force interpolation output size.")
+            forward_upsample_size = True
+
+        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension
+        # expects mask of shape:
+        #   [batch, key_tokens]
+        # adds singleton query_tokens dimension:
+        #   [batch,                    1, key_tokens]
+        # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+        #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+        #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+        if attention_mask is not None:
+            # assume that mask is expressed as:
+            #   (1 = keep,      0 = discard)
+            # convert mask into a bias that can be added to attention scores:
+            #       (keep = +0,     discard = -10000.0)
+            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None:
+            encoder_attention_mask = (1 - encoder_attention_mask.to(sample.dtype)) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # 0. center input if necessary
+        if self.config.center_input_sample:
+            sample = 2 * sample - 1.0
+
+        # 1. time (skip)
+        emb = None
+
+        if self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_proj":
+            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states)
+        elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_image_proj":
+            # Kadinsky 2.1 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states, image_embeds)
+        elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "image_proj":
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            encoder_hidden_states = self.encoder_hid_proj(image_embeds)
+        # 2. pre-process
+        sample = self.conv_in(sample)
+
+        # 3. down
+
+        is_controlnet = mid_block_additional_residual is not None and down_block_additional_residuals is not None
+        is_adapter = mid_block_additional_residual is None and down_block_additional_residuals is not None
+
+        down_block_res_samples = (sample,)
+        for downsample_block in self.down_blocks:
+            if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
+                # For t2i-adapter CrossAttnDownBlock2D
+                additional_residuals = {}
+                if is_adapter and len(down_block_additional_residuals) > 0:
+                    additional_residuals["additional_residuals"] = down_block_additional_residuals.pop(0)
+
+                sample, res_samples = downsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    encoder_attention_mask=encoder_attention_mask,
+                    **additional_residuals,
+                )
+            else:
+                sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
+
+                if is_adapter and len(down_block_additional_residuals) > 0:
+                    sample += down_block_additional_residuals.pop(0)
+
+            down_block_res_samples += res_samples
+
+        if is_controlnet:
+            new_down_block_res_samples = ()
+
+            for down_block_res_sample, down_block_additional_residual in zip(
+                down_block_res_samples, down_block_additional_residuals
+            ):
+                down_block_res_sample = down_block_res_sample + down_block_additional_residual
+                new_down_block_res_samples = new_down_block_res_samples + (down_block_res_sample,)
+
+            down_block_res_samples = new_down_block_res_samples
+
+        # 4. mid
+        if self.mid_block is not None:
+            sample = self.mid_block(
+                sample,
+                emb,
+                encoder_hidden_states=encoder_hidden_states,
+                attention_mask=attention_mask,
+                cross_attention_kwargs=cross_attention_kwargs,
+                encoder_attention_mask=encoder_attention_mask,
+            )
+
+        if is_controlnet:
+            sample = sample + mid_block_additional_residual
+
+        # 5. up
+        for i, upsample_block in enumerate(self.up_blocks):
+            is_final_block = i == len(self.up_blocks) - 1
+
+            res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+            down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
+
+            # if we have not reached the final block and need to forward the
+            # upsample size, we do it here
+            if not is_final_block and forward_upsample_size:
+                upsample_size = down_block_res_samples[-1].shape[2:]
+
+            if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    upsample_size=upsample_size,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                )
+            else:
+                sample = upsample_block(
+                    hidden_states=sample, temb=emb, res_hidden_states_tuple=res_samples, upsample_size=upsample_size
+                )
+
+        # 6. post-process
+        if self.conv_norm_out:
+            sample = self.conv_norm_out(sample)
+            sample = self.conv_act(sample)
+        sample = self.conv_out(sample)
+
+        if not return_dict:
+            return (sample,)
+
+        return UNet2DConditionOutput(sample=sample)

diffusers_patch/pipelines/oms/__init__.py

@@ -0,0 +1 @@
+from .pipeline_oms import OMSPipeline

diffusers_patch/pipelines/oms/pipeline_oms.py

@@ -0,0 +1,655 @@
+import json
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import torch
+from transformers import CLIPTextModel, CLIPTokenizer
+
+from diffusers.loaders import FromSingleFileMixin
+
+from diffusers.utils import (
+    USE_PEFT_BACKEND,
+    deprecate,
+    logging,
+)
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.pipelines.pipeline_utils import *
+from diffusers.pipelines.pipeline_utils import _get_pipeline_class
+from diffusers.models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT
+
+from diffusers_patch.models.unet_2d_condition_woct import UNet2DConditionWoCTModel
+
+from diffusers_patch.pipelines.oms.utils import SDXLTextEncoder, SDXLTokenizer
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+def load_sub_model_oms(
+    library_name: str,
+    class_name: str,
+    importable_classes: List[Any],
+    pipelines: Any,
+    is_pipeline_module: bool,
+    pipeline_class: Any,
+    torch_dtype: torch.dtype,
+    provider: Any,
+    sess_options: Any,
+    device_map: Optional[Union[Dict[str, torch.device], str]],
+    max_memory: Optional[Dict[Union[int, str], Union[int, str]]],
+    offload_folder: Optional[Union[str, os.PathLike]],
+    offload_state_dict: bool,
+    model_variants: Dict[str, str],
+    name: str,
+    from_flax: bool,
+    variant: str,
+    low_cpu_mem_usage: bool,
+    cached_folder: Union[str, os.PathLike],
+):
+    """Helper method to load the module `name` from `library_name` and `class_name`"""
+    # retrieve class candidates
+    class_obj, class_candidates = get_class_obj_and_candidates(
+        library_name,
+        class_name,
+        importable_classes,
+        pipelines,
+        is_pipeline_module,
+        component_name=name,
+        cache_dir=cached_folder,
+    )
+
+    load_method_name = None
+    # retrive load method name
+    for class_name, class_candidate in class_candidates.items():
+        if class_candidate is not None and issubclass(class_obj, class_candidate):
+            load_method_name = importable_classes[class_name][1]
+
+    # if load method name is None, then we have a dummy module -> raise Error
+    if load_method_name is None:
+        none_module = class_obj.__module__
+        is_dummy_path = none_module.startswith(DUMMY_MODULES_FOLDER) or none_module.startswith(
+            TRANSFORMERS_DUMMY_MODULES_FOLDER
+        )
+        if is_dummy_path and "dummy" in none_module:
+            # call class_obj for nice error message of missing requirements
+            class_obj()
+
+        raise ValueError(
+            f"The component {class_obj} of {pipeline_class} cannot be loaded as it does not seem to have"
+            f" any of the loading methods defined in {ALL_IMPORTABLE_CLASSES}."
+        )
+
+    load_method = getattr(class_obj, load_method_name)
+
+    # add kwargs to loading method
+    import diffusers
+    loading_kwargs = {}
+    if issubclass(class_obj, torch.nn.Module):
+        loading_kwargs["torch_dtype"] = torch_dtype
+    if issubclass(class_obj, diffusers.OnnxRuntimeModel):
+        loading_kwargs["provider"] = provider
+        loading_kwargs["sess_options"] = sess_options
+
+    is_diffusers_model = issubclass(class_obj, diffusers.ModelMixin)
+
+    if is_transformers_available():
+        transformers_version = version.parse(version.parse(transformers.__version__).base_version)
+    else:
+        transformers_version = "N/A"
+
+    is_transformers_model = (
+        is_transformers_available()
+        and issubclass(class_obj, PreTrainedModel)
+        and transformers_version >= version.parse("4.20.0")
+    )
+
+    # When loading a transformers model, if the device_map is None, the weights will be initialized as opposed to diffusers.
+    # To make default loading faster we set the `low_cpu_mem_usage=low_cpu_mem_usage` flag which is `True` by default.
+    # This makes sure that the weights won't be initialized which significantly speeds up loading.
+    if is_diffusers_model or is_transformers_model:
+        loading_kwargs["device_map"] = device_map
+        loading_kwargs["max_memory"] = max_memory
+        loading_kwargs["offload_folder"] = offload_folder
+        loading_kwargs["offload_state_dict"] = offload_state_dict
+        loading_kwargs["variant"] = model_variants.pop(name, None)
+        if from_flax:
+            loading_kwargs["from_flax"] = True
+
+        # the following can be deleted once the minimum required `transformers` version
+        # is higher than 4.27
+        if (
+            is_transformers_model
+            and loading_kwargs["variant"] is not None
+            and transformers_version < version.parse("4.27.0")
+        ):
+            raise ImportError(
+                f"When passing `variant='{variant}'`, please make sure to upgrade your `transformers` version to at least 4.27.0.dev0"
+            )
+        elif is_transformers_model and loading_kwargs["variant"] is None:
+            loading_kwargs.pop("variant")
+
+        # if `from_flax` and model is transformer model, can currently not load with `low_cpu_mem_usage`
+        if not (from_flax and is_transformers_model):
+            loading_kwargs["low_cpu_mem_usage"] = low_cpu_mem_usage
+        else:
+            loading_kwargs["low_cpu_mem_usage"] = False
+    # check if oms directory
+    if 'oms' in name:
+        config_name = os.path.join(cached_folder, name, 'config.json')
+        with open(config_name, "r", encoding="utf-8") as f:
+            index = json.load(f)
+        file_path_or_name = index['_name_or_path']
+        if 'SDXL' in index.get('_class_name', 'CLIP'):
+            loaded_sub_model = load_method(file_path_or_name, **loading_kwargs)
+        elif 'subfolder' in index.keys():
+            loading_kwargs["subfolder"] = index["subfolder"]
+            loaded_sub_model = load_method(file_path_or_name, **loading_kwargs)
+    else:
+        # check if the module is in a subdirectory
+        if os.path.isdir(os.path.join(cached_folder, name)):
+            loaded_sub_model = load_method(os.path.join(cached_folder, name), **loading_kwargs)
+        else:
+            # else load from the root directory
+            loaded_sub_model = load_method(cached_folder, **loading_kwargs)
+
+    return loaded_sub_model
+
+class OMSPipeline(DiffusionPipeline, FromSingleFileMixin):
+
+    def __init__(
+        self,
+        oms_module: UNet2DConditionWoCTModel,
+        sd_pipeline: DiffusionPipeline,
+        oms_text_encoder:Optional[Union[CLIPTextModel, SDXLTextEncoder]],
+        oms_tokenizer:Optional[Union[CLIPTokenizer, SDXLTokenizer]],
+        sd_scheduler = None
+    ):
+        # assert sd_pipeline is not None
+
+        if oms_tokenizer is None:
+            oms_tokenizer = sd_pipeline.tokenizer
+        if oms_text_encoder is None:
+            oms_text_encoder = sd_pipeline.text_encoder
+
+        # For OMS with SDXL text encoders 
+        if 'SDXL' in oms_text_encoder.__class__.__name__:
+            self.is_dual_text_encoder = True
+        else:
+            self.is_dual_text_encoder = False 
+
+        self.register_modules(
+            oms_module=oms_module,
+            oms_text_encoder=oms_text_encoder,
+            oms_tokenizer=oms_tokenizer,
+            sd_pipeline = sd_pipeline
+        )
+
+        if sd_scheduler is None:
+            self.scheduler = sd_pipeline.scheduler 
+        else:
+            self.scheduler = sd_scheduler
+            sd_pipeline.scheduler = sd_scheduler
+
+        self.vae_scale_factor = 2 ** (len(sd_pipeline.vae.config.block_out_channels) - 1)
+        self.default_sample_size = sd_pipeline.unet.config.sample_size
+
+    # 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 oms_step(self, predict_v, latents, do_classifier_free_guidance_for_oms, oms_guidance_scale, generator, alpha_prod_t_prev):
+        if do_classifier_free_guidance_for_oms:
+            pred_uncond, pred_text = predict_v.chunk(2)
+            predict_v = pred_uncond + oms_guidance_scale * (pred_text - pred_uncond)
+        # so fking dirty but keep it for now
+        alpha_prod_t = torch.zeros_like(alpha_prod_t_prev)
+        beta_prod_t = 1 - alpha_prod_t
+        beta_prod_t_prev = 1 - alpha_prod_t_prev
+        current_alpha_t = alpha_prod_t / alpha_prod_t_prev
+        current_beta_t = 1 - current_alpha_t
+        pred_original_sample = (alpha_prod_t**0.5) * latents - (beta_prod_t**0.5) * predict_v
+        # pred_original_sample = - predict_v
+        pred_original_sample_coeff = (alpha_prod_t_prev ** (0.5) * current_beta_t) / beta_prod_t
+        current_sample_coeff = current_alpha_t ** (0.5) * beta_prod_t_prev / beta_prod_t
+        pred_prev_sample = pred_original_sample_coeff * pred_original_sample + current_sample_coeff * latents
+
+        pred_prev_sample = pred_prev_sample
+        # TODO unit variance but seem dont need it
+
+        device = latents.device
+        variance_noise = randn_tensor(
+            latents.shape, generator=generator, device=device, dtype=latents.dtype
+        )
+        variance = (1 - alpha_prod_t_prev) / (1 - alpha_prod_t) * current_beta_t 
+        variance = torch.clamp(variance, min=1e-20) * variance_noise
+
+        latents = pred_prev_sample + variance
+        return latents
+
+    def oms_text_encode(self, prompt, num_images_per_prompt, device):
+        max_length = None if self.is_dual_text_encoder else self.oms_tokenizer.model_max_length
+        if self.is_dual_text_encoder:
+            tokenized_prompts = self.oms_tokenizer(prompt,
+                                    padding='max_length',
+                                    max_length=max_length,
+                                    truncation=True,
+                                    return_tensors='pt').input_ids
+            tokenized_prompts = torch.stack([tokenized_prompts[0], tokenized_prompts[1]], dim=1)
+            text_embeddings, _ = self.oms_text_encoder( [tokenized_prompts[:, 0, :].to(device), tokenized_prompts[:, 1, :].to(device)])  # type: ignore
+        elif 'clip' in self.oms_text_encoder.config_class.model_type:
+            tokenized_prompts = self.oms_tokenizer(prompt,
+                                               padding='max_length',
+                                               max_length=max_length,
+                                               truncation=True,
+                                               return_tensors='pt').input_ids
+            text_embeddings = self.oms_text_encoder(tokenized_prompts.to(device))[0]  # type: ignore
+        else: # T5
+            tokenized_prompts = self.oms_tokenizer(prompt,
+                                               padding='max_length',
+                                               max_length=max_length,
+                                               truncation=True,
+                                               add_special_tokens=True,
+                                               return_tensors='pt').input_ids
+            # Note: t5 text encoder outputs "None" under fp16
+            with torch.cuda.amp.autocast(dtype=torch.float32):
+                text_embeddings = self.text_encoder(tokenized_prompts.to(device))[0] 
+
+        # duplicate text embeddings for each generation per prompt
+        bs_embed, seq_len, _ = text_embeddings.shape
+        text_embeddings = text_embeddings.repeat(1, num_images_per_prompt, 1)  # type: ignore
+        text_embeddings = text_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+        return text_embeddings
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs):
+        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
+        resume_download = kwargs.pop("resume_download", False)
+        force_download = kwargs.pop("force_download", False)
+        proxies = kwargs.pop("proxies", None)
+        local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
+        use_auth_token = kwargs.pop("use_auth_token", None)
+        revision = kwargs.pop("revision", None)
+        from_flax = kwargs.pop("from_flax", False)
+        torch_dtype = kwargs.pop("torch_dtype", None)
+        custom_pipeline = kwargs.pop("custom_pipeline", None)
+        custom_revision = kwargs.pop("custom_revision", None)
+        provider = kwargs.pop("provider", None)
+        sess_options = kwargs.pop("sess_options", None)
+        device_map = kwargs.pop("device_map", None)
+        max_memory = kwargs.pop("max_memory", None)
+        offload_folder = kwargs.pop("offload_folder", None)
+        offload_state_dict = kwargs.pop("offload_state_dict", False)
+        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT)
+        variant = kwargs.pop("variant", None)
+        use_safetensors = kwargs.pop("use_safetensors", None)
+        load_connected_pipeline = kwargs.pop("load_connected_pipeline", False)
+
+        # 1. Download the checkpoints and configs
+        # use snapshot download here to get it working from from_pretrained
+        if not os.path.isdir(pretrained_model_name_or_path):
+            if pretrained_model_name_or_path.count("/") > 1:
+                raise ValueError(
+                    f'The provided pretrained_model_name_or_path "{pretrained_model_name_or_path}"'
+                    " is neither a valid local path nor a valid repo id. Please check the parameter."
+                )
+            cached_folder = cls.download(
+                pretrained_model_name_or_path,
+                cache_dir=cache_dir,
+                resume_download=resume_download,
+                force_download=force_download,
+                proxies=proxies,
+                local_files_only=local_files_only,
+                use_auth_token=use_auth_token,
+                revision=revision,
+                from_flax=from_flax,
+                use_safetensors=use_safetensors,
+                custom_pipeline=custom_pipeline,
+                custom_revision=custom_revision,
+                variant=variant,
+                load_connected_pipeline=load_connected_pipeline,
+                **kwargs,
+            )
+        else:
+            cached_folder = pretrained_model_name_or_path
+
+        config_dict = cls.load_config(cached_folder)
+
+        # pop out "_ignore_files" as it is only needed for download
+        config_dict.pop("_ignore_files", None)
+
+        # 2. Define which model components should load variants
+        # We retrieve the information by matching whether variant
+        # model checkpoints exist in the subfolders
+        model_variants = {}
+        if variant is not None:
+            for folder in os.listdir(cached_folder):
+                folder_path = os.path.join(cached_folder, folder)
+                is_folder = os.path.isdir(folder_path) and folder in config_dict
+                variant_exists = is_folder and any(
+                    p.split(".")[1].startswith(variant) for p in os.listdir(folder_path)
+                )
+                if variant_exists:
+                    model_variants[folder] = variant
+
+        # 3. Load the pipeline class, if using custom module then load it from the hub
+        # if we load from explicit class, let's use it
+        pipeline_class = _get_pipeline_class(
+            cls,
+            config_dict,
+            load_connected_pipeline=load_connected_pipeline,
+            custom_pipeline=custom_pipeline,
+            cache_dir=cache_dir,
+            revision=custom_revision,
+        )
+
+        # DEPRECATED: To be removed in 1.0.0
+        if pipeline_class.__name__ == "StableDiffusionInpaintPipeline" and version.parse(
+            version.parse(config_dict["_diffusers_version"]).base_version
+        ) <= version.parse("0.5.1"):
+            from diffusers import StableDiffusionInpaintPipeline, StableDiffusionInpaintPipelineLegacy
+
+            pipeline_class = StableDiffusionInpaintPipelineLegacy
+
+            deprecation_message = (
+                "You are using a legacy checkpoint for inpainting with Stable Diffusion, therefore we are loading the"
+                f" {StableDiffusionInpaintPipelineLegacy} class instead of {StableDiffusionInpaintPipeline}. For"
+                " better inpainting results, we strongly suggest using Stable Diffusion's official inpainting"
+                " checkpoint: https://huggingface.co/runwayml/stable-diffusion-inpainting instead or adapting your"
+                f" checkpoint {pretrained_model_name_or_path} to the format of"
+                " https://huggingface.co/runwayml/stable-diffusion-inpainting. Note that we do not actively maintain"
+                " the {StableDiffusionInpaintPipelineLegacy} class and will likely remove it in version 1.0.0."
+            )
+            deprecate("StableDiffusionInpaintPipelineLegacy", "1.0.0", deprecation_message, standard_warn=False)
+
+        # 4. Define expected modules given pipeline signature
+        # and define non-None initialized modules (=`init_kwargs`)
+
+        # some modules can be passed directly to the init
+        # in this case they are already instantiated in `kwargs`
+        # extract them here
+        expected_modules, optional_kwargs = cls._get_signature_keys(pipeline_class)
+        passed_class_obj = {k: kwargs.pop(k) for k in expected_modules if k in kwargs}
+        passed_pipe_kwargs = {k: kwargs.pop(k) for k in optional_kwargs if k in kwargs}
+
+        init_dict, unused_kwargs, _ = pipeline_class.extract_init_dict(config_dict, **kwargs)
+
+        # define init kwargs and make sure that optional component modules are filtered out
+        init_kwargs = {
+            k: init_dict.pop(k)
+            for k in optional_kwargs
+            if k in init_dict and k not in pipeline_class._optional_components
+        }
+        init_kwargs = {**init_kwargs, **passed_pipe_kwargs}
+
+        # remove `null` components
+        def load_module(name, value):
+            if value[0] is None:
+                return False
+            if name in passed_class_obj and passed_class_obj[name] is None:
+                return False
+            return True
+
+        init_dict = {k: v for k, v in init_dict.items() if load_module(k, v)}
+
+        # Special case: safety_checker must be loaded separately when using `from_flax`
+        if from_flax and "safety_checker" in init_dict and "safety_checker" not in passed_class_obj:
+            raise NotImplementedError(
+                "The safety checker cannot be automatically loaded when loading weights `from_flax`."
+                " Please, pass `safety_checker=None` to `from_pretrained`, and load the safety checker"
+                " separately if you need it."
+            )
+
+        # 5. Throw nice warnings / errors for fast accelerate loading
+        if len(unused_kwargs) > 0:
+            logger.warning(
+                f"Keyword arguments {unused_kwargs} are not expected by {pipeline_class.__name__} and will be ignored."
+            )
+
+        if low_cpu_mem_usage and not is_accelerate_available():
+            low_cpu_mem_usage = False
+            logger.warning(
+                "Cannot initialize model with low cpu memory usage because `accelerate` was not found in the"
+                " environment. Defaulting to `low_cpu_mem_usage=False`. It is strongly recommended to install"
+                " `accelerate` for faster and less memory-intense model loading. You can do so with: \n```\npip"
+                " install accelerate\n```\n."
+            )
+
+        if device_map is not None and not is_torch_version(">=", "1.9.0"):
+            raise NotImplementedError(
+                "Loading and dispatching requires torch >= 1.9.0. Please either update your PyTorch version or set"
+                " `device_map=None`."
+            )
+
+        if low_cpu_mem_usage is True and not is_torch_version(">=", "1.9.0"):
+            raise NotImplementedError(
+                "Low memory initialization requires torch >= 1.9.0. Please either update your PyTorch version or set"
+                " `low_cpu_mem_usage=False`."
+            )
+
+        if low_cpu_mem_usage is False and device_map is not None:
+            raise ValueError(
+                f"You cannot set `low_cpu_mem_usage` to False while using device_map={device_map} for loading and"
+                " dispatching. Please make sure to set `low_cpu_mem_usage=True`."
+            )
+
+        # import it here to avoid circular import
+        from diffusers import pipelines
+
+        # 6. Load each module in the pipeline
+        for name, (library_name, class_name) in logging.tqdm(init_dict.items(), desc="Loading pipeline components..."):
+            # 6.1 - now that JAX/Flax is an official framework of the library, we might load from Flax names
+            class_name = class_name[4:] if class_name.startswith("Flax") else class_name
+
+            # 6.2 Define all importable classes
+            is_pipeline_module = hasattr(pipelines, library_name) 
+            importable_classes = ALL_IMPORTABLE_CLASSES
+            loaded_sub_model = None
+
+            # 6.3 Use passed sub model or load class_name from library_name
+            if name in passed_class_obj:
+                # if the model is in a pipeline module, then we load it from the pipeline
+                # check that passed_class_obj has correct parent class
+                maybe_raise_or_warn(
+                    library_name, library, class_name, importable_classes, passed_class_obj, name, is_pipeline_module
+                )
+
+                loaded_sub_model = passed_class_obj[name]
+            else:
+                # load sub model
+                loaded_sub_model = load_sub_model_oms(
+                    library_name=library_name,
+                    class_name=class_name,
+                    importable_classes=importable_classes,
+                    pipelines=pipelines,
+                    is_pipeline_module=is_pipeline_module,
+                    pipeline_class=pipeline_class,
+                    torch_dtype=torch_dtype,
+                    provider=provider,
+                    sess_options=sess_options,
+                    device_map=device_map,
+                    max_memory=max_memory,
+                    offload_folder=offload_folder,
+                    offload_state_dict=offload_state_dict,
+                    model_variants=model_variants,
+                    name=name,
+                    from_flax=from_flax,
+                    variant=variant,
+                    low_cpu_mem_usage=low_cpu_mem_usage,
+                    cached_folder=cached_folder,
+                )
+                logger.info(
+                    f"Loaded {name} as {class_name} from `{name}` subfolder of {pretrained_model_name_or_path}."
+                )
+
+            init_kwargs[name] = loaded_sub_model  # UNet(...), # DiffusionSchedule(...)
+
+        if pipeline_class._load_connected_pipes and os.path.isfile(os.path.join(cached_folder, "README.md")):
+            modelcard = ModelCard.load(os.path.join(cached_folder, "README.md"))
+            connected_pipes = {prefix: getattr(modelcard.data, prefix, [None])[0] for prefix in CONNECTED_PIPES_KEYS}
+            load_kwargs = {
+                "cache_dir": cache_dir,
+                "resume_download": resume_download,
+                "force_download": force_download,
+                "proxies": proxies,
+                "local_files_only": local_files_only,
+                "use_auth_token": use_auth_token,
+                "revision": revision,
+                "torch_dtype": torch_dtype,
+                "custom_pipeline": custom_pipeline,
+                "custom_revision": custom_revision,
+                "provider": provider,
+                "sess_options": sess_options,
+                "device_map": device_map,
+                "max_memory": max_memory,
+                "offload_folder": offload_folder,
+                "offload_state_dict": offload_state_dict,
+                "low_cpu_mem_usage": low_cpu_mem_usage,
+                "variant": variant,
+                "use_safetensors": use_safetensors,
+            }
+            connected_pipes = {
+                prefix: DiffusionPipeline.from_pretrained(repo_id, **load_kwargs.copy())
+                for prefix, repo_id in connected_pipes.items()
+                if repo_id is not None
+            }
+
+            for prefix, connected_pipe in connected_pipes.items():
+                # add connected pipes to `init_kwargs` with <prefix>_<component_name>, e.g. "prior_text_encoder"
+                init_kwargs.update(
+                    {"_".join([prefix, name]): component for name, component in connected_pipe.components.items()}
+                )
+
+        # 7. Potentially add passed objects if expected
+        missing_modules = set(expected_modules) - set(init_kwargs.keys())
+        passed_modules = list(passed_class_obj.keys())
+        optional_modules = pipeline_class._optional_components
+        if len(missing_modules) > 0 and missing_modules <= set(passed_modules + optional_modules):
+            for module in missing_modules:
+                init_kwargs[module] = passed_class_obj.get(module, None)
+        elif len(missing_modules) > 0:
+            passed_modules = set(list(init_kwargs.keys()) + list(passed_class_obj.keys())) - optional_kwargs
+            raise ValueError(
+                f"Pipeline {pipeline_class} expected {expected_modules}, but only {passed_modules} were passed."
+            )
+
+        # 8. Instantiate the pipeline
+        model = pipeline_class(**init_kwargs)
+
+        # 9. Save where the model was instantiated from
+        model.register_to_config(_name_or_path=pretrained_model_name_or_path)
+        return model
+
+    @torch.no_grad()
+    # @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        oms_prompt: Union[str, List[str]] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 50,
+        num_images_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        oms_guidance_scale: float = 1.0,
+        oms_flag: bool = True,
+        **kwargs,
+    ):
+        """Pseudo-doc for OMS"""
+
+        if oms_flag is True:
+            if oms_prompt is not None:
+                sd_prompt = prompt
+                prompt = oms_prompt
+
+            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)
+
+
+            height = height or self.default_sample_size * self.vae_scale_factor
+            width = width or self.default_sample_size * self.vae_scale_factor
+            device = self._execution_device
+            ## Guidance flag for OMS
+            if oms_guidance_scale is not None:
+                do_classifier_free_guidance_for_oms = True
+            else:
+                do_classifier_free_guidance_for_oms = False
+
+
+            oms_prompt_emb = self.oms_text_encode(prompt,num_images_per_prompt,device)
+            if do_classifier_free_guidance_for_oms:
+                oms_negative_prompt = ([''] * (batch_size // num_images_per_prompt))
+                oms_negative_prompt_emb = self.oms_text_encode(oms_negative_prompt,num_images_per_prompt,device)
+
+            # 4. Prepare timesteps
+            self.scheduler.set_timesteps(num_inference_steps, device=device)
+
+            timesteps = self.scheduler.timesteps
+
+            # 5. Prepare latent variables
+            num_channels_latents = self.oms_module.config.in_channels
+            latents = self.prepare_latents(
+                batch_size * num_images_per_prompt,
+                num_channels_latents,
+                height,
+                width,
+                oms_prompt_emb.dtype,
+                device,
+                generator,
+                latents=None,
+            )
+
+            ## OMS CFG
+            if do_classifier_free_guidance_for_oms:
+                oms_prompt_emb = torch.cat([oms_negative_prompt_emb, oms_prompt_emb], dim=0)
+
+        
+            ## OMS to device
+            oms_prompt_emb = oms_prompt_emb.to(device)
+
+
+            ## Perform OMS
+            alphas_cumprod = self.scheduler.alphas_cumprod.to(device)
+            alpha_prod_t_prev =  alphas_cumprod[int(timesteps[0].item())] 
+            latent_input_oms = torch.cat([latents] * 2) if do_classifier_free_guidance_for_oms else latents
+            v_pred_oms = self.oms_module(latent_input_oms, oms_prompt_emb)['sample']
+            latents = self.oms_step(v_pred_oms, latents, do_classifier_free_guidance_for_oms, oms_guidance_scale, generator, alpha_prod_t_prev)
+            
+
+            if oms_prompt is not None:
+                prompt = sd_prompt
+
+            print('OMS Completed')
+        else:
+            print("OMS unloaded")
+            latents = None
+        output = self.sd_pipeline(
+                prompt = prompt,
+                height = height,
+                width = width,
+                num_inference_steps = num_inference_steps,
+                num_images_per_prompt = num_images_per_prompt,
+                generator = generator,
+                latents = latents,
+                **kwargs
+                )
+
+        return output

diffusers_patch/pipelines/oms/utils.py

@@ -0,0 +1,70 @@
+import torch
+from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer
+
+
+class SDXLTextEncoder(torch.nn.Module):
+    """Wrapper around HuggingFace text encoders for SDXL.
+
+    Creates two text encoders (a CLIPTextModel and CLIPTextModelWithProjection) that behave like one.
+
+    Args:
+        model_name (str): Name of the model's text encoders to load. Defaults to 'stabilityai/stable-diffusion-xl-base-1.0'.
+        encode_latents_in_fp16 (bool): Whether to encode latents in fp16. Defaults to True.
+    """
+
+    def __init__(self, model_name='stabilityai/stable-diffusion-xl-base-1.0', encode_latents_in_fp16=True, torch_dtype=None):
+        super().__init__()
+        if torch_dtype is None:
+            torch_dtype = torch.float16 if encode_latents_in_fp16 else None
+        self.text_encoder = CLIPTextModel.from_pretrained(model_name, subfolder='text_encoder', torch_dtype=torch_dtype)
+        self.text_encoder_2 = CLIPTextModelWithProjection.from_pretrained(model_name,
+                                                                          subfolder='text_encoder_2',
+                                                                          torch_dtype=torch_dtype)
+
+    @property
+    def device(self):
+        return self.text_encoder.device
+
+    def forward(self, tokenized_text):
+        # first text encoder
+        conditioning = self.text_encoder(tokenized_text[0], output_hidden_states=True).hidden_states[-2]
+        # second text encoder
+        text_encoder_2_out = self.text_encoder_2(tokenized_text[1], output_hidden_states=True)
+        pooled_conditioning = text_encoder_2_out[0]  # (batch_size, 1280)
+        conditioning_2 = text_encoder_2_out.hidden_states[-2]  # (batch_size, 77, 1280)
+
+        conditioning = torch.concat([conditioning, conditioning_2], dim=-1)
+        return conditioning, pooled_conditioning
+
+
+class SDXLTokenizer:
+    """Wrapper around HuggingFace tokenizers for SDXL.
+
+    Tokenizes prompt with two tokenizers and returns the joined output.
+
+    Args:
+        model_name (str): Name of the model's text encoders to load. Defaults to 'stabilityai/stable-diffusion-xl-base-1.0'.
+    """
+
+    def __init__(self, model_name='stabilityai/stable-diffusion-xl-base-1.0'):
+        self.tokenizer = CLIPTokenizer.from_pretrained(model_name, subfolder='tokenizer')
+        self.tokenizer_2 = CLIPTokenizer.from_pretrained(model_name, subfolder='tokenizer_2')
+
+    def __call__(self, prompt, padding, truncation, return_tensors, max_length=None):
+        tokenized_output = self.tokenizer(
+            prompt,
+            padding=padding,
+            max_length=self.tokenizer.model_max_length if max_length is None else max_length,
+            truncation=truncation,
+            return_tensors=return_tensors)
+        tokenized_output_2 = self.tokenizer_2(
+            prompt,
+            padding=padding,
+            max_length=self.tokenizer_2.model_max_length if max_length is None else max_length,
+            truncation=truncation,
+            return_tensors=return_tensors)
+
+        # Add second tokenizer output to first tokenizer
+        for key in tokenized_output.keys():
+            tokenized_output[key] = [tokenized_output[key], tokenized_output_2[key]]
+        return tokenized_output

requirements.txt

@@ -0,0 +1,6 @@
+diffusers==0.23.1
+transformers==4.35.2
+accelerate==0.24.1
+gradio==4.7.1
+pydantic==1.10.13
+spacy==3.7.2