marigold folder added

Marigold/LICENSE.txt

@@ -0,0 +1,177 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS

Marigold/README.md

@@ -0,0 +1 @@
+Code is copied from https://github.com/prs-eth/Marigold.

Marigold/resnet/__init__.py

@@ -0,0 +1,14 @@
+from .resnet import ResnetBlock2D, ResnetBlockCondNorm2D
+from .downsampling import Downsample2D, FirDownsample2D, KDownsample2D
+from .upsampling import Upsample2D, FirUpsample2D, KUpsample2D
+
+__all__ = [
+    "ResnetBlock2D", 
+    "ResnetBlockCondNorm2D",
+    "Downsample2D",
+    "FirDownsample2D", 
+    "KDownsample2D",
+    "Upsample2D",
+    "FirUpsample2D",
+    "KUpsample2D"
+]

Marigold/resnet/downsampling.py

@@ -0,0 +1,406 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from diffusers.utils import deprecate
+from diffusers.models.normalization import RMSNorm
+from Marigold.resnet.upsampling import upfirdn2d_native
+from einops import rearrange
+
+
+class Downsample1D(nn.Module):
+    """A 1D downsampling layer with an optional convolution.
+
+    Parameters:
+        channels (`int`):
+            number of channels in the inputs and outputs.
+        use_conv (`bool`, default `False`):
+            option to use a convolution.
+        out_channels (`int`, optional):
+            number of output channels. Defaults to `channels`.
+        padding (`int`, default `1`):
+            padding for the convolution.
+        name (`str`, default `conv`):
+            name of the downsampling 1D layer.
+    """
+
+    def __init__(
+        self,
+        channels: int,
+        use_conv: bool = False,
+        out_channels: Optional[int] = None,
+        padding: int = 1,
+        name: str = "conv",
+    ):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.padding = padding
+        stride = 2
+        self.name = name
+
+        if use_conv:
+            self.conv = nn.Conv1d(self.channels, self.out_channels, 3, stride=stride, padding=padding)
+        else:
+            assert self.channels == self.out_channels
+            self.conv = nn.AvgPool1d(kernel_size=stride, stride=stride)
+
+    def forward(self, inputs: torch.Tensor) -> torch.Tensor:
+        assert inputs.shape[1] == self.channels
+        return self.conv(inputs)
+
+
+class Downsample2D(nn.Module):
+    """A 2D downsampling layer with an optional convolution.
+
+    Parameters:
+        channels (`int`):
+            number of channels in the inputs and outputs.
+        use_conv (`bool`, default `False`):
+            option to use a convolution.
+        out_channels (`int`, optional):
+            number of output channels. Defaults to `channels`.
+        padding (`int`, default `1`):
+            padding for the convolution.
+        name (`str`, default `conv`):
+            name of the downsampling 2D layer.
+    """
+
+    def __init__(
+        self,
+        channels: int,
+        use_conv: bool = False,
+        out_channels: Optional[int] = None,
+        padding: int = 1,
+        name: str = "conv",
+        kernel_size=3,
+        norm_type=None,
+        eps=None,
+        elementwise_affine=None,
+        bias=True,
+    ):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.padding = padding
+        stride = 2
+        self.name = name
+
+        if norm_type == "ln_norm":
+            self.norm = nn.LayerNorm(channels, eps, elementwise_affine)
+        elif norm_type == "rms_norm":
+            self.norm = RMSNorm(channels, eps, elementwise_affine)
+        elif norm_type is None:
+            self.norm = None
+        else:
+            raise ValueError(f"unknown norm_type: {norm_type}")
+
+        if use_conv:
+            conv = nn.Conv2d(
+                self.channels, self.out_channels, kernel_size=kernel_size, stride=stride, padding=padding, bias=bias,
+            )
+        else:
+            assert self.channels == self.out_channels
+            conv = nn.AvgPool2d(kernel_size=stride, stride=stride)
+
+        # TODO(Suraj, Patrick) - clean up after weight dicts are correctly renamed
+        if name == "conv":
+            self.Conv2d_0 = conv
+            self.conv = conv
+        elif name == "Conv2d_0":
+            self.conv = conv
+        else:
+            self.conv = conv
+
+    def forward(self, hidden_states: torch.Tensor, *args, **kwargs) -> torch.Tensor:
+        if len(args) > 0 or kwargs.get("scale", None) is not None:
+            deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`."
+            deprecate("scale", "1.0.0", deprecation_message)
+        assert hidden_states.shape[1] == self.channels
+
+        if self.norm is not None:
+            hidden_states_permuted = hidden_states.permute(0, 2, 3, 1)  # [N, C, H, W] -> [N, H, W, C]
+            b, c, h, w = hidden_states_permuted.shape
+            hidden_states_permuted = rearrange(hidden_states_permuted, "(b t) c h w -> b c (h w t)", b=1, h=h, w=w)
+            hidden_states = self.norm(hidden_states_permuted)
+            hidden_states = rearrange(hidden_states, "b c (h w t) -> (b t) c h w", b=1, h=h, w=w)
+            hidden_states = hidden_states.permute(0, 3, 1, 2)  # [N, H, W, C] -> [N, C, H, W]
+        if self.use_conv and self.padding == 0:
+            pad = (0, 1, 0, 1)
+            hidden_states = F.pad(hidden_states, pad, mode="constant", value=0)
+
+        assert hidden_states.shape[1] == self.channels
+
+        hidden_states = self.conv(hidden_states)
+
+        return hidden_states
+
+
+class FirDownsample2D(nn.Module):
+    """A 2D FIR downsampling layer with an optional convolution.
+
+    Parameters:
+        channels (`int`):
+            number of channels in the inputs and outputs.
+        use_conv (`bool`, default `False`):
+            option to use a convolution.
+        out_channels (`int`, optional):
+            number of output channels. Defaults to `channels`.
+        fir_kernel (`tuple`, default `(1, 3, 3, 1)`):
+            kernel for the FIR filter.
+    """
+
+    def __init__(
+        self,
+        channels: Optional[int] = None,
+        out_channels: Optional[int] = None,
+        use_conv: bool = False,
+        fir_kernel: Tuple[int, int, int, int] = (1, 3, 3, 1),
+    ):
+        super().__init__()
+        out_channels = out_channels if out_channels else channels
+        if use_conv:
+            self.Conv2d_0 = nn.Conv2d(channels, out_channels, kernel_size=3, stride=1, padding=1)
+        self.fir_kernel = fir_kernel
+        self.use_conv = use_conv
+        self.out_channels = out_channels
+
+    def _downsample_2d(
+        self,
+        hidden_states: torch.Tensor,
+        weight: Optional[torch.Tensor] = None,
+        kernel: Optional[torch.Tensor] = None,
+        factor: int = 2,
+        gain: float = 1,
+    ) -> torch.Tensor:
+        """Fused `Conv2d()` followed by `downsample_2d()`.
+        Padding is performed only once at the beginning, not between the operations. The fused op is considerably more
+        efficient than performing the same calculation using standard TensorFlow ops. It supports gradients of
+        arbitrary order.
+
+        Args:
+            hidden_states (`torch.Tensor`):
+                Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
+            weight (`torch.Tensor`, *optional*):
+                Weight tensor of the shape `[filterH, filterW, inChannels, outChannels]`. Grouped convolution can be
+                performed by `inChannels = x.shape[0] // numGroups`.
+            kernel (`torch.Tensor`, *optional*):
+                FIR filter of the shape `[firH, firW]` or `[firN]` (separable). The default is `[1] * factor`, which
+                corresponds to average pooling.
+            factor (`int`, *optional*, default to `2`):
+                Integer downsampling factor.
+            gain (`float`, *optional*, default to `1.0`):
+                Scaling factor for signal magnitude.
+
+        Returns:
+            output (`torch.Tensor`):
+                Tensor of the shape `[N, C, H // factor, W // factor]` or `[N, H // factor, W // factor, C]`, and same
+                datatype as `x`.
+        """
+
+        assert isinstance(factor, int) and factor >= 1
+        if kernel is None:
+            kernel = [1] * factor
+
+        # setup kernel
+        kernel = torch.tensor(kernel, dtype=torch.float32)
+        if kernel.ndim == 1:
+            kernel = torch.outer(kernel, kernel)
+        kernel /= torch.sum(kernel)
+
+        kernel = kernel * gain
+
+        if self.use_conv:
+            _, _, convH, convW = weight.shape
+            pad_value = (kernel.shape[0] - factor) + (convW - 1)
+            stride_value = [factor, factor]
+            upfirdn_input = upfirdn2d_native(
+                hidden_states,
+                torch.tensor(kernel, device=hidden_states.device),
+                pad=((pad_value + 1) // 2, pad_value // 2),
+            )
+            output = F.conv2d(upfirdn_input, weight, stride=stride_value, padding=0)
+        else:
+            pad_value = kernel.shape[0] - factor
+            output = upfirdn2d_native(
+                hidden_states,
+                torch.tensor(kernel, device=hidden_states.device),
+                down=factor,
+                pad=((pad_value + 1) // 2, pad_value // 2),
+            )
+
+        return output
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        if self.use_conv:
+            downsample_input = self._downsample_2d(hidden_states, weight=self.Conv2d_0.weight, kernel=self.fir_kernel)
+            hidden_states = downsample_input + self.Conv2d_0.bias.reshape(1, -1, 1, 1)
+        else:
+            hidden_states = self._downsample_2d(hidden_states, kernel=self.fir_kernel, factor=2)
+
+        return hidden_states
+
+
+# downsample/upsample layer used in k-upscaler, might be able to use FirDownsample2D/DirUpsample2D instead
+class KDownsample2D(nn.Module):
+    r"""A 2D K-downsampling layer.
+
+    Parameters:
+        pad_mode (`str`, *optional*, default to `"reflect"`): the padding mode to use.
+    """
+
+    def __init__(self, pad_mode: str = "reflect"):
+        super().__init__()
+        self.pad_mode = pad_mode
+        kernel_1d = torch.tensor([[1 / 8, 3 / 8, 3 / 8, 1 / 8]])
+        self.pad = kernel_1d.shape[1] // 2 - 1
+        self.register_buffer("kernel", kernel_1d.T @ kernel_1d, persistent=False)
+
+    def forward(self, inputs: torch.Tensor) -> torch.Tensor:
+        inputs = F.pad(inputs, (self.pad,) * 4, self.pad_mode)
+        weight = inputs.new_zeros(
+            [
+                inputs.shape[1],
+                inputs.shape[1],
+                self.kernel.shape[0],
+                self.kernel.shape[1],
+            ]
+        )
+        indices = torch.arange(inputs.shape[1], device=inputs.device)
+        kernel = self.kernel.to(weight)[None, :].expand(inputs.shape[1], -1, -1)
+        weight[indices, indices] = kernel
+        return F.conv2d(inputs, weight, stride=2)
+
+
+class CogVideoXDownsample3D(nn.Module):
+    # Todo: Wait for paper relase.
+    r"""
+    A 3D Downsampling layer using in [CogVideoX]() by Tsinghua University & ZhipuAI
+
+    Args:
+        in_channels (`int`):
+            Number of channels in the input image.
+        out_channels (`int`):
+            Number of channels produced by the convolution.
+        kernel_size (`int`, defaults to `3`):
+            Size of the convolving kernel.
+        stride (`int`, defaults to `2`):
+            Stride of the convolution.
+        padding (`int`, defaults to `0`):
+            Padding added to all four sides of the input.
+        compress_time (`bool`, defaults to `False`):
+            Whether or not to compress the time dimension.
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        kernel_size: int = 3,
+        stride: int = 2,
+        padding: int = 0,
+        compress_time: bool = False,
+    ):
+        super().__init__()
+
+        self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=padding)
+        self.compress_time = compress_time
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        if self.compress_time:
+            batch_size, channels, frames, height, width = x.shape
+
+            # (batch_size, channels, frames, height, width) -> (batch_size, height, width, channels, frames) -> (batch_size * height * width, channels, frames)
+            x = x.permute(0, 3, 4, 1, 2).reshape(batch_size * height * width, channels, frames)
+
+            if x.shape[-1] % 2 == 1:
+                x_first, x_rest = x[..., 0], x[..., 1:]
+                if x_rest.shape[-1] > 0:
+                    # (batch_size * height * width, channels, frames - 1) -> (batch_size * height * width, channels, (frames - 1) // 2)
+                    x_rest = F.avg_pool1d(x_rest, kernel_size=2, stride=2)
+
+                x = torch.cat([x_first[..., None], x_rest], dim=-1)
+                # (batch_size * height * width, channels, (frames // 2) + 1) -> (batch_size, height, width, channels, (frames // 2) + 1) -> (batch_size, channels, (frames // 2) + 1, height, width)
+                x = x.reshape(batch_size, height, width, channels, x.shape[-1]).permute(0, 3, 4, 1, 2)
+            else:
+                # (batch_size * height * width, channels, frames) -> (batch_size * height * width, channels, frames // 2)
+                x = F.avg_pool1d(x, kernel_size=2, stride=2)
+                # (batch_size * height * width, channels, frames // 2) -> (batch_size, height, width, channels, frames // 2) -> (batch_size, channels, frames // 2, height, width)
+                x = x.reshape(batch_size, height, width, channels, x.shape[-1]).permute(0, 3, 4, 1, 2)
+
+        # Pad the tensor
+        pad = (0, 1, 0, 1)
+        x = F.pad(x, pad, mode="constant", value=0)
+        batch_size, channels, frames, height, width = x.shape
+        # (batch_size, channels, frames, height, width) -> (batch_size, frames, channels, height, width) -> (batch_size * frames, channels, height, width)
+        x = x.permute(0, 2, 1, 3, 4).reshape(batch_size * frames, channels, height, width)
+        x = self.conv(x)
+        # (batch_size * frames, channels, height, width) -> (batch_size, frames, channels, height, width) -> (batch_size, channels, frames, height, width)
+        x = x.reshape(batch_size, frames, x.shape[1], x.shape[2], x.shape[3]).permute(0, 2, 1, 3, 4)
+        return x
+
+
+def downsample_2d(
+    hidden_states: torch.Tensor,
+    kernel: Optional[torch.Tensor] = None,
+    factor: int = 2,
+    gain: float = 1,
+) -> torch.Tensor:
+    r"""Downsample2D a batch of 2D images with the given filter.
+    Accepts a batch of 2D images of the shape `[N, C, H, W]` or `[N, H, W, C]` and downsamples each image with the
+    given filter. The filter is normalized so that if the input pixels are constant, they will be scaled by the
+    specified `gain`. Pixels outside the image are assumed to be zero, and the filter is padded with zeros so that its
+    shape is a multiple of the downsampling factor.
+
+    Args:
+        hidden_states (`torch.Tensor`)
+            Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
+        kernel (`torch.Tensor`, *optional*):
+            FIR filter of the shape `[firH, firW]` or `[firN]` (separable). The default is `[1] * factor`, which
+            corresponds to average pooling.
+        factor (`int`, *optional*, default to `2`):
+            Integer downsampling factor.
+        gain (`float`, *optional*, default to `1.0`):
+            Scaling factor for signal magnitude.
+
+    Returns:
+        output (`torch.Tensor`):
+            Tensor of the shape `[N, C, H // factor, W // factor]`
+    """
+
+    assert isinstance(factor, int) and factor >= 1
+    if kernel is None:
+        kernel = [1] * factor
+
+    kernel = torch.tensor(kernel, dtype=torch.float32)
+    if kernel.ndim == 1:
+        kernel = torch.outer(kernel, kernel)
+    kernel /= torch.sum(kernel)
+
+    kernel = kernel * gain
+    pad_value = kernel.shape[0] - factor
+    output = upfirdn2d_native(
+        hidden_states,
+        kernel.to(device=hidden_states.device),
+        down=factor,
+        pad=((pad_value + 1) // 2, pad_value // 2),
+    )
+    return output

Marigold/resnet/resnet.py

@@ -0,0 +1,802 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+# `TemporalConvLayer` Copyright 2024 Alibaba DAMO-VILAB, The ModelScope Team and 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 functools import partial
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from diffusers.utils import deprecate
+from diffusers.models.activations import get_activation
+from diffusers.models.attention_processor import SpatialNorm
+from Marigold.resnet.downsampling import (  # noqa
+    Downsample2D,
+    downsample_2d,
+)
+from diffusers.models.normalization import AdaGroupNorm
+from Marigold.resnet.upsampling import (  # noqa
+    Upsample2D,
+    upsample_2d,
+)
+from einops import rearrange
+
+class ResnetBlockCondNorm2D(nn.Module):
+    r"""
+    A Resnet block that use normalization layer that incorporate conditioning information.
+
+    Parameters:
+        in_channels (`int`): The number of channels in the input.
+        out_channels (`int`, *optional*, default to be `None`):
+            The number of output channels for the first conv2d layer. If None, same as `in_channels`.
+        dropout (`float`, *optional*, defaults to `0.0`): The dropout probability to use.
+        temb_channels (`int`, *optional*, default to `512`): the number of channels in timestep embedding.
+        groups (`int`, *optional*, default to `32`): The number of groups to use for the first normalization layer.
+        groups_out (`int`, *optional*, default to None):
+            The number of groups to use for the second normalization layer. if set to None, same as `groups`.
+        eps (`float`, *optional*, defaults to `1e-6`): The epsilon to use for the normalization.
+        non_linearity (`str`, *optional*, default to `"swish"`): the activation function to use.
+        time_embedding_norm (`str`, *optional*, default to `"ada_group"` ):
+            The normalization layer for time embedding `temb`. Currently only support "ada_group" or "spatial".
+        kernel (`torch.Tensor`, optional, default to None): FIR filter, see
+            [`~models.resnet.FirUpsample2D`] and [`~models.resnet.FirDownsample2D`].
+        output_scale_factor (`float`, *optional*, default to be `1.0`): the scale factor to use for the output.
+        use_in_shortcut (`bool`, *optional*, default to `True`):
+            If `True`, add a 1x1 nn.conv2d layer for skip-connection.
+        up (`bool`, *optional*, default to `False`): If `True`, add an upsample layer.
+        down (`bool`, *optional*, default to `False`): If `True`, add a downsample layer.
+        conv_shortcut_bias (`bool`, *optional*, default to `True`):  If `True`, adds a learnable bias to the
+            `conv_shortcut` output.
+        conv_2d_out_channels (`int`, *optional*, default to `None`): the number of channels in the output.
+            If None, same as `out_channels`.
+    """
+
+    def __init__(
+        self,
+        *,
+        in_channels: int,
+        out_channels: Optional[int] = None,
+        conv_shortcut: bool = False,
+        dropout: float = 0.0,
+        temb_channels: int = 512,
+        groups: int = 32,
+        groups_out: Optional[int] = None,
+        eps: float = 1e-6,
+        non_linearity: str = "swish",
+        time_embedding_norm: str = "ada_group",  # ada_group, spatial
+        output_scale_factor: float = 1.0,
+        use_in_shortcut: Optional[bool] = None,
+        up: bool = False,
+        down: bool = False,
+        conv_shortcut_bias: bool = True,
+        conv_2d_out_channels: Optional[int] = None,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+        self.use_conv_shortcut = conv_shortcut
+        self.up = up
+        self.down = down
+        self.output_scale_factor = output_scale_factor
+        self.time_embedding_norm = time_embedding_norm
+
+        if groups_out is None:
+            groups_out = groups
+
+        if self.time_embedding_norm == "ada_group":  # ada_group
+            self.norm1 = AdaGroupNorm(temb_channels, in_channels, groups, eps=eps)
+        elif self.time_embedding_norm == "spatial":
+            self.norm1 = SpatialNorm(in_channels, temb_channels)
+        else:
+            raise ValueError(f" unsupported time_embedding_norm: {self.time_embedding_norm}")
+
+        self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1)
+
+        if self.time_embedding_norm == "ada_group":  # ada_group
+            self.norm2 = AdaGroupNorm(temb_channels, out_channels, groups_out, eps=eps)
+        elif self.time_embedding_norm == "spatial":  # spatial
+            self.norm2 = SpatialNorm(out_channels, temb_channels)
+        else:
+            raise ValueError(f" unsupported time_embedding_norm: {self.time_embedding_norm}")
+
+        self.dropout = torch.nn.Dropout(dropout)
+
+        conv_2d_out_channels = conv_2d_out_channels or out_channels
+        self.conv2 = nn.Conv2d(out_channels, conv_2d_out_channels, kernel_size=3, stride=1, padding=1)
+
+        self.nonlinearity = get_activation(non_linearity)
+
+        self.upsample = self.downsample = None
+        if self.up:
+            self.upsample = Upsample2D(in_channels, use_conv=False)
+        elif self.down:
+            self.downsample = Downsample2D(in_channels, use_conv=False, padding=1, name="op")
+
+        self.use_in_shortcut = self.in_channels != conv_2d_out_channels if use_in_shortcut is None else use_in_shortcut
+
+        self.conv_shortcut = None
+        if self.use_in_shortcut:
+            self.conv_shortcut = nn.Conv2d(
+                in_channels,
+                conv_2d_out_channels,
+                kernel_size=1,
+                stride=1,
+                padding=0,
+                bias=conv_shortcut_bias,
+            )
+
+    def forward(self, input_tensor: torch.Tensor, temb: torch.Tensor, *args, **kwargs) -> torch.Tensor:
+        if len(args) > 0 or kwargs.get("scale", None) is not None:
+            deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`."
+            deprecate("scale", "1.0.0", deprecation_message)
+
+        hidden_states = input_tensor
+
+        hidden_states = self.norm1(hidden_states, temb)
+
+        hidden_states = self.nonlinearity(hidden_states)
+
+        if self.upsample is not None:
+            # upsample_nearest_nhwc fails with large batch sizes. see https://github.com/huggingface/diffusers/issues/984
+            if hidden_states.shape[0] >= 64:
+                input_tensor = input_tensor.contiguous()
+                hidden_states = hidden_states.contiguous()
+            input_tensor = self.upsample(input_tensor)
+            hidden_states = self.upsample(hidden_states)
+
+        elif self.downsample is not None:
+            input_tensor = self.downsample(input_tensor)
+            hidden_states = self.downsample(hidden_states)
+
+        hidden_states = self.conv1(hidden_states)
+
+        hidden_states = self.norm2(hidden_states, temb)
+
+        hidden_states = self.nonlinearity(hidden_states)
+
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.conv2(hidden_states)
+
+        if self.conv_shortcut is not None:
+            input_tensor = self.conv_shortcut(input_tensor)
+
+        output_tensor = (input_tensor + hidden_states) / self.output_scale_factor
+
+        return output_tensor
+
+
+class ResnetBlock2D(nn.Module):
+    r"""
+    A Resnet block.
+
+    Parameters:
+        in_channels (`int`): The number of channels in the input.
+        out_channels (`int`, *optional*, default to be `None`):
+            The number of output channels for the first conv2d layer. If None, same as `in_channels`.
+        dropout (`float`, *optional*, defaults to `0.0`): The dropout probability to use.
+        temb_channels (`int`, *optional*, default to `512`): the number of channels in timestep embedding.
+        groups (`int`, *optional*, default to `32`): The number of groups to use for the first normalization layer.
+        groups_out (`int`, *optional*, default to None):
+            The number of groups to use for the second normalization layer. if set to None, same as `groups`.
+        eps (`float`, *optional*, defaults to `1e-6`): The epsilon to use for the normalization.
+        non_linearity (`str`, *optional*, default to `"swish"`): the activation function to use.
+        time_embedding_norm (`str`, *optional*, default to `"default"` ): Time scale shift config.
+            By default, apply timestep embedding conditioning with a simple shift mechanism. Choose "scale_shift" for a
+            stronger conditioning with scale and shift.
+        kernel (`torch.Tensor`, optional, default to None): FIR filter, see
+            [`~models.resnet.FirUpsample2D`] and [`~models.resnet.FirDownsample2D`].
+        output_scale_factor (`float`, *optional*, default to be `1.0`): the scale factor to use for the output.
+        use_in_shortcut (`bool`, *optional*, default to `True`):
+            If `True`, add a 1x1 nn.conv2d layer for skip-connection.
+        up (`bool`, *optional*, default to `False`): If `True`, add an upsample layer.
+        down (`bool`, *optional*, default to `False`): If `True`, add a downsample layer.
+        conv_shortcut_bias (`bool`, *optional*, default to `True`):  If `True`, adds a learnable bias to the
+            `conv_shortcut` output.
+        conv_2d_out_channels (`int`, *optional*, default to `None`): the number of channels in the output.
+            If None, same as `out_channels`.
+    """
+
+    def __init__(
+        self,
+        *,
+        in_channels: int,
+        out_channels: Optional[int] = None,
+        conv_shortcut: bool = False,
+        dropout: float = 0.0,
+        temb_channels: int = 512,
+        groups: int = 32,
+        groups_out: Optional[int] = None,
+        pre_norm: bool = True,
+        eps: float = 1e-6,
+        non_linearity: str = "swish",
+        skip_time_act: bool = False,
+        time_embedding_norm: str = "default",  # default, scale_shift,
+        kernel: Optional[torch.Tensor] = None,
+        output_scale_factor: float = 1.0,
+        use_in_shortcut: Optional[bool] = None,
+        up: bool = False,
+        down: bool = False,
+        conv_shortcut_bias: bool = True,
+        conv_2d_out_channels: Optional[int] = None,
+    ):
+        super().__init__()
+        if time_embedding_norm == "ada_group":
+            raise ValueError(
+                "This class cannot be used with `time_embedding_norm==ada_group`, please use `ResnetBlockCondNorm2D` instead",
+            )
+        if time_embedding_norm == "spatial":
+            raise ValueError(
+                "This class cannot be used with `time_embedding_norm==spatial`, please use `ResnetBlockCondNorm2D` instead",
+            )
+
+        self.pre_norm = True
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+        self.use_conv_shortcut = conv_shortcut
+        self.up = up
+        self.down = down
+        self.output_scale_factor = output_scale_factor
+        self.time_embedding_norm = time_embedding_norm
+        self.skip_time_act = skip_time_act
+
+        if groups_out is None:
+            groups_out = groups
+
+        self.norm1 = torch.nn.GroupNorm(num_groups=groups, num_channels=in_channels, eps=eps, affine=True)
+
+        self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1)
+
+        if temb_channels is not None:
+            if self.time_embedding_norm == "default":
+                self.time_emb_proj = nn.Linear(temb_channels, out_channels)
+            elif self.time_embedding_norm == "scale_shift":
+                self.time_emb_proj = nn.Linear(temb_channels, 2 * out_channels)
+            else:
+                raise ValueError(f"unknown time_embedding_norm : {self.time_embedding_norm} ")
+        else:
+            self.time_emb_proj = None
+
+        self.norm2 = torch.nn.GroupNorm(num_groups=groups_out, num_channels=out_channels, eps=eps, affine=True)
+
+        self.dropout = torch.nn.Dropout(dropout)
+        conv_2d_out_channels = conv_2d_out_channels or out_channels
+        self.conv2 = nn.Conv2d(out_channels, conv_2d_out_channels, kernel_size=3, stride=1, padding=1)
+
+        self.nonlinearity = get_activation(non_linearity)
+
+        self.upsample = self.downsample = None
+        if self.up:
+            if kernel == "fir":
+                fir_kernel = (1, 3, 3, 1)
+                self.upsample = lambda x: upsample_2d(x, kernel=fir_kernel)
+            elif kernel == "sde_vp":
+                self.upsample = partial(F.interpolate, scale_factor=2.0, mode="nearest")
+            else:
+                self.upsample = Upsample2D(in_channels, use_conv=False)
+        elif self.down:
+            if kernel == "fir":
+                fir_kernel = (1, 3, 3, 1)
+                self.downsample = lambda x: downsample_2d(x, kernel=fir_kernel)
+            elif kernel == "sde_vp":
+                self.downsample = partial(F.avg_pool2d, kernel_size=2, stride=2)
+            else:
+                self.downsample = Downsample2D(in_channels, use_conv=False, padding=1, name="op")
+
+        self.use_in_shortcut = self.in_channels != conv_2d_out_channels if use_in_shortcut is None else use_in_shortcut
+
+        self.conv_shortcut = None
+        if self.use_in_shortcut:
+            self.conv_shortcut = nn.Conv2d(
+                in_channels,
+                conv_2d_out_channels,
+                kernel_size=1,
+                stride=1,
+                padding=0,
+                bias=conv_shortcut_bias,
+            )
+
+    def forward(self, input_tensor: torch.Tensor, temb: torch.Tensor, *args, **kwargs) -> torch.Tensor:
+        if len(args) > 0 or kwargs.get("scale", None) is not None:
+            deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`."
+            deprecate("scale", "1.0.0", deprecation_message)
+
+        hidden_states = input_tensor
+
+        b, c, h, w = hidden_states.shape
+        hidden_states = rearrange(hidden_states, "(b t) c h w -> b c (h w t)", t=6, h=h, w=w)
+        hidden_states = self.norm1(hidden_states)
+        hidden_states = rearrange(hidden_states, "b c (h w t) -> (b t) c h w", t=6, h=h, w=w)
+        hidden_states = self.nonlinearity(hidden_states)
+
+        if self.upsample is not None:
+            # upsample_nearest_nhwc fails with large batch sizes. see https://github.com/huggingface/diffusers/issues/984
+            if hidden_states.shape[0] >= 64:
+                input_tensor = input_tensor.contiguous()
+                hidden_states = hidden_states.contiguous()
+            input_tensor = self.upsample(input_tensor)
+            hidden_states = self.upsample(hidden_states)
+        elif self.downsample is not None:
+            input_tensor = self.downsample(input_tensor)
+            hidden_states = self.downsample(hidden_states)
+
+        hidden_states = self.conv1(hidden_states)
+
+        if self.time_emb_proj is not None:
+            if not self.skip_time_act:
+                temb = self.nonlinearity(temb)
+            temb = self.time_emb_proj(temb)[:, :, None, None]
+
+        if self.time_embedding_norm == "default":
+            if temb is not None:
+                hidden_states = hidden_states + temb
+            b, c, h, w = hidden_states.shape
+            hidden_states = rearrange(hidden_states, "(b t) c h w -> b c (h w t)", t=6, h=h, w=w)
+            hidden_states = self.norm2(hidden_states)
+            hidden_states = rearrange(hidden_states, "b c (h w t) -> (b t) c h w", t=6, h=h, w=w)
+        elif self.time_embedding_norm == "scale_shift":
+            if temb is None:
+                raise ValueError(
+                    f" `temb` should not be None when `time_embedding_norm` is {self.time_embedding_norm}"
+                )
+            time_scale, time_shift = torch.chunk(temb, 2, dim=1)
+            b, c, h, w = hidden_states.shape
+            hidden_states = rearrange(hidden_states, "(b t) c h w -> b c (h w t)", t=6, h=h, w=w)
+            hidden_states = self.norm2(hidden_states)
+            hidden_states = rearrange(hidden_states, "b c (h w t) -> (b t) c h w", t=6, h=h, w=w)
+            hidden_states = hidden_states * (1 + time_scale) + time_shift
+        else:
+            b, c, h, w = hidden_states.shape
+            hidden_states = rearrange(hidden_states, "(b t) c h w -> b c (h w t)", t=6, h=h, w=w)
+            hidden_states = self.norm2(hidden_states)
+            hidden_states = rearrange(hidden_states, "b c (h w t) -> (b t) c h w", t=6, h=h, w=w)
+
+        hidden_states = self.nonlinearity(hidden_states)
+
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.conv2(hidden_states)
+
+        if self.conv_shortcut is not None:
+            input_tensor = self.conv_shortcut(input_tensor)
+
+        output_tensor = (input_tensor + hidden_states) / self.output_scale_factor
+
+        return output_tensor
+
+
+# unet_rl.py
+def rearrange_dims(tensor: torch.Tensor) -> torch.Tensor:
+    if len(tensor.shape) == 2:
+        return tensor[:, :, None]
+    if len(tensor.shape) == 3:
+        return tensor[:, :, None, :]
+    elif len(tensor.shape) == 4:
+        return tensor[:, :, 0, :]
+    else:
+        raise ValueError(f"`len(tensor)`: {len(tensor)} has to be 2, 3 or 4.")
+
+
+class Conv1dBlock(nn.Module):
+    """
+    Conv1d --> GroupNorm --> Mish
+
+    Parameters:
+        inp_channels (`int`): Number of input channels.
+        out_channels (`int`): Number of output channels.
+        kernel_size (`int` or `tuple`): Size of the convolving kernel.
+        n_groups (`int`, default `8`): Number of groups to separate the channels into.
+        activation (`str`, defaults to `mish`): Name of the activation function.
+    """
+
+    def __init__(
+        self,
+        inp_channels: int,
+        out_channels: int,
+        kernel_size: Union[int, Tuple[int, int]],
+        n_groups: int = 8,
+        activation: str = "mish",
+    ):
+        super().__init__()
+
+        self.conv1d = nn.Conv1d(inp_channels, out_channels, kernel_size, padding=kernel_size // 2)
+        self.group_norm = nn.GroupNorm(n_groups, out_channels)
+        self.mish = get_activation(activation)
+
+    def forward(self, inputs: torch.Tensor) -> torch.Tensor:
+        intermediate_repr = self.conv1d(inputs)
+        intermediate_repr = rearrange_dims(intermediate_repr)
+        intermediate_repr = self.group_norm(intermediate_repr)
+        intermediate_repr = rearrange_dims(intermediate_repr)
+        output = self.mish(intermediate_repr)
+        return output
+
+
+# unet_rl.py
+class ResidualTemporalBlock1D(nn.Module):
+    """
+    Residual 1D block with temporal convolutions.
+
+    Parameters:
+        inp_channels (`int`): Number of input channels.
+        out_channels (`int`): Number of output channels.
+        embed_dim (`int`): Embedding dimension.
+        kernel_size (`int` or `tuple`): Size of the convolving kernel.
+        activation (`str`, defaults `mish`): It is possible to choose the right activation function.
+    """
+
+    def __init__(
+        self,
+        inp_channels: int,
+        out_channels: int,
+        embed_dim: int,
+        kernel_size: Union[int, Tuple[int, int]] = 5,
+        activation: str = "mish",
+    ):
+        super().__init__()
+        self.conv_in = Conv1dBlock(inp_channels, out_channels, kernel_size)
+        self.conv_out = Conv1dBlock(out_channels, out_channels, kernel_size)
+
+        self.time_emb_act = get_activation(activation)
+        self.time_emb = nn.Linear(embed_dim, out_channels)
+
+        self.residual_conv = (
+            nn.Conv1d(inp_channels, out_channels, 1) if inp_channels != out_channels else nn.Identity()
+        )
+
+    def forward(self, inputs: torch.Tensor, t: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+            inputs : [ batch_size x inp_channels x horizon ]
+            t : [ batch_size x embed_dim ]
+
+        returns:
+            out : [ batch_size x out_channels x horizon ]
+        """
+        t = self.time_emb_act(t)
+        t = self.time_emb(t)
+        out = self.conv_in(inputs) + rearrange_dims(t)
+        out = self.conv_out(out)
+        return out + self.residual_conv(inputs)
+
+
+class TemporalConvLayer(nn.Module):
+    """
+    Temporal convolutional layer that can be used for video (sequence of images) input Code mostly copied from:
+    https://github.com/modelscope/modelscope/blob/1509fdb973e5871f37148a4b5e5964cafd43e64d/modelscope/models/multi_modal/video_synthesis/unet_sd.py#L1016
+
+    Parameters:
+        in_dim (`int`): Number of input channels.
+        out_dim (`int`): Number of output channels.
+        dropout (`float`, *optional*, defaults to `0.0`): The dropout probability to use.
+    """
+
+    def __init__(
+        self,
+        in_dim: int,
+        out_dim: Optional[int] = None,
+        dropout: float = 0.0,
+        norm_num_groups: int = 32,
+    ):
+        super().__init__()
+        out_dim = out_dim or in_dim
+        self.in_dim = in_dim
+        self.out_dim = out_dim
+
+        # conv layers
+        self.conv1 = nn.Sequential(
+            nn.GroupNorm(norm_num_groups, in_dim),
+            nn.SiLU(),
+            nn.Conv3d(in_dim, out_dim, (3, 1, 1), padding=(1, 0, 0)),
+        )
+        self.conv2 = nn.Sequential(
+            nn.GroupNorm(norm_num_groups, out_dim),
+            nn.SiLU(),
+            nn.Dropout(dropout),
+            nn.Conv3d(out_dim, in_dim, (3, 1, 1), padding=(1, 0, 0)),
+        )
+        self.conv3 = nn.Sequential(
+            nn.GroupNorm(norm_num_groups, out_dim),
+            nn.SiLU(),
+            nn.Dropout(dropout),
+            nn.Conv3d(out_dim, in_dim, (3, 1, 1), padding=(1, 0, 0)),
+        )
+        self.conv4 = nn.Sequential(
+            nn.GroupNorm(norm_num_groups, out_dim),
+            nn.SiLU(),
+            nn.Dropout(dropout),
+            nn.Conv3d(out_dim, in_dim, (3, 1, 1), padding=(1, 0, 0)),
+        )
+
+        # zero out the last layer params,so the conv block is identity
+        nn.init.zeros_(self.conv4[-1].weight)
+        nn.init.zeros_(self.conv4[-1].bias)
+
+    def forward(self, hidden_states: torch.Tensor, num_frames: int = 1) -> torch.Tensor:
+        hidden_states = (
+            hidden_states[None, :].reshape((-1, num_frames) + hidden_states.shape[1:]).permute(0, 2, 1, 3, 4)
+        )
+
+        identity = hidden_states
+        hidden_states = self.conv1(hidden_states)
+        hidden_states = self.conv2(hidden_states)
+        hidden_states = self.conv3(hidden_states)
+        hidden_states = self.conv4(hidden_states)
+
+        hidden_states = identity + hidden_states
+
+        hidden_states = hidden_states.permute(0, 2, 1, 3, 4).reshape(
+            (hidden_states.shape[0] * hidden_states.shape[2], -1) + hidden_states.shape[3:]
+        )
+        return hidden_states
+
+
+class TemporalResnetBlock(nn.Module):
+    r"""
+    A Resnet block.
+
+    Parameters:
+        in_channels (`int`): The number of channels in the input.
+        out_channels (`int`, *optional*, default to be `None`):
+            The number of output channels for the first conv2d layer. If None, same as `in_channels`.
+        temb_channels (`int`, *optional*, default to `512`): the number of channels in timestep embedding.
+        eps (`float`, *optional*, defaults to `1e-6`): The epsilon to use for the normalization.
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: Optional[int] = None,
+        temb_channels: int = 512,
+        eps: float = 1e-6,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+
+        kernel_size = (3, 1, 1)
+        padding = [k // 2 for k in kernel_size]
+
+        self.norm1 = torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=eps, affine=True)
+        self.conv1 = nn.Conv3d(
+            in_channels,
+            out_channels,
+            kernel_size=kernel_size,
+            stride=1,
+            padding=padding,
+        )
+
+        if temb_channels is not None:
+            self.time_emb_proj = nn.Linear(temb_channels, out_channels)
+        else:
+            self.time_emb_proj = None
+
+        self.norm2 = torch.nn.GroupNorm(num_groups=32, num_channels=out_channels, eps=eps, affine=True)
+
+        self.dropout = torch.nn.Dropout(0.0)
+        self.conv2 = nn.Conv3d(
+            out_channels,
+            out_channels,
+            kernel_size=kernel_size,
+            stride=1,
+            padding=padding,
+        )
+
+        self.nonlinearity = get_activation("silu")
+
+        self.use_in_shortcut = self.in_channels != out_channels
+
+        self.conv_shortcut = None
+        if self.use_in_shortcut:
+            self.conv_shortcut = nn.Conv3d(
+                in_channels,
+                out_channels,
+                kernel_size=1,
+                stride=1,
+                padding=0,
+            )
+
+    def forward(self, input_tensor: torch.Tensor, temb: torch.Tensor) -> torch.Tensor:
+        hidden_states = input_tensor
+
+        hidden_states = self.norm1(hidden_states)
+        hidden_states = self.nonlinearity(hidden_states)
+        hidden_states = self.conv1(hidden_states)
+
+        if self.time_emb_proj is not None:
+            temb = self.nonlinearity(temb)
+            temb = self.time_emb_proj(temb)[:, :, :, None, None]
+            temb = temb.permute(0, 2, 1, 3, 4)
+            hidden_states = hidden_states + temb
+
+        hidden_states = self.norm2(hidden_states)
+        hidden_states = self.nonlinearity(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.conv2(hidden_states)
+
+        if self.conv_shortcut is not None:
+            input_tensor = self.conv_shortcut(input_tensor)
+
+        output_tensor = input_tensor + hidden_states
+
+        return output_tensor
+
+
+# VideoResBlock
+class SpatioTemporalResBlock(nn.Module):
+    r"""
+    A SpatioTemporal Resnet block.
+
+    Parameters:
+        in_channels (`int`): The number of channels in the input.
+        out_channels (`int`, *optional*, default to be `None`):
+            The number of output channels for the first conv2d layer. If None, same as `in_channels`.
+        temb_channels (`int`, *optional*, default to `512`): the number of channels in timestep embedding.
+        eps (`float`, *optional*, defaults to `1e-6`): The epsilon to use for the spatial resenet.
+        temporal_eps (`float`, *optional*, defaults to `eps`): The epsilon to use for the temporal resnet.
+        merge_factor (`float`, *optional*, defaults to `0.5`): The merge factor to use for the temporal mixing.
+        merge_strategy (`str`, *optional*, defaults to `learned_with_images`):
+            The merge strategy to use for the temporal mixing.
+        switch_spatial_to_temporal_mix (`bool`, *optional*, defaults to `False`):
+            If `True`, switch the spatial and temporal mixing.
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: Optional[int] = None,
+        temb_channels: int = 512,
+        eps: float = 1e-6,
+        temporal_eps: Optional[float] = None,
+        merge_factor: float = 0.5,
+        merge_strategy="learned_with_images",
+        switch_spatial_to_temporal_mix: bool = False,
+    ):
+        super().__init__()
+
+        self.spatial_res_block = ResnetBlock2D(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            eps=eps,
+        )
+
+        self.temporal_res_block = TemporalResnetBlock(
+            in_channels=out_channels if out_channels is not None else in_channels,
+            out_channels=out_channels if out_channels is not None else in_channels,
+            temb_channels=temb_channels,
+            eps=temporal_eps if temporal_eps is not None else eps,
+        )
+
+        self.time_mixer = AlphaBlender(
+            alpha=merge_factor,
+            merge_strategy=merge_strategy,
+            switch_spatial_to_temporal_mix=switch_spatial_to_temporal_mix,
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        temb: Optional[torch.Tensor] = None,
+        image_only_indicator: Optional[torch.Tensor] = None,
+    ):
+        num_frames = image_only_indicator.shape[-1]
+        hidden_states = self.spatial_res_block(hidden_states, temb)
+
+        batch_frames, channels, height, width = hidden_states.shape
+        batch_size = batch_frames // num_frames
+
+        hidden_states_mix = (
+            hidden_states[None, :].reshape(batch_size, num_frames, channels, height, width).permute(0, 2, 1, 3, 4)
+        )
+        hidden_states = (
+            hidden_states[None, :].reshape(batch_size, num_frames, channels, height, width).permute(0, 2, 1, 3, 4)
+        )
+
+        if temb is not None:
+            temb = temb.reshape(batch_size, num_frames, -1)
+
+        hidden_states = self.temporal_res_block(hidden_states, temb)
+        hidden_states = self.time_mixer(
+            x_spatial=hidden_states_mix,
+            x_temporal=hidden_states,
+            image_only_indicator=image_only_indicator,
+        )
+
+        hidden_states = hidden_states.permute(0, 2, 1, 3, 4).reshape(batch_frames, channels, height, width)
+        return hidden_states
+
+
+class AlphaBlender(nn.Module):
+    r"""
+    A module to blend spatial and temporal features.
+
+    Parameters:
+        alpha (`float`): The initial value of the blending factor.
+        merge_strategy (`str`, *optional*, defaults to `learned_with_images`):
+            The merge strategy to use for the temporal mixing.
+        switch_spatial_to_temporal_mix (`bool`, *optional*, defaults to `False`):
+            If `True`, switch the spatial and temporal mixing.
+    """
+
+    strategies = ["learned", "fixed", "learned_with_images"]
+
+    def __init__(
+        self,
+        alpha: float,
+        merge_strategy: str = "learned_with_images",
+        switch_spatial_to_temporal_mix: bool = False,
+    ):
+        super().__init__()
+        self.merge_strategy = merge_strategy
+        self.switch_spatial_to_temporal_mix = switch_spatial_to_temporal_mix  # For TemporalVAE
+
+        if merge_strategy not in self.strategies:
+            raise ValueError(f"merge_strategy needs to be in {self.strategies}")
+
+        if self.merge_strategy == "fixed":
+            self.register_buffer("mix_factor", torch.Tensor([alpha]))
+        elif self.merge_strategy == "learned" or self.merge_strategy == "learned_with_images":
+            self.register_parameter("mix_factor", torch.nn.Parameter(torch.Tensor([alpha])))
+        else:
+            raise ValueError(f"Unknown merge strategy {self.merge_strategy}")
+
+    def get_alpha(self, image_only_indicator: torch.Tensor, ndims: int) -> torch.Tensor:
+        if self.merge_strategy == "fixed":
+            alpha = self.mix_factor
+
+        elif self.merge_strategy == "learned":
+            alpha = torch.sigmoid(self.mix_factor)
+
+        elif self.merge_strategy == "learned_with_images":
+            if image_only_indicator is None:
+                raise ValueError("Please provide image_only_indicator to use learned_with_images merge strategy")
+
+            alpha = torch.where(
+                image_only_indicator.bool(),
+                torch.ones(1, 1, device=image_only_indicator.device),
+                torch.sigmoid(self.mix_factor)[..., None],
+            )
+
+            # (batch, channel, frames, height, width)
+            if ndims == 5:
+                alpha = alpha[:, None, :, None, None]
+            # (batch*frames, height*width, channels)
+            elif ndims == 3:
+                alpha = alpha.reshape(-1)[:, None, None]
+            else:
+                raise ValueError(f"Unexpected ndims {ndims}. Dimensions should be 3 or 5")
+
+        else:
+            raise NotImplementedError
+
+        return alpha
+
+    def forward(
+        self,
+        x_spatial: torch.Tensor,
+        x_temporal: torch.Tensor,
+        image_only_indicator: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        alpha = self.get_alpha(image_only_indicator, x_spatial.ndim)
+        alpha = alpha.to(x_spatial.dtype)
+
+        if self.switch_spatial_to_temporal_mix:
+            alpha = 1.0 - alpha
+
+        x = alpha * x_spatial + (1.0 - alpha) * x_temporal
+        return x

Marigold/resnet/upsampling.py

@@ -0,0 +1,515 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from diffusers.utils import deprecate
+from diffusers.models.normalization import RMSNorm
+from einops import rearrange
+
+
+class Upsample1D(nn.Module):
+    """A 1D upsampling layer with an optional convolution.
+
+    Parameters:
+        channels (`int`):
+            number of channels in the inputs and outputs.
+        use_conv (`bool`, default `False`):
+            option to use a convolution.
+        use_conv_transpose (`bool`, default `False`):
+            option to use a convolution transpose.
+        out_channels (`int`, optional):
+            number of output channels. Defaults to `channels`.
+        name (`str`, default `conv`):
+            name of the upsampling 1D layer.
+    """
+
+    def __init__(
+        self,
+        channels: int,
+        use_conv: bool = False,
+        use_conv_transpose: bool = False,
+        out_channels: Optional[int] = None,
+        name: str = "conv",
+    ):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.use_conv_transpose = use_conv_transpose
+        self.name = name
+
+        self.conv = None
+        if use_conv_transpose:
+            self.conv = nn.ConvTranspose1d(channels, self.out_channels, 4, 2, 1)
+        elif use_conv:
+            self.conv = nn.Conv1d(self.channels, self.out_channels, 3, padding=1)
+
+    def forward(self, inputs: torch.Tensor) -> torch.Tensor:
+        assert inputs.shape[1] == self.channels
+        if self.use_conv_transpose:
+            return self.conv(inputs)
+
+        outputs = F.interpolate(inputs, scale_factor=2.0, mode="nearest")
+
+        if self.use_conv:
+            outputs = self.conv(outputs)
+
+        return outputs
+
+
+class Upsample2D(nn.Module):
+    """A 2D upsampling layer with an optional convolution.
+
+    Parameters:
+        channels (`int`):
+            number of channels in the inputs and outputs.
+        use_conv (`bool`, default `False`):
+            option to use a convolution.
+        use_conv_transpose (`bool`, default `False`):
+            option to use a convolution transpose.
+        out_channels (`int`, optional):
+            number of output channels. Defaults to `channels`.
+        name (`str`, default `conv`):
+            name of the upsampling 2D layer.
+    """
+
+    def __init__(
+        self,
+        channels: int,
+        use_conv: bool = False,
+        use_conv_transpose: bool = False,
+        out_channels: Optional[int] = None,
+        name: str = "conv",
+        kernel_size: Optional[int] = None,
+        padding=1,
+        norm_type=None,
+        eps=None,
+        elementwise_affine=None,
+        bias=True,
+        interpolate=True,
+    ):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.use_conv_transpose = use_conv_transpose
+        self.name = name
+        self.interpolate = interpolate
+
+        if norm_type == "ln_norm":
+            self.norm = nn.LayerNorm(channels, eps, elementwise_affine)
+        elif norm_type == "rms_norm":
+            self.norm = RMSNorm(channels, eps, elementwise_affine)
+        elif norm_type is None:
+            self.norm = None
+        else:
+            raise ValueError(f"unknown norm_type: {norm_type}")
+
+        conv = None
+        if use_conv_transpose:
+            if kernel_size is None:
+                kernel_size = 4
+            conv = nn.ConvTranspose2d(
+                channels, self.out_channels, kernel_size=kernel_size, stride=2, padding=padding, bias=bias
+            )
+        elif use_conv:
+            if kernel_size is None:
+                kernel_size = 3
+            conv = nn.Conv2d(self.channels, self.out_channels, kernel_size=kernel_size, padding=padding, bias=bias)
+
+        # TODO(Suraj, Patrick) - clean up after weight dicts are correctly renamed
+        if name == "conv":
+            self.conv = conv
+        else:
+            self.Conv2d_0 = conv
+
+    def forward(self, hidden_states: torch.Tensor, output_size: Optional[int] = None, *args, **kwargs) -> torch.Tensor:
+        if len(args) > 0 or kwargs.get("scale", None) is not None:
+            deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`."
+            deprecate("scale", "1.0.0", deprecation_message)
+
+        assert hidden_states.shape[1] == self.channels
+
+        if self.norm is not None:
+            hidden_states_permuted = hidden_states.permute(0, 2, 3, 1)  # [N, C, H, W] -> [N, H, W, C]
+            b, c, h, w = hidden_states_permuted.shape
+            hidden_states_permuted = rearrange(hidden_states_permuted, "(b t) c h w -> b c (h w t)", b=1, h=h, w=w)
+            hidden_states = self.norm(hidden_states_permuted)
+            hidden_states = rearrange(hidden_states, "b c (h w t) -> (b t) c h w", b=1, h=h, w=w)
+            hidden_states = hidden_states.permute(0, 3, 1, 2)  # [N, H, W, C] -> [N, C, H, W]
+
+        if self.use_conv_transpose:
+            return self.conv(hidden_states)
+
+        # Cast to float32 to as 'upsample_nearest2d_out_frame' op does not support bfloat16
+        # TODO(Suraj): Remove this cast once the issue is fixed in PyTorch
+        # https://github.com/pytorch/pytorch/issues/86679
+        dtype = hidden_states.dtype
+        if dtype == torch.bfloat16:
+            hidden_states = hidden_states.to(torch.float32)
+
+        # upsample_nearest_nhwc fails with large batch sizes. see https://github.com/huggingface/diffusers/issues/984
+        if hidden_states.shape[0] >= 64:
+            hidden_states = hidden_states.contiguous()
+
+        # if `output_size` is passed we force the interpolation output
+        # size and do not make use of `scale_factor=2`
+        if self.interpolate:
+            if output_size is None:
+                hidden_states = F.interpolate(hidden_states, scale_factor=2.0, mode="nearest")
+            else:
+                hidden_states = F.interpolate(hidden_states, size=output_size, mode="nearest")
+
+        # If the input is bfloat16, we cast back to bfloat16
+        if dtype == torch.bfloat16:
+            hidden_states = hidden_states.to(dtype)
+
+        # TODO(Suraj, Patrick) - clean up after weight dicts are correctly renamed
+        if self.use_conv:
+            if self.name == "conv":
+                hidden_states = self.conv(hidden_states)
+            else:
+                hidden_states = self.Conv2d_0(hidden_states)
+
+        return hidden_states
+
+
+class FirUpsample2D(nn.Module):
+    """A 2D FIR upsampling layer with an optional convolution.
+
+    Parameters:
+        channels (`int`, optional):
+            number of channels in the inputs and outputs.
+        use_conv (`bool`, default `False`):
+            option to use a convolution.
+        out_channels (`int`, optional):
+            number of output channels. Defaults to `channels`.
+        fir_kernel (`tuple`, default `(1, 3, 3, 1)`):
+            kernel for the FIR filter.
+    """
+
+    def __init__(
+        self,
+        channels: Optional[int] = None,
+        out_channels: Optional[int] = None,
+        use_conv: bool = False,
+        fir_kernel: Tuple[int, int, int, int] = (1, 3, 3, 1),
+    ):
+        super().__init__()
+        out_channels = out_channels if out_channels else channels
+        if use_conv:
+            self.Conv2d_0 = nn.Conv2d(channels, out_channels, kernel_size=3, stride=1, padding=1)
+        self.use_conv = use_conv
+        self.fir_kernel = fir_kernel
+        self.out_channels = out_channels
+
+    def _upsample_2d(
+        self,
+        hidden_states: torch.Tensor,
+        weight: Optional[torch.Tensor] = None,
+        kernel: Optional[torch.Tensor] = None,
+        factor: int = 2,
+        gain: float = 1,
+    ) -> torch.Tensor:
+        """Fused `upsample_2d()` followed by `Conv2d()`.
+
+        Padding is performed only once at the beginning, not between the operations. The fused op is considerably more
+        efficient than performing the same calculation using standard TensorFlow ops. It supports gradients of
+        arbitrary order.
+
+        Args:
+            hidden_states (`torch.Tensor`):
+                Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
+            weight (`torch.Tensor`, *optional*):
+                Weight tensor of the shape `[filterH, filterW, inChannels, outChannels]`. Grouped convolution can be
+                performed by `inChannels = x.shape[0] // numGroups`.
+            kernel (`torch.Tensor`, *optional*):
+                FIR filter of the shape `[firH, firW]` or `[firN]` (separable). The default is `[1] * factor`, which
+                corresponds to nearest-neighbor upsampling.
+            factor (`int`, *optional*): Integer upsampling factor (default: 2).
+            gain (`float`, *optional*): Scaling factor for signal magnitude (default: 1.0).
+
+        Returns:
+            output (`torch.Tensor`):
+                Tensor of the shape `[N, C, H * factor, W * factor]` or `[N, H * factor, W * factor, C]`, and same
+                datatype as `hidden_states`.
+        """
+
+        assert isinstance(factor, int) and factor >= 1
+
+        # Setup filter kernel.
+        if kernel is None:
+            kernel = [1] * factor
+
+        # setup kernel
+        kernel = torch.tensor(kernel, dtype=torch.float32)
+        if kernel.ndim == 1:
+            kernel = torch.outer(kernel, kernel)
+        kernel /= torch.sum(kernel)
+
+        kernel = kernel * (gain * (factor**2))
+
+        if self.use_conv:
+            convH = weight.shape[2]
+            convW = weight.shape[3]
+            inC = weight.shape[1]
+
+            pad_value = (kernel.shape[0] - factor) - (convW - 1)
+
+            stride = (factor, factor)
+            # Determine data dimensions.
+            output_shape = (
+                (hidden_states.shape[2] - 1) * factor + convH,
+                (hidden_states.shape[3] - 1) * factor + convW,
+            )
+            output_padding = (
+                output_shape[0] - (hidden_states.shape[2] - 1) * stride[0] - convH,
+                output_shape[1] - (hidden_states.shape[3] - 1) * stride[1] - convW,
+            )
+            assert output_padding[0] >= 0 and output_padding[1] >= 0
+            num_groups = hidden_states.shape[1] // inC
+
+            # Transpose weights.
+            weight = torch.reshape(weight, (num_groups, -1, inC, convH, convW))
+            weight = torch.flip(weight, dims=[3, 4]).permute(0, 2, 1, 3, 4)
+            weight = torch.reshape(weight, (num_groups * inC, -1, convH, convW))
+
+            inverse_conv = F.conv_transpose2d(
+                hidden_states,
+                weight,
+                stride=stride,
+                output_padding=output_padding,
+                padding=0,
+            )
+
+            output = upfirdn2d_native(
+                inverse_conv,
+                torch.tensor(kernel, device=inverse_conv.device),
+                pad=((pad_value + 1) // 2 + factor - 1, pad_value // 2 + 1),
+            )
+        else:
+            pad_value = kernel.shape[0] - factor
+            output = upfirdn2d_native(
+                hidden_states,
+                torch.tensor(kernel, device=hidden_states.device),
+                up=factor,
+                pad=((pad_value + 1) // 2 + factor - 1, pad_value // 2),
+            )
+
+        return output
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        if self.use_conv:
+            height = self._upsample_2d(hidden_states, self.Conv2d_0.weight, kernel=self.fir_kernel)
+            height = height + self.Conv2d_0.bias.reshape(1, -1, 1, 1)
+        else:
+            height = self._upsample_2d(hidden_states, kernel=self.fir_kernel, factor=2)
+
+        return height
+
+
+class KUpsample2D(nn.Module):
+    r"""A 2D K-upsampling layer.
+
+    Parameters:
+        pad_mode (`str`, *optional*, default to `"reflect"`): the padding mode to use.
+    """
+
+    def __init__(self, pad_mode: str = "reflect"):
+        super().__init__()
+        self.pad_mode = pad_mode
+        kernel_1d = torch.tensor([[1 / 8, 3 / 8, 3 / 8, 1 / 8]]) * 2
+        self.pad = kernel_1d.shape[1] // 2 - 1
+        self.register_buffer("kernel", kernel_1d.T @ kernel_1d, persistent=False)
+
+    def forward(self, inputs: torch.Tensor) -> torch.Tensor:
+        inputs = F.pad(inputs, ((self.pad + 1) // 2,) * 4, self.pad_mode)
+        weight = inputs.new_zeros(
+            [
+                inputs.shape[1],
+                inputs.shape[1],
+                self.kernel.shape[0],
+                self.kernel.shape[1],
+            ]
+        )
+        indices = torch.arange(inputs.shape[1], device=inputs.device)
+        kernel = self.kernel.to(weight)[None, :].expand(inputs.shape[1], -1, -1)
+        weight[indices, indices] = kernel
+        return F.conv_transpose2d(inputs, weight, stride=2, padding=self.pad * 2 + 1)
+
+
+class CogVideoXUpsample3D(nn.Module):
+    r"""
+    A 3D Upsample layer using in CogVideoX by Tsinghua University & ZhipuAI # Todo: Wait for paper relase.
+
+    Args:
+        in_channels (`int`):
+            Number of channels in the input image.
+        out_channels (`int`):
+            Number of channels produced by the convolution.
+        kernel_size (`int`, defaults to `3`):
+            Size of the convolving kernel.
+        stride (`int`, defaults to `1`):
+            Stride of the convolution.
+        padding (`int`, defaults to `1`):
+            Padding added to all four sides of the input.
+        compress_time (`bool`, defaults to `False`):
+            Whether or not to compress the time dimension.
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        kernel_size: int = 3,
+        stride: int = 1,
+        padding: int = 1,
+        compress_time: bool = False,
+    ) -> None:
+        super().__init__()
+
+        self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=padding)
+        self.compress_time = compress_time
+
+    def forward(self, inputs: torch.Tensor) -> torch.Tensor:
+        if self.compress_time:
+            if inputs.shape[2] > 1 and inputs.shape[2] % 2 == 1:
+                # split first frame
+                x_first, x_rest = inputs[:, :, 0], inputs[:, :, 1:]
+
+                x_first = F.interpolate(x_first, scale_factor=2.0)
+                x_rest = F.interpolate(x_rest, scale_factor=2.0)
+                x_first = x_first[:, :, None, :, :]
+                inputs = torch.cat([x_first, x_rest], dim=2)
+            elif inputs.shape[2] > 1:
+                inputs = F.interpolate(inputs, scale_factor=2.0)
+            else:
+                inputs = inputs.squeeze(2)
+                inputs = F.interpolate(inputs, scale_factor=2.0)
+                inputs = inputs[:, :, None, :, :]
+        else:
+            # only interpolate 2D
+            b, c, t, h, w = inputs.shape
+            inputs = inputs.permute(0, 2, 1, 3, 4).reshape(b * t, c, h, w)
+            inputs = F.interpolate(inputs, scale_factor=2.0)
+            inputs = inputs.reshape(b, t, c, *inputs.shape[2:]).permute(0, 2, 1, 3, 4)
+
+        b, c, t, h, w = inputs.shape
+        inputs = inputs.permute(0, 2, 1, 3, 4).reshape(b * t, c, h, w)
+        inputs = self.conv(inputs)
+        inputs = inputs.reshape(b, t, *inputs.shape[1:]).permute(0, 2, 1, 3, 4)
+
+        return inputs
+
+
+def upfirdn2d_native(
+    tensor: torch.Tensor,
+    kernel: torch.Tensor,
+    up: int = 1,
+    down: int = 1,
+    pad: Tuple[int, int] = (0, 0),
+) -> torch.Tensor:
+    up_x = up_y = up
+    down_x = down_y = down
+    pad_x0 = pad_y0 = pad[0]
+    pad_x1 = pad_y1 = pad[1]
+
+    _, channel, in_h, in_w = tensor.shape
+    tensor = tensor.reshape(-1, in_h, in_w, 1)
+
+    _, in_h, in_w, minor = tensor.shape
+    kernel_h, kernel_w = kernel.shape
+
+    out = tensor.view(-1, in_h, 1, in_w, 1, minor)
+    out = F.pad(out, [0, 0, 0, up_x - 1, 0, 0, 0, up_y - 1])
+    out = out.view(-1, in_h * up_y, in_w * up_x, minor)
+
+    out = F.pad(out, [0, 0, max(pad_x0, 0), max(pad_x1, 0), max(pad_y0, 0), max(pad_y1, 0)])
+    out = out.to(tensor.device)  # Move back to mps if necessary
+    out = out[
+        :,
+        max(-pad_y0, 0) : out.shape[1] - max(-pad_y1, 0),
+        max(-pad_x0, 0) : out.shape[2] - max(-pad_x1, 0),
+        :,
+    ]
+
+    out = out.permute(0, 3, 1, 2)
+    out = out.reshape([-1, 1, in_h * up_y + pad_y0 + pad_y1, in_w * up_x + pad_x0 + pad_x1])
+    w = torch.flip(kernel, [0, 1]).view(1, 1, kernel_h, kernel_w)
+    out = F.conv2d(out, w)
+    out = out.reshape(
+        -1,
+        minor,
+        in_h * up_y + pad_y0 + pad_y1 - kernel_h + 1,
+        in_w * up_x + pad_x0 + pad_x1 - kernel_w + 1,
+    )
+    out = out.permute(0, 2, 3, 1)
+    out = out[:, ::down_y, ::down_x, :]
+
+    out_h = (in_h * up_y + pad_y0 + pad_y1 - kernel_h) // down_y + 1
+    out_w = (in_w * up_x + pad_x0 + pad_x1 - kernel_w) // down_x + 1
+
+    return out.view(-1, channel, out_h, out_w)
+
+
+def upsample_2d(
+    hidden_states: torch.Tensor,
+    kernel: Optional[torch.Tensor] = None,
+    factor: int = 2,
+    gain: float = 1,
+) -> torch.Tensor:
+    r"""Upsample2D a batch of 2D images with the given filter.
+    Accepts a batch of 2D images of the shape `[N, C, H, W]` or `[N, H, W, C]` and upsamples each image with the given
+    filter. The filter is normalized so that if the input pixels are constant, they will be scaled by the specified
+    `gain`. Pixels outside the image are assumed to be zero, and the filter is padded with zeros so that its shape is
+    a: multiple of the upsampling factor.
+
+    Args:
+        hidden_states (`torch.Tensor`):
+            Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
+        kernel (`torch.Tensor`, *optional*):
+            FIR filter of the shape `[firH, firW]` or `[firN]` (separable). The default is `[1] * factor`, which
+            corresponds to nearest-neighbor upsampling.
+        factor (`int`, *optional*, default to `2`):
+            Integer upsampling factor.
+        gain (`float`, *optional*, default to `1.0`):
+            Scaling factor for signal magnitude (default: 1.0).
+
+    Returns:
+        output (`torch.Tensor`):
+            Tensor of the shape `[N, C, H * factor, W * factor]`
+    """
+    assert isinstance(factor, int) and factor >= 1
+    if kernel is None:
+        kernel = [1] * factor
+
+    kernel = torch.tensor(kernel, dtype=torch.float32)
+    if kernel.ndim == 1:
+        kernel = torch.outer(kernel, kernel)
+    kernel /= torch.sum(kernel)
+
+    kernel = kernel * (gain * (factor**2))
+    pad_value = kernel.shape[0] - factor
+    output = upfirdn2d_native(
+        hidden_states,
+        kernel.to(device=hidden_states.device),
+        up=factor,
+        pad=((pad_value + 1) // 2 + factor - 1, pad_value // 2),
+    )
+    return output

Marigold/unet/attention.py

@@ -0,0 +1,1342 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Any, Dict, List, Optional, Tuple
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from diffusers.utils import deprecate, logging
+from diffusers.utils.torch_utils import maybe_allow_in_graph
+from diffusers.models.activations import GEGLU, GELU, ApproximateGELU, FP32SiLU, SwiGLU
+from diffusers.models.attention_processor import JointAttnProcessor2_0
+from diffusers.models.embeddings import SinusoidalPositionalEmbedding
+from diffusers.models.normalization import AdaLayerNorm, AdaLayerNormContinuous, AdaLayerNormZero, RMSNorm
+
+from Marigold.unet.attention_processor import Attention
+from einops import rearrange
+
+logger = logging.get_logger(__name__)
+
+
+def _chunked_feed_forward(ff: nn.Module, hidden_states: torch.Tensor, chunk_dim: int, chunk_size: int):
+    # "feed_forward_chunk_size" can be used to save memory
+    if hidden_states.shape[chunk_dim] % chunk_size != 0:
+        raise ValueError(
+            f"`hidden_states` dimension to be chunked: {hidden_states.shape[chunk_dim]} has to be divisible by chunk size: {chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`."
+        )
+
+    num_chunks = hidden_states.shape[chunk_dim] // chunk_size
+    ff_output = torch.cat(
+        [ff(hid_slice) for hid_slice in hidden_states.chunk(num_chunks, dim=chunk_dim)],
+        dim=chunk_dim,
+    )
+    return ff_output
+
+
+@maybe_allow_in_graph
+class GatedSelfAttentionDense(nn.Module):
+    r"""
+    A gated self-attention dense layer that combines visual features and object features.
+
+    Parameters:
+        query_dim (`int`): The number of channels in the query.
+        context_dim (`int`): The number of channels in the context.
+        n_heads (`int`): The number of heads to use for attention.
+        d_head (`int`): The number of channels in each head.
+    """
+
+    def __init__(self, query_dim: int, context_dim: int, n_heads: int, d_head: int):
+        super().__init__()
+
+        # we need a linear projection since we need cat visual feature and obj feature
+        self.linear = nn.Linear(context_dim, query_dim)
+
+        self.attn = Attention(query_dim=query_dim, heads=n_heads, dim_head=d_head)
+        self.ff = FeedForward(query_dim, activation_fn="geglu")
+
+        self.norm1 = nn.LayerNorm(query_dim)
+        self.norm2 = nn.LayerNorm(query_dim)
+
+        self.register_parameter("alpha_attn", nn.Parameter(torch.tensor(0.0)))
+        self.register_parameter("alpha_dense", nn.Parameter(torch.tensor(0.0)))
+
+        self.enabled = True
+
+    def forward(self, x: torch.Tensor, objs: torch.Tensor) -> torch.Tensor:
+        if not self.enabled:
+            return x
+
+        n_visual = x.shape[1]
+        objs = self.linear(objs)
+
+        x = x + self.alpha_attn.tanh() * self.attn(self.norm1(torch.cat([x, objs], dim=1)))[:, :n_visual, :]
+        x = x + self.alpha_dense.tanh() * self.ff(self.norm2(x))
+
+        return x
+
+
+@maybe_allow_in_graph
+class JointTransformerBlock(nn.Module):
+    r"""
+    A Transformer block following the MMDiT architecture, introduced in Stable Diffusion 3.
+
+    Reference: https://arxiv.org/abs/2403.03206
+
+    Parameters:
+        dim (`int`): The number of channels in the input and output.
+        num_attention_heads (`int`): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`): The number of channels in each head.
+        context_pre_only (`bool`): Boolean to determine if we should add some blocks associated with the
+            processing of `context` conditions.
+    """
+
+    def __init__(self, dim, num_attention_heads, attention_head_dim, context_pre_only=False):
+        super().__init__()
+
+        self.context_pre_only = context_pre_only
+        context_norm_type = "ada_norm_continous" if context_pre_only else "ada_norm_zero"
+
+        self.norm1 = AdaLayerNormZero(dim)
+
+        if context_norm_type == "ada_norm_continous":
+            self.norm1_context = AdaLayerNormContinuous(
+                dim, dim, elementwise_affine=False, eps=1e-6, bias=True, norm_type="layer_norm"
+            )
+        elif context_norm_type == "ada_norm_zero":
+            self.norm1_context = AdaLayerNormZero(dim)
+        else:
+            raise ValueError(
+                f"Unknown context_norm_type: {context_norm_type}, currently only support `ada_norm_continous`, `ada_norm_zero`"
+            )
+        if hasattr(F, "scaled_dot_product_attention"):
+            processor = JointAttnProcessor2_0()
+        else:
+            raise ValueError(
+                "The current PyTorch version does not support the `scaled_dot_product_attention` function."
+            )
+        self.attn = Attention(
+            query_dim=dim,
+            cross_attention_dim=None,
+            added_kv_proj_dim=dim,
+            dim_head=attention_head_dim,
+            heads=num_attention_heads,
+            out_dim=dim,
+            context_pre_only=context_pre_only,
+            bias=True,
+            processor=processor,
+        )
+
+        self.norm2 = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-6)
+        self.ff = FeedForward(dim=dim, dim_out=dim, activation_fn="gelu-approximate")
+
+        if not context_pre_only:
+            self.norm2_context = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-6)
+            self.ff_context = FeedForward(dim=dim, dim_out=dim, activation_fn="gelu-approximate")
+        else:
+            self.norm2_context = None
+            self.ff_context = None
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = 0
+
+    # Copied from diffusers.models.attention.BasicTransformerBlock.set_chunk_feed_forward
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0):
+        # Sets chunk feed-forward
+        self._chunk_size = chunk_size
+        self._chunk_dim = dim
+
+    def forward(
+        self, hidden_states: torch.FloatTensor, encoder_hidden_states: torch.FloatTensor, temb: torch.FloatTensor
+    ):
+        norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(hidden_states, emb=temb)
+
+        if self.context_pre_only:
+            norm_encoder_hidden_states = self.norm1_context(encoder_hidden_states, temb)
+        else:
+            norm_encoder_hidden_states, c_gate_msa, c_shift_mlp, c_scale_mlp, c_gate_mlp = self.norm1_context(
+                encoder_hidden_states, emb=temb
+            )
+
+        # Attention.
+        attn_output, context_attn_output = self.attn(
+            hidden_states=norm_hidden_states, encoder_hidden_states=norm_encoder_hidden_states
+        )
+
+        # Process attention outputs for the `hidden_states`.
+        attn_output = gate_msa.unsqueeze(1) * attn_output
+        hidden_states = hidden_states + attn_output
+
+        norm_hidden_states = self.norm2(hidden_states)
+        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
+            ff_output = _chunked_feed_forward(self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size)
+        else:
+            ff_output = self.ff(norm_hidden_states)
+        ff_output = gate_mlp.unsqueeze(1) * ff_output
+
+        hidden_states = hidden_states + ff_output
+
+        # Process attention outputs for the `encoder_hidden_states`.
+        if self.context_pre_only:
+            encoder_hidden_states = None
+        else:
+            context_attn_output = c_gate_msa.unsqueeze(1) * context_attn_output
+            encoder_hidden_states = encoder_hidden_states + context_attn_output
+
+            norm_encoder_hidden_states = self.norm2_context(encoder_hidden_states)
+            norm_encoder_hidden_states = norm_encoder_hidden_states * (1 + c_scale_mlp[:, None]) + c_shift_mlp[:, None]
+            if self._chunk_size is not None:
+                # "feed_forward_chunk_size" can be used to save memory
+                context_ff_output = _chunked_feed_forward(
+                    self.ff_context, norm_encoder_hidden_states, self._chunk_dim, self._chunk_size
+                )
+            else:
+                context_ff_output = self.ff_context(norm_encoder_hidden_states)
+            encoder_hidden_states = encoder_hidden_states + c_gate_mlp.unsqueeze(1) * context_ff_output
+
+        return encoder_hidden_states, hidden_states
+
+
+
+@maybe_allow_in_graph
+class BasicTransformerBlock(nn.Module):
+    r"""
+    A basic Transformer block.
+
+    Parameters:
+        dim (`int`): The number of channels in the input and output.
+        num_attention_heads (`int`): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`): The number of channels in each head.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        num_embeds_ada_norm (:
+            obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`.
+        attention_bias (:
+            obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
+        only_cross_attention (`bool`, *optional*):
+            Whether to use only cross-attention layers. In this case two cross attention layers are used.
+        double_self_attention (`bool`, *optional*):
+            Whether to use two self-attention layers. In this case no cross attention layers are used.
+        upcast_attention (`bool`, *optional*):
+            Whether to upcast the attention computation to float32. This is useful for mixed precision training.
+        norm_elementwise_affine (`bool`, *optional*, defaults to `True`):
+            Whether to use learnable elementwise affine parameters for normalization.
+        norm_type (`str`, *optional*, defaults to `"layer_norm"`):
+            The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`.
+        final_dropout (`bool` *optional*, defaults to False):
+            Whether to apply a final dropout after the last feed-forward layer.
+        attention_type (`str`, *optional*, defaults to `"default"`):
+            The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`.
+        positional_embeddings (`str`, *optional*, defaults to `None`):
+            The type of positional embeddings to apply to.
+        num_positional_embeddings (`int`, *optional*, defaults to `None`):
+            The maximum number of positional embeddings to apply.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        attention_bias: bool = False,
+        only_cross_attention: bool = False,
+        double_self_attention: bool = False,
+        upcast_attention: bool = False,
+        norm_elementwise_affine: bool = True,
+        norm_type: str = "layer_norm",  # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single', 'ada_norm_continuous', 'layer_norm_i2vgen'
+        norm_eps: float = 1e-5,
+        final_dropout: bool = False,
+        attention_type: str = "default",
+        positional_embeddings: Optional[str] = None,
+        num_positional_embeddings: Optional[int] = None,
+        ada_norm_continous_conditioning_embedding_dim: Optional[int] = None,
+        ada_norm_bias: Optional[int] = None,
+        ff_inner_dim: Optional[int] = None,
+        ff_bias: bool = True,
+        attention_out_bias: bool = True,
+        use_RoPE: bool = False,
+    ):
+        super().__init__()
+        self.dim = dim
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        self.dropout = dropout
+        self.cross_attention_dim = cross_attention_dim
+        self.activation_fn = activation_fn
+        self.attention_bias = attention_bias
+        self.double_self_attention = double_self_attention
+        self.norm_elementwise_affine = norm_elementwise_affine
+        self.positional_embeddings = positional_embeddings
+        self.num_positional_embeddings = num_positional_embeddings
+        self.only_cross_attention = only_cross_attention
+        self.use_RoPE = use_RoPE
+
+        # We keep these boolean flags for backward-compatibility.
+        self.use_ada_layer_norm_zero = (num_embeds_ada_norm is not None) and norm_type == "ada_norm_zero"
+        self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm"
+        self.use_ada_layer_norm_single = norm_type == "ada_norm_single"
+        self.use_layer_norm = norm_type == "layer_norm"
+        self.use_ada_layer_norm_continuous = norm_type == "ada_norm_continuous"
+
+        if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
+            raise ValueError(
+                f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
+                f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
+            )
+
+        self.norm_type = norm_type
+        self.num_embeds_ada_norm = num_embeds_ada_norm
+
+        if positional_embeddings and (num_positional_embeddings is None):
+            raise ValueError(
+                "If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined."
+            )
+
+        if positional_embeddings == "sinusoidal":
+            self.pos_embed = SinusoidalPositionalEmbedding(dim, max_seq_length=num_positional_embeddings)
+        else:
+            self.pos_embed = None
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Self-Attn
+        if norm_type == "ada_norm":
+            self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
+        elif norm_type == "ada_norm_zero":
+            self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
+        elif norm_type == "ada_norm_continuous":
+            self.norm1 = AdaLayerNormContinuous(
+                dim,
+                ada_norm_continous_conditioning_embedding_dim,
+                norm_elementwise_affine,
+                norm_eps,
+                ada_norm_bias,
+                "rms_norm",
+            )
+        else:
+            self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
+
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            cross_attention_dim=cross_attention_dim if only_cross_attention else None,
+            upcast_attention=upcast_attention,
+            out_bias=attention_out_bias,
+            use_RoPE=use_RoPE,
+        )
+
+        # 2. Cross-Attn
+        if cross_attention_dim is not None or double_self_attention:
+            # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
+            # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
+            # the second cross attention block.
+            if norm_type == "ada_norm":
+                self.norm2 = AdaLayerNorm(dim, num_embeds_ada_norm)
+            elif norm_type == "ada_norm_continuous":
+                self.norm2 = AdaLayerNormContinuous(
+                    dim,
+                    ada_norm_continous_conditioning_embedding_dim,
+                    norm_elementwise_affine,
+                    norm_eps,
+                    ada_norm_bias,
+                    "rms_norm",
+                )
+            else:
+                self.norm2 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)
+
+            self.attn2 = Attention(
+                query_dim=dim,
+                cross_attention_dim=cross_attention_dim if not double_self_attention else None,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+                out_bias=attention_out_bias,
+                use_RoPE=use_RoPE,
+            )  # is self-attn if encoder_hidden_states is none
+        else:
+            if norm_type == "ada_norm_single":  # For Latte
+                self.norm2 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)
+            else:
+                self.norm2 = None
+            self.attn2 = None
+
+        # 3. Feed-forward
+        if norm_type == "ada_norm_continuous":
+            self.norm3 = AdaLayerNormContinuous(
+                dim,
+                ada_norm_continous_conditioning_embedding_dim,
+                norm_elementwise_affine,
+                norm_eps,
+                ada_norm_bias,
+                "layer_norm",
+            )
+
+        elif norm_type in ["ada_norm_zero", "ada_norm", "layer_norm"]:
+            self.norm3 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)
+        elif norm_type == "layer_norm_i2vgen":
+            self.norm3 = None
+
+        self.ff = FeedForward(
+            dim,
+            dropout=dropout,
+            activation_fn=activation_fn,
+            final_dropout=final_dropout,
+            inner_dim=ff_inner_dim,
+            bias=ff_bias,
+        )
+
+        # 4. Fuser
+        if attention_type == "gated" or attention_type == "gated-text-image":
+            self.fuser = GatedSelfAttentionDense(dim, cross_attention_dim, num_attention_heads, attention_head_dim)
+
+        # 5. Scale-shift for PixArt-Alpha.
+        if norm_type == "ada_norm_single":
+            self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5)
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = 0
+
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0):
+        # Sets chunk feed-forward
+        self._chunk_size = chunk_size
+        self._chunk_dim = dim
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        class_labels: Optional[torch.LongTensor] = None,
+        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+    ) -> torch.Tensor:
+        if cross_attention_kwargs is not None:
+            if cross_attention_kwargs.get("scale", None) is not None:
+                logger.warning("Passing `scale` to `cross_attention_kwargs` is deprecated. `scale` will be ignored.")
+
+        # Notice that normalization is always applied before the real computation in the following blocks.
+        # 0. Self-Attention
+        batch_size = hidden_states.shape[0]
+        if self.norm_type == "ada_norm":
+            norm_hidden_states = self.norm1(hidden_states, timestep)
+        elif self.norm_type == "ada_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
+            )
+        elif self.norm_type in ["layer_norm", "layer_norm_i2vgen"]:
+            norm_hidden_states = self.norm1(hidden_states)
+        elif self.norm_type == "ada_norm_continuous":
+            norm_hidden_states = self.norm1(hidden_states, added_cond_kwargs["pooled_text_emb"])
+        elif self.norm_type == "ada_norm_single":
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
+                self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1)
+            ).chunk(6, dim=1)
+            norm_hidden_states = self.norm1(hidden_states)
+            norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
+        else:
+            raise ValueError("Incorrect norm used")
+        if self.pos_embed is not None:
+            norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+        # 1. 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.norm_type == "ada_norm_zero":
+            attn_output = gate_msa.unsqueeze(1) * attn_output
+        elif self.norm_type == "ada_norm_single":
+            attn_output = gate_msa * attn_output
+
+
+        hidden_states = attn_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        # 1.2 GLIGEN Control
+        if gligen_kwargs is not None:
+            hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"])
+
+        # 3. Cross-Attention
+        if self.attn2 is not None:
+            if self.norm_type == "ada_norm":
+                norm_hidden_states = self.norm2(hidden_states, timestep)
+            elif self.norm_type in ["ada_norm_zero", "layer_norm", "layer_norm_i2vgen"]:
+                norm_hidden_states = self.norm2(hidden_states)
+            elif self.norm_type == "ada_norm_single":
+                # For PixArt norm2 isn't applied here:
+                # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103
+                norm_hidden_states = hidden_states
+            elif self.norm_type == "ada_norm_continuous":
+                norm_hidden_states = self.norm2(hidden_states, added_cond_kwargs["pooled_text_emb"])
+            else:
+                raise ValueError("Incorrect norm")
+            if self.pos_embed is not None and self.norm_type != "ada_norm_single":
+                norm_hidden_states = self.pos_embed(norm_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
+        # i2vgen doesn't have this norm 🤷‍♂️
+        if self.norm_type == "ada_norm_continuous":
+            norm_hidden_states = self.norm3(hidden_states, added_cond_kwargs["pooled_text_emb"])
+        elif not self.norm_type == "ada_norm_single":
+            norm_hidden_states = self.norm3(hidden_states)
+
+        if self.norm_type == "ada_norm_zero":
+            norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+
+        if self.norm_type == "ada_norm_single":
+            norm_hidden_states = self.norm2(hidden_states)
+            norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp
+
+        if self._chunk_size is not None:
+            # "feed_forward_chunk_size" can be used to save memory
+            ff_output = _chunked_feed_forward(self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size)
+        else:
+            ff_output = self.ff(norm_hidden_states)
+
+        if self.norm_type == "ada_norm_zero":
+            ff_output = gate_mlp.unsqueeze(1) * ff_output
+        elif self.norm_type == "ada_norm_single":
+            ff_output = gate_mlp * ff_output
+
+        hidden_states = ff_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        return hidden_states
+
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0):
+        # Sets chunk feed-forward
+        self._chunk_size = chunk_size
+        self._chunk_dim = dim
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        class_labels: Optional[torch.LongTensor] = None,
+        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+        num_cube_faces = 6,
+    ) -> torch.Tensor:
+        if cross_attention_kwargs is not None:
+            if cross_attention_kwargs.get("scale", None) is not None:
+                logger.warning("Passing `scale` to `cross_attention_kwargs` is deprecated. `scale` will be ignored.")
+
+        # Notice that normalization is always applied before the real computation in the following blocks.
+        # 0. Self-Attention
+        batch_size = hidden_states.shape[0]
+        
+        if self.norm_type == "ada_norm":
+            # sync_norm:
+            hidden_states = rearrange(hidden_states, '(b t) l c -> b (t l) c', t=6)
+            norm_hidden_states = self.norm1(hidden_states, timestep)
+            norm_hidden_states = rearrange(norm_hidden_states, 'b (t l) c -> (b t) l c', t=6)
+            hidden_states = rearrange(hidden_states, 'b (t l) c -> (b t) l c', t=6)
+        elif self.norm_type == "ada_norm_zero":
+            # sync_norm:
+            hidden_states = rearrange(hidden_states, '(b t) l c -> b (t l) c', t=6)
+            norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
+            hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
+            )
+            norm_hidden_states = rearrange(norm_hidden_states, 'b (t l) c -> (b t) l c', t=6)
+            hidden_states = rearrange(hidden_states, 'b (t l) c -> (b t) l c', t=6)
+        elif self.norm_type in ["layer_norm", "layer_norm_i2vgen"]:
+            # sync_norm:
+            hidden_states = rearrange(hidden_states, '(b t) l c -> b (t l) c', t=6)
+            norm_hidden_states = self.norm1(hidden_states)
+            norm_hidden_states = rearrange(norm_hidden_states, 'b (t l) c -> (b t) l c', t=6)
+            hidden_states = rearrange(hidden_states, 'b (t l) c -> (b t) l c', t=6)
+        elif self.norm_type == "ada_norm_continuous":
+            # sync_norm:
+            hidden_states = rearrange(hidden_states, '(b t) l c -> b (t l) c', t=6)
+            norm_hidden_states = self.norm1(hidden_states, added_cond_kwargs["pooled_text_emb"])
+            norm_hidden_states = rearrange(norm_hidden_states, 'b (t l) c -> (b t) l c', t=6)
+            hidden_states = rearrange(hidden_states, 'b (t l) c -> (b t) l c', t=6)
+        elif self.norm_type == "ada_norm_single":
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
+                self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1)
+            ).chunk(6, dim=1)
+            # sync_norm:
+            hidden_states = rearrange(hidden_states, '(b t) l c -> b (t l) c', t=6)
+            norm_hidden_states = self.norm1(hidden_states)
+            norm_hidden_states = rearrange(norm_hidden_states, 'b (t l) c -> (b t) l c', t=6)
+            hidden_states = rearrange(hidden_states, 'b (t l) c -> (b t) l c', t=6)
+            norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
+        else:
+            raise ValueError("Incorrect norm used")
+
+        if self.pos_embed is not None:
+            norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+        # 1. 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.norm_type == "ada_norm_zero":
+            attn_output = gate_msa.unsqueeze(1) * attn_output
+        elif self.norm_type == "ada_norm_single":
+            attn_output = gate_msa * attn_output
+
+        hidden_states = attn_output + hidden_states
+
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        # 1.2 GLIGEN Control
+        if gligen_kwargs is not None:
+            hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"])
+
+        # 3. Cross-Attention
+        if self.attn2 is not None: 
+            if self.norm_type == "ada_norm":
+                # sync_norm:
+                hidden_states = rearrange(hidden_states, '(b t) l c -> b (t l) c', t=6)
+                norm_hidden_states = self.norm2(hidden_states, timestep)
+                norm_hidden_states = rearrange(norm_hidden_states, 'b (t l) c -> (b t) l c', t=6)
+                hidden_states = rearrange(hidden_states, 'b (t l) c -> (b t) l c', t=6)
+            elif self.norm_type in ["ada_norm_zero", "layer_norm", "layer_norm_i2vgen"]:
+                # sync_norm:
+                hidden_states = rearrange(hidden_states, '(b t) l c -> b (t l) c', t=6)
+                norm_hidden_states = self.norm2(hidden_states)
+                norm_hidden_states = rearrange(norm_hidden_states, 'b (t l) c -> (b t) l c', t=6)
+                hidden_states = rearrange(hidden_states, 'b (t l) c -> (b t) l c', t=6)
+            elif self.norm_type == "ada_norm_single":
+                # For PixArt norm2 isn't applied here:
+                # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103
+                norm_hidden_states = hidden_states
+            elif self.norm_type == "ada_norm_continuous":
+                # sync_norm:
+                hidden_states = rearrange(hidden_states, '(b t) l c -> b (t l) c', t=6)
+                norm_hidden_states = self.norm2(hidden_states, added_cond_kwargs["pooled_text_emb"])
+                norm_hidden_states = rearrange(norm_hidden_states, 'b (t l) c -> (b t) l c', t=6)
+                hidden_states = rearrange(hidden_states, 'b (t l) c -> (b t) l c', t=6)
+            else:
+                raise ValueError("Incorrect norm")
+
+            if self.pos_embed is not None and self.norm_type != "ada_norm_single":
+                norm_hidden_states = self.pos_embed(norm_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
+        # i2vgen doesn't have this norm 🤷‍♂️
+        if self.norm_type == "ada_norm_continuous":
+            # sync_norm:
+            hidden_states = rearrange(hidden_states, '(b t) l c -> b (t l) c', t=num_cube_faces)
+            norm_hidden_states = self.norm3(hidden_states, added_cond_kwargs["pooled_text_emb"])
+            norm_hidden_states = rearrange(norm_hidden_states, 'b (t l) c -> (b t) l c', t=num_cube_faces)
+            hidden_states = rearrange(hidden_states, 'b (t l) c -> (b t) l c', t=num_cube_faces)
+        elif not self.norm_type == "ada_norm_single":
+            # sync_norm:
+            hidden_states = rearrange(hidden_states, '(b t) l c -> b (t l) c', t=num_cube_faces)
+            norm_hidden_states = self.norm3(hidden_states)
+            norm_hidden_states = rearrange(norm_hidden_states, 'b (t l) c -> (b t) l c', t=num_cube_faces)
+            hidden_states = rearrange(hidden_states, 'b (t l) c -> (b t) l c', t=num_cube_faces)
+        if self.norm_type == "ada_norm_zero":
+            norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+
+        if self.norm_type == "ada_norm_single":
+            # sync_norm:
+            hidden_states = rearrange(hidden_states, '(b t) l c -> b (t l) c', t=6)
+            norm_hidden_states = self.norm2(hidden_states)
+            hidden_states = rearrange(hidden_states, '(b t) l c -> (b t) l c', t=6)
+            norm_hidden_states = rearrange(norm_hidden_states, 'b (t l) c -> (b t) l c', t=6)
+            norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp
+        if self._chunk_size is not None:
+            # "feed_forward_chunk_size" can be used to save memory
+            ff_output = _chunked_feed_forward(self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size)
+        else:
+            ff_output = self.ff(norm_hidden_states)
+
+        if self.norm_type == "ada_norm_zero":
+            ff_output = gate_mlp.unsqueeze(1) * ff_output
+        elif self.norm_type == "ada_norm_single":
+            ff_output = gate_mlp * ff_output
+
+        hidden_states = ff_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        return hidden_states
+
+class LuminaFeedForward(nn.Module):
+    r"""
+    A feed-forward layer.
+
+    Parameters:
+        hidden_size (`int`):
+            The dimensionality of the hidden layers in the model. This parameter determines the width of the model's
+            hidden representations.
+        intermediate_size (`int`): The intermediate dimension of the feedforward layer.
+        multiple_of (`int`, *optional*): Value to ensure hidden dimension is a multiple
+            of this value.
+        ffn_dim_multiplier (float, *optional*): Custom multiplier for hidden
+            dimension. Defaults to None.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        inner_dim: int,
+        multiple_of: Optional[int] = 256,
+        ffn_dim_multiplier: Optional[float] = None,
+    ):
+        super().__init__()
+        inner_dim = int(2 * inner_dim / 3)
+        # custom hidden_size factor multiplier
+        if ffn_dim_multiplier is not None:
+            inner_dim = int(ffn_dim_multiplier * inner_dim)
+        inner_dim = multiple_of * ((inner_dim + multiple_of - 1) // multiple_of)
+
+        self.linear_1 = nn.Linear(
+            dim,
+            inner_dim,
+            bias=False,
+        )
+        self.linear_2 = nn.Linear(
+            inner_dim,
+            dim,
+            bias=False,
+        )
+        self.linear_3 = nn.Linear(
+            dim,
+            inner_dim,
+            bias=False,
+        )
+        self.silu = FP32SiLU()
+
+    def forward(self, x):
+        return self.linear_2(self.silu(self.linear_1(x)) * self.linear_3(x))
+
+
+@maybe_allow_in_graph
+class TemporalBasicTransformerBlock(nn.Module):
+    r"""
+    A basic Transformer block for video like data.
+
+    Parameters:
+        dim (`int`): The number of channels in the input and output.
+        time_mix_inner_dim (`int`): The number of channels for temporal attention.
+        num_attention_heads (`int`): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`): The number of channels in each head.
+        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        time_mix_inner_dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        cross_attention_dim: Optional[int] = None,
+    ):
+        super().__init__()
+        self.is_res = dim == time_mix_inner_dim
+
+        self.norm_in = nn.LayerNorm(dim)
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Self-Attn
+        self.ff_in = FeedForward(
+            dim,
+            dim_out=time_mix_inner_dim,
+            activation_fn="geglu",
+        )
+
+        self.norm1 = nn.LayerNorm(time_mix_inner_dim)
+        self.attn1 = Attention(
+            query_dim=time_mix_inner_dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            cross_attention_dim=None,
+        )
+
+        # 2. Cross-Attn
+        if cross_attention_dim is not None:
+            # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
+            # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
+            # the second cross attention block.
+            self.norm2 = nn.LayerNorm(time_mix_inner_dim)
+            self.attn2 = Attention(
+                query_dim=time_mix_inner_dim,
+                cross_attention_dim=cross_attention_dim,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+            )  # is self-attn if encoder_hidden_states is none
+        else:
+            self.norm2 = None
+            self.attn2 = None
+
+        # 3. Feed-forward
+        self.norm3 = nn.LayerNorm(time_mix_inner_dim)
+        self.ff = FeedForward(time_mix_inner_dim, activation_fn="geglu")
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = None
+
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], **kwargs):
+        # Sets chunk feed-forward
+        self._chunk_size = chunk_size
+        # chunk dim should be hardcoded to 1 to have better speed vs. memory trade-off
+        self._chunk_dim = 1
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        num_frames: int,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        # Notice that normalization is always applied before the real computation in the following blocks.
+        # 0. Self-Attention
+        batch_size = hidden_states.shape[0]
+
+        batch_frames, seq_length, channels = hidden_states.shape
+        batch_size = batch_frames // num_frames
+
+        hidden_states = hidden_states[None, :].reshape(batch_size, num_frames, seq_length, channels)
+        hidden_states = hidden_states.permute(0, 2, 1, 3)
+        hidden_states = hidden_states.reshape(batch_size * seq_length, num_frames, channels)
+
+        residual = hidden_states
+        hidden_states = self.norm_in(hidden_states)
+
+        if self._chunk_size is not None:
+            hidden_states = _chunked_feed_forward(self.ff_in, hidden_states, self._chunk_dim, self._chunk_size)
+        else:
+            hidden_states = self.ff_in(hidden_states)
+
+        if self.is_res:
+            hidden_states = hidden_states + residual
+
+        norm_hidden_states = self.norm1(hidden_states)
+        attn_output = self.attn1(norm_hidden_states, encoder_hidden_states=None)
+        hidden_states = attn_output + hidden_states
+
+        # 3. Cross-Attention
+        if self.attn2 is not None:
+            norm_hidden_states = self.norm2(hidden_states)
+            attn_output = self.attn2(norm_hidden_states, encoder_hidden_states=encoder_hidden_states)
+            hidden_states = attn_output + hidden_states
+
+        # 4. Feed-forward
+        norm_hidden_states = self.norm3(hidden_states)
+
+        if self._chunk_size is not None:
+            ff_output = _chunked_feed_forward(self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size)
+        else:
+            ff_output = self.ff(norm_hidden_states)
+
+        if self.is_res:
+            hidden_states = ff_output + hidden_states
+        else:
+            hidden_states = ff_output
+
+        hidden_states = hidden_states[None, :].reshape(batch_size, seq_length, num_frames, channels)
+        hidden_states = hidden_states.permute(0, 2, 1, 3)
+        hidden_states = hidden_states.reshape(batch_size * num_frames, seq_length, channels)
+
+        return hidden_states
+
+
+class SkipFFTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        kv_input_dim: int,
+        kv_input_dim_proj_use_bias: bool,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        attention_bias: bool = False,
+        attention_out_bias: bool = True,
+    ):
+        super().__init__()
+        if kv_input_dim != dim:
+            self.kv_mapper = nn.Linear(kv_input_dim, dim, kv_input_dim_proj_use_bias)
+        else:
+            self.kv_mapper = None
+
+        self.norm1 = RMSNorm(dim, 1e-06)
+
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            cross_attention_dim=cross_attention_dim,
+            out_bias=attention_out_bias,
+        )
+
+        self.norm2 = RMSNorm(dim, 1e-06)
+
+        self.attn2 = Attention(
+            query_dim=dim,
+            cross_attention_dim=cross_attention_dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            out_bias=attention_out_bias,
+        )
+
+    def forward(self, hidden_states, encoder_hidden_states, cross_attention_kwargs):
+        cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
+
+        if self.kv_mapper is not None:
+            encoder_hidden_states = self.kv_mapper(F.silu(encoder_hidden_states))
+
+        norm_hidden_states = self.norm1(hidden_states)
+
+        attn_output = self.attn1(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states,
+            **cross_attention_kwargs,
+        )
+
+        hidden_states = attn_output + hidden_states
+
+        norm_hidden_states = self.norm2(hidden_states)
+
+        attn_output = self.attn2(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states,
+            **cross_attention_kwargs,
+        )
+
+        hidden_states = attn_output + hidden_states
+
+        return hidden_states
+
+
+@maybe_allow_in_graph
+class FreeNoiseTransformerBlock(nn.Module):
+    r"""
+    A FreeNoise Transformer block.
+
+    Parameters:
+        dim (`int`):
+            The number of channels in the input and output.
+        num_attention_heads (`int`):
+            The number of heads to use for multi-head attention.
+        attention_head_dim (`int`):
+            The number of channels in each head.
+        dropout (`float`, *optional*, defaults to 0.0):
+            The dropout probability to use.
+        cross_attention_dim (`int`, *optional*):
+            The size of the encoder_hidden_states vector for cross attention.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`):
+            Activation function to be used in feed-forward.
+        num_embeds_ada_norm (`int`, *optional*):
+            The number of diffusion steps used during training. See `Transformer2DModel`.
+        attention_bias (`bool`, defaults to `False`):
+            Configure if the attentions should contain a bias parameter.
+        only_cross_attention (`bool`, defaults to `False`):
+            Whether to use only cross-attention layers. In this case two cross attention layers are used.
+        double_self_attention (`bool`, defaults to `False`):
+            Whether to use two self-attention layers. In this case no cross attention layers are used.
+        upcast_attention (`bool`, defaults to `False`):
+            Whether to upcast the attention computation to float32. This is useful for mixed precision training.
+        norm_elementwise_affine (`bool`, defaults to `True`):
+            Whether to use learnable elementwise affine parameters for normalization.
+        norm_type (`str`, defaults to `"layer_norm"`):
+            The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`.
+        final_dropout (`bool` defaults to `False`):
+            Whether to apply a final dropout after the last feed-forward layer.
+        attention_type (`str`, defaults to `"default"`):
+            The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`.
+        positional_embeddings (`str`, *optional*):
+            The type of positional embeddings to apply to.
+        num_positional_embeddings (`int`, *optional*, defaults to `None`):
+            The maximum number of positional embeddings to apply.
+        ff_inner_dim (`int`, *optional*):
+            Hidden dimension of feed-forward MLP.
+        ff_bias (`bool`, defaults to `True`):
+            Whether or not to use bias in feed-forward MLP.
+        attention_out_bias (`bool`, defaults to `True`):
+            Whether or not to use bias in attention output project layer.
+        context_length (`int`, defaults to `16`):
+            The maximum number of frames that the FreeNoise block processes at once.
+        context_stride (`int`, defaults to `4`):
+            The number of frames to be skipped before starting to process a new batch of `context_length` frames.
+        weighting_scheme (`str`, defaults to `"pyramid"`):
+            The weighting scheme to use for weighting averaging of processed latent frames. As described in the
+            Equation 9. of the [FreeNoise](https://arxiv.org/abs/2310.15169) paper, "pyramid" is the default setting
+            used.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout: float = 0.0,
+        cross_attention_dim: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        attention_bias: bool = False,
+        only_cross_attention: bool = False,
+        double_self_attention: bool = False,
+        upcast_attention: bool = False,
+        norm_elementwise_affine: bool = True,
+        norm_type: str = "layer_norm",
+        norm_eps: float = 1e-5,
+        final_dropout: bool = False,
+        positional_embeddings: Optional[str] = None,
+        num_positional_embeddings: Optional[int] = None,
+        ff_inner_dim: Optional[int] = None,
+        ff_bias: bool = True,
+        attention_out_bias: bool = True,
+        context_length: int = 16,
+        context_stride: int = 4,
+        weighting_scheme: str = "pyramid",
+    ):
+        super().__init__()
+        self.dim = dim
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        self.dropout = dropout
+        self.cross_attention_dim = cross_attention_dim
+        self.activation_fn = activation_fn
+        self.attention_bias = attention_bias
+        self.double_self_attention = double_self_attention
+        self.norm_elementwise_affine = norm_elementwise_affine
+        self.positional_embeddings = positional_embeddings
+        self.num_positional_embeddings = num_positional_embeddings
+        self.only_cross_attention = only_cross_attention
+
+        self.set_free_noise_properties(context_length, context_stride, weighting_scheme)
+
+        # We keep these boolean flags for backward-compatibility.
+        self.use_ada_layer_norm_zero = (num_embeds_ada_norm is not None) and norm_type == "ada_norm_zero"
+        self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm"
+        self.use_ada_layer_norm_single = norm_type == "ada_norm_single"
+        self.use_layer_norm = norm_type == "layer_norm"
+        self.use_ada_layer_norm_continuous = norm_type == "ada_norm_continuous"
+
+        if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
+            raise ValueError(
+                f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
+                f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
+            )
+
+        self.norm_type = norm_type
+        self.num_embeds_ada_norm = num_embeds_ada_norm
+
+        if positional_embeddings and (num_positional_embeddings is None):
+            raise ValueError(
+                "If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined."
+            )
+
+        if positional_embeddings == "sinusoidal":
+            self.pos_embed = SinusoidalPositionalEmbedding(dim, max_seq_length=num_positional_embeddings)
+        else:
+            self.pos_embed = None
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Self-Attn
+        self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
+
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            cross_attention_dim=cross_attention_dim if only_cross_attention else None,
+            upcast_attention=upcast_attention,
+            out_bias=attention_out_bias,
+        )
+
+        # 2. Cross-Attn
+        if cross_attention_dim is not None or double_self_attention:
+            self.norm2 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)
+
+            self.attn2 = Attention(
+                query_dim=dim,
+                cross_attention_dim=cross_attention_dim if not double_self_attention else None,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+                out_bias=attention_out_bias,
+            )  # is self-attn if encoder_hidden_states is none
+
+        # 3. Feed-forward
+        self.ff = FeedForward(
+            dim,
+            dropout=dropout,
+            activation_fn=activation_fn,
+            final_dropout=final_dropout,
+            inner_dim=ff_inner_dim,
+            bias=ff_bias,
+        )
+
+        self.norm3 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = 0
+
+    def _get_frame_indices(self, num_frames: int) -> List[Tuple[int, int]]:
+        frame_indices = []
+        for i in range(0, num_frames - self.context_length + 1, self.context_stride):
+            window_start = i
+            window_end = min(num_frames, i + self.context_length)
+            frame_indices.append((window_start, window_end))
+        return frame_indices
+
+    def _get_frame_weights(self, num_frames: int, weighting_scheme: str = "pyramid") -> List[float]:
+        if weighting_scheme == "pyramid":
+            if num_frames % 2 == 0:
+                # num_frames = 4 => [1, 2, 2, 1]
+                weights = list(range(1, num_frames // 2 + 1))
+                weights = weights + weights[::-1]
+            else:
+                # num_frames = 5 => [1, 2, 3, 2, 1]
+                weights = list(range(1, num_frames // 2 + 1))
+                weights = weights + [num_frames // 2 + 1] + weights[::-1]
+        else:
+            raise ValueError(f"Unsupported value for weighting_scheme={weighting_scheme}")
+
+        return weights
+
+    def set_free_noise_properties(
+        self, context_length: int, context_stride: int, weighting_scheme: str = "pyramid"
+    ) -> None:
+        self.context_length = context_length
+        self.context_stride = context_stride
+        self.weighting_scheme = weighting_scheme
+
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0) -> None:
+        # Sets chunk feed-forward
+        self._chunk_size = chunk_size
+        self._chunk_dim = dim
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        *args,
+        **kwargs,
+    ) -> torch.Tensor:
+        if cross_attention_kwargs is not None:
+            if cross_attention_kwargs.get("scale", None) is not None:
+                logger.warning("Passing `scale` to `cross_attention_kwargs` is deprecated. `scale` will be ignored.")
+
+        cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
+
+        # hidden_states: [B x H x W, F, C]
+        device = hidden_states.device
+        dtype = hidden_states.dtype
+
+        num_frames = hidden_states.size(1)
+        frame_indices = self._get_frame_indices(num_frames)
+        frame_weights = self._get_frame_weights(self.context_length, self.weighting_scheme)
+        frame_weights = torch.tensor(frame_weights, device=device, dtype=dtype).unsqueeze(0).unsqueeze(-1)
+        is_last_frame_batch_complete = frame_indices[-1][1] == num_frames
+
+        # Handle out-of-bounds case if num_frames isn't perfectly divisible by context_length
+        # For example, num_frames=25, context_length=16, context_stride=4, then we expect the ranges:
+        #    [(0, 16), (4, 20), (8, 24), (10, 26)]
+        if not is_last_frame_batch_complete:
+            if num_frames < self.context_length:
+                raise ValueError(f"Expected {num_frames=} to be greater or equal than {self.context_length=}")
+            last_frame_batch_length = num_frames - frame_indices[-1][1]
+            frame_indices.append((num_frames - self.context_length, num_frames))
+
+        num_times_accumulated = torch.zeros((1, num_frames, 1), device=device)
+        accumulated_values = torch.zeros_like(hidden_states)
+
+        for i, (frame_start, frame_end) in enumerate(frame_indices):
+            # The reason for slicing here is to ensure that if (frame_end - frame_start) is to handle
+            # cases like frame_indices=[(0, 16), (16, 20)], if the user provided a video with 19 frames, or
+            # essentially a non-multiple of `context_length`.
+            weights = torch.ones_like(num_times_accumulated[:, frame_start:frame_end])
+            weights *= frame_weights
+
+            hidden_states_chunk = hidden_states[:, frame_start:frame_end]
+
+            # Notice that normalization is always applied before the real computation in the following blocks.
+            # 1. Self-Attention
+            norm_hidden_states = self.norm1(hidden_states_chunk)
+
+            if self.pos_embed is not None:
+                norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+            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,
+            )
+
+            hidden_states_chunk = attn_output + hidden_states_chunk
+            if hidden_states_chunk.ndim == 4:
+                hidden_states_chunk = hidden_states_chunk.squeeze(1)
+
+            # 2. Cross-Attention
+            if self.attn2 is not None:
+                norm_hidden_states = self.norm2(hidden_states_chunk)
+
+                if self.pos_embed is not None and self.norm_type != "ada_norm_single":
+                    norm_hidden_states = self.pos_embed(norm_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_chunk = attn_output + hidden_states_chunk
+
+            if i == len(frame_indices) - 1 and not is_last_frame_batch_complete:
+                accumulated_values[:, -last_frame_batch_length:] += (
+                    hidden_states_chunk[:, -last_frame_batch_length:] * weights[:, -last_frame_batch_length:]
+                )
+                num_times_accumulated[:, -last_frame_batch_length:] += weights[:, -last_frame_batch_length]
+            else:
+                accumulated_values[:, frame_start:frame_end] += hidden_states_chunk * weights
+                num_times_accumulated[:, frame_start:frame_end] += weights
+
+        hidden_states = torch.where(
+            num_times_accumulated > 0, accumulated_values / num_times_accumulated, accumulated_values
+        ).to(dtype)
+
+        # 3. Feed-forward
+        norm_hidden_states = self.norm3(hidden_states)
+
+        if self._chunk_size is not None:
+            ff_output = _chunked_feed_forward(self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size)
+        else:
+            ff_output = self.ff(norm_hidden_states)
+
+        hidden_states = ff_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        return hidden_states
+
+
+class FeedForward(nn.Module):
+    r"""
+    A feed-forward layer.
+
+    Parameters:
+        dim (`int`): The number of channels in the input.
+        dim_out (`int`, *optional*): The number of channels in the output. If not given, defaults to `dim`.
+        mult (`int`, *optional*, defaults to 4): The multiplier to use for the hidden dimension.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        final_dropout (`bool` *optional*, defaults to False): Apply a final dropout.
+        bias (`bool`, defaults to True): Whether to use a bias in the linear layer.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        dim_out: Optional[int] = None,
+        mult: int = 4,
+        dropout: float = 0.0,
+        activation_fn: str = "geglu",
+        final_dropout: bool = False,
+        inner_dim=None,
+        bias: bool = True,
+    ):
+        super().__init__()
+        if inner_dim is None:
+            inner_dim = int(dim * mult)
+        dim_out = dim_out if dim_out is not None else dim
+
+        if activation_fn == "gelu":
+            act_fn = GELU(dim, inner_dim, bias=bias)
+        if activation_fn == "gelu-approximate":
+            act_fn = GELU(dim, inner_dim, approximate="tanh", bias=bias)
+        elif activation_fn == "geglu":
+            act_fn = GEGLU(dim, inner_dim, bias=bias)
+        elif activation_fn == "geglu-approximate":
+            act_fn = ApproximateGELU(dim, inner_dim, bias=bias)
+        elif activation_fn == "swiglu":
+            act_fn = SwiGLU(dim, inner_dim, bias=bias)
+
+        self.net = nn.ModuleList([])
+        # project in
+        self.net.append(act_fn)
+        # project dropout
+        self.net.append(nn.Dropout(dropout))
+        # project out
+        self.net.append(nn.Linear(inner_dim, dim_out, bias=bias))
+        # FF as used in Vision Transformer, MLP-Mixer, etc. have a final dropout
+        if final_dropout:
+            self.net.append(nn.Dropout(dropout))
+
+    def forward(self, hidden_states: torch.Tensor, *args, **kwargs) -> torch.Tensor:
+        if len(args) > 0 or kwargs.get("scale", None) is not None:
+            deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`."
+            deprecate("scale", "1.0.0", deprecation_message)
+        for module in self.net:
+            hidden_states = module(hidden_states)
+        return hidden_states

Marigold/unet/transformer_2d.py

@@ -0,0 +1,585 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Any, Dict, Optional
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from diffusers.configuration_utils import LegacyConfigMixin, register_to_config
+from diffusers.utils import deprecate, is_torch_version, logging
+from diffusers.models.embeddings import ImagePositionalEmbeddings, PatchEmbed, PixArtAlphaTextProjection
+from diffusers.models.modeling_outputs import Transformer2DModelOutput
+from diffusers.models.modeling_utils import LegacyModelMixin
+from diffusers.models.normalization import AdaLayerNormSingle
+from einops import rearrange
+from Marigold.unet.attention import BasicTransformerBlock
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class Transformer2DModelOutput(Transformer2DModelOutput):
+    def __init__(self, *args, **kwargs):
+        deprecation_message = "Importing `Transformer2DModelOutput` from `diffusers.models.transformer_2d` is deprecated and this will be removed in a future version. Please use `from diffusers.models.modeling_outputs import Transformer2DModelOutput`, instead."
+        deprecate("Transformer2DModelOutput", "1.0.0", deprecation_message)
+        super().__init__(*args, **kwargs)
+
+
+class Transformer2DModel(LegacyModelMixin, LegacyConfigMixin):
+    """
+    A 2D Transformer model for image-like data.
+
+    Parameters:
+        num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
+        in_channels (`int`, *optional*):
+            The number of channels in the input and output (specify if the input is **continuous**).
+        num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
+        sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**).
+            This is fixed during training since it is used to learn a number of position embeddings.
+        num_vector_embeds (`int`, *optional*):
+            The number of classes of the vector embeddings of the latent pixels (specify if the input is **discrete**).
+            Includes the class for the masked latent pixel.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to use in feed-forward.
+        num_embeds_ada_norm ( `int`, *optional*):
+            The number of diffusion steps used during training. Pass if at least one of the norm_layers is
+            `AdaLayerNorm`. This is fixed during training since it is used to learn a number of embeddings that are
+            added to the hidden states.
+
+            During inference, you can denoise for up to but not more steps than `num_embeds_ada_norm`.
+        attention_bias (`bool`, *optional*):
+            Configure if the `TransformerBlocks` attention should contain a bias parameter.
+    """
+
+    _supports_gradient_checkpointing = True
+    _no_split_modules = ["BasicTransformerBlock"]
+
+    @register_to_config
+    def __init__(
+        self,
+        num_attention_heads: int = 16,
+        attention_head_dim: int = 88,
+        in_channels: Optional[int] = None,
+        out_channels: Optional[int] = None,
+        num_layers: int = 1,
+        dropout: float = 0.0,
+        norm_num_groups: int = 32,
+        cross_attention_dim: Optional[int] = None,
+        attention_bias: bool = False,
+        sample_size: Optional[int] = None,
+        num_vector_embeds: Optional[int] = None,
+        patch_size: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        use_linear_projection: bool = False,
+        only_cross_attention: bool = False,
+        double_self_attention: bool = False,
+        upcast_attention: bool = False,
+        norm_type: str = "layer_norm",  # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single', 'ada_norm_continuous', 'layer_norm_i2vgen'
+        norm_elementwise_affine: bool = True,
+        norm_eps: float = 1e-5,
+        attention_type: str = "default",
+        caption_channels: int = None,
+        interpolation_scale: float = None,
+        use_additional_conditions: Optional[bool] = None,
+        use_RoPE: bool = False,
+    ):
+        super().__init__()
+
+        # Validate inputs.
+        if patch_size is not None:
+            if norm_type not in ["ada_norm", "ada_norm_zero", "ada_norm_single"]:
+                raise NotImplementedError(
+                    f"Forward pass is not implemented when `patch_size` is not None and `norm_type` is '{norm_type}'."
+                )
+            elif norm_type in ["ada_norm", "ada_norm_zero"] and num_embeds_ada_norm is None:
+                raise ValueError(
+                    f"When using a `patch_size` and this `norm_type` ({norm_type}), `num_embeds_ada_norm` cannot be None."
+                )
+
+        # 1. Transformer2DModel can process both standard continuous images of shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of shape `(batch_size, num_image_vectors)`
+        # Define whether input is continuous or discrete depending on configuration
+        self.is_input_continuous = (in_channels is not None) and (patch_size is None)
+        self.is_input_vectorized = num_vector_embeds is not None
+        self.is_input_patches = in_channels is not None and patch_size is not None
+
+        if self.is_input_continuous and self.is_input_vectorized:
+            raise ValueError(
+                f"Cannot define both `in_channels`: {in_channels} and `num_vector_embeds`: {num_vector_embeds}. Make"
+                " sure that either `in_channels` or `num_vector_embeds` is None."
+            )
+        elif self.is_input_vectorized and self.is_input_patches:
+            raise ValueError(
+                f"Cannot define both `num_vector_embeds`: {num_vector_embeds} and `patch_size`: {patch_size}. Make"
+                " sure that either `num_vector_embeds` or `num_patches` is None."
+            )
+        elif not self.is_input_continuous and not self.is_input_vectorized and not self.is_input_patches:
+            raise ValueError(
+                f"Has to define `in_channels`: {in_channels}, `num_vector_embeds`: {num_vector_embeds}, or patch_size:"
+                f" {patch_size}. Make sure that `in_channels`, `num_vector_embeds` or `num_patches` is not None."
+            )
+
+        if norm_type == "layer_norm" and num_embeds_ada_norm is not None:
+            deprecation_message = (
+                f"The configuration file of this model: {self.__class__} is outdated. `norm_type` is either not set or"
+                " incorrectly set to `'layer_norm'`. Make sure to set `norm_type` to `'ada_norm'` in the config."
+                " Please make sure to update the config accordingly as leaving `norm_type` might led to incorrect"
+                " results in future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it"
+                " would be very nice if you could open a Pull request for the `transformer/config.json` file"
+            )
+            deprecate("norm_type!=num_embeds_ada_norm", "1.0.0", deprecation_message, standard_warn=False)
+            norm_type = "ada_norm"
+
+        # Set some common variables used across the board.
+        self.use_linear_projection = use_linear_projection
+        self.interpolation_scale = interpolation_scale
+        self.caption_channels = caption_channels
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        self.inner_dim = self.config.num_attention_heads * self.config.attention_head_dim
+        self.in_channels = in_channels
+        self.out_channels = in_channels if out_channels is None else out_channels
+        self.gradient_checkpointing = False
+        self.use_RoPE = use_RoPE
+
+
+        if use_additional_conditions is None:
+            if norm_type == "ada_norm_single" and sample_size == 128:
+                use_additional_conditions = True
+            else:
+                use_additional_conditions = False
+        self.use_additional_conditions = use_additional_conditions
+
+        # 2. Initialize the right blocks.
+        # These functions follow a common structure:
+        # a. Initialize the input blocks. b. Initialize the transformer blocks.
+        # c. Initialize the output blocks and other projection blocks when necessary.
+        if self.is_input_continuous:
+            self._init_continuous_input(norm_type=norm_type)
+        elif self.is_input_vectorized:
+            self._init_vectorized_inputs(norm_type=norm_type)
+        elif self.is_input_patches:
+            self._init_patched_inputs(norm_type=norm_type)
+
+    def _init_continuous_input(self, norm_type):
+        self.norm = torch.nn.GroupNorm(
+            num_groups=self.config.norm_num_groups, num_channels=self.in_channels, eps=1e-6, affine=True
+        )
+        if self.use_linear_projection:
+            self.proj_in = torch.nn.Linear(self.in_channels, self.inner_dim)
+        else:
+            self.proj_in = torch.nn.Conv2d(self.in_channels, self.inner_dim, kernel_size=1, stride=1, padding=0)
+
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    self.inner_dim,
+                    self.config.num_attention_heads,
+                    self.config.attention_head_dim,
+                    dropout=self.config.dropout,
+                    cross_attention_dim=self.config.cross_attention_dim,
+                    activation_fn=self.config.activation_fn,
+                    num_embeds_ada_norm=self.config.num_embeds_ada_norm,
+                    attention_bias=self.config.attention_bias,
+                    only_cross_attention=self.config.only_cross_attention,
+                    double_self_attention=self.config.double_self_attention,
+                    upcast_attention=self.config.upcast_attention,
+                    norm_type=norm_type,
+                    norm_elementwise_affine=self.config.norm_elementwise_affine,
+                    norm_eps=self.config.norm_eps,
+                    attention_type=self.config.attention_type,
+                    use_RoPE=self.use_RoPE,
+                )
+                for _ in range(self.config.num_layers)
+            ]
+        )
+
+        if self.use_linear_projection:
+            self.proj_out = torch.nn.Linear(self.inner_dim, self.out_channels)
+        else:
+            self.proj_out = torch.nn.Conv2d(self.inner_dim, self.out_channels, kernel_size=1, stride=1, padding=0)
+
+    def _init_vectorized_inputs(self, norm_type):
+        assert self.config.sample_size is not None, "Transformer2DModel over discrete input must provide sample_size"
+        assert (
+            self.config.num_vector_embeds is not None
+        ), "Transformer2DModel over discrete input must provide num_embed"
+
+        self.height = self.config.sample_size
+        self.width = self.config.sample_size
+        self.num_latent_pixels = self.height * self.width
+
+        self.latent_image_embedding = ImagePositionalEmbeddings(
+            num_embed=self.config.num_vector_embeds, embed_dim=self.inner_dim, height=self.height, width=self.width
+        )
+
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    self.inner_dim,
+                    self.config.num_attention_heads,
+                    self.config.attention_head_dim,
+                    dropout=self.config.dropout,
+                    cross_attention_dim=self.config.cross_attention_dim,
+                    activation_fn=self.config.activation_fn,
+                    num_embeds_ada_norm=self.config.num_embeds_ada_norm,
+                    attention_bias=self.config.attention_bias,
+                    only_cross_attention=self.config.only_cross_attention,
+                    double_self_attention=self.config.double_self_attention,
+                    upcast_attention=self.config.upcast_attention,
+                    norm_type=norm_type,
+                    norm_elementwise_affine=self.config.norm_elementwise_affine,
+                    norm_eps=self.config.norm_eps,
+                    attention_type=self.config.attention_type,
+                )
+                for _ in range(self.config.num_layers)
+            ]
+        )
+
+        self.norm_out = nn.LayerNorm(self.inner_dim)
+        self.out = nn.Linear(self.inner_dim, self.config.num_vector_embeds - 1)
+
+    def _init_patched_inputs(self, norm_type):
+        assert self.config.sample_size is not None, "Transformer2DModel over patched input must provide sample_size"
+
+        self.height = self.config.sample_size
+        self.width = self.config.sample_size
+
+        self.patch_size = self.config.patch_size
+        interpolation_scale = (
+            self.config.interpolation_scale
+            if self.config.interpolation_scale is not None
+            else max(self.config.sample_size // 64, 1)
+        )
+        self.pos_embed = PatchEmbed(
+            height=self.config.sample_size,
+            width=self.config.sample_size,
+            patch_size=self.config.patch_size,
+            in_channels=self.in_channels,
+            embed_dim=self.inner_dim,
+            interpolation_scale=interpolation_scale,
+        )
+
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    self.inner_dim,
+                    self.config.num_attention_heads,
+                    self.config.attention_head_dim,
+                    dropout=self.config.dropout,
+                    cross_attention_dim=self.config.cross_attention_dim,
+                    activation_fn=self.config.activation_fn,
+                    num_embeds_ada_norm=self.config.num_embeds_ada_norm,
+                    attention_bias=self.config.attention_bias,
+                    only_cross_attention=self.config.only_cross_attention,
+                    double_self_attention=self.config.double_self_attention,
+                    upcast_attention=self.config.upcast_attention,
+                    norm_type=norm_type,
+                    norm_elementwise_affine=self.config.norm_elementwise_affine,
+                    norm_eps=self.config.norm_eps,
+                    attention_type=self.config.attention_type,
+                )
+                for _ in range(self.config.num_layers)
+            ]
+        )
+
+        if self.config.norm_type != "ada_norm_single":
+            self.norm_out = nn.LayerNorm(self.inner_dim, elementwise_affine=False, eps=1e-6)
+            self.proj_out_1 = nn.Linear(self.inner_dim, 2 * self.inner_dim)
+            self.proj_out_2 = nn.Linear(
+                self.inner_dim, self.config.patch_size * self.config.patch_size * self.out_channels
+            )
+        elif self.config.norm_type == "ada_norm_single":
+            self.norm_out = nn.LayerNorm(self.inner_dim, elementwise_affine=False, eps=1e-6)
+            self.scale_shift_table = nn.Parameter(torch.randn(2, self.inner_dim) / self.inner_dim**0.5)
+            self.proj_out = nn.Linear(
+                self.inner_dim, self.config.patch_size * self.config.patch_size * self.out_channels
+            )
+
+        # PixArt-Alpha blocks.
+        self.adaln_single = None
+        if self.config.norm_type == "ada_norm_single":
+            # TODO(Sayak, PVP) clean this, for now we use sample size to determine whether to use
+            # additional conditions until we find better name
+            self.adaln_single = AdaLayerNormSingle(
+                self.inner_dim, use_additional_conditions=self.use_additional_conditions
+            )
+
+        self.caption_projection = None
+        if self.caption_channels is not None:
+            self.caption_projection = PixArtAlphaTextProjection(
+                in_features=self.caption_channels, hidden_size=self.inner_dim
+            )
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        added_cond_kwargs: Dict[str, torch.Tensor] = None,
+        class_labels: Optional[torch.LongTensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ):
+        """
+        The [`Transformer2DModel`] forward method.
+
+        Args:
+            hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.Tensor` of shape `(batch size, channel, height, width)` if continuous):
+                Input `hidden_states`.
+            encoder_hidden_states ( `torch.Tensor` of shape `(batch size, sequence len, embed dims)`, *optional*):
+                Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
+                self-attention.
+            timestep ( `torch.LongTensor`, *optional*):
+                Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`.
+            class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*):
+                Used to indicate class labels conditioning. Optional class labels to be applied as an embedding in
+                `AdaLayerZeroNorm`.
+            cross_attention_kwargs ( `Dict[str, Any]`, *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).
+            attention_mask ( `torch.Tensor`, *optional*):
+                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+                negative values to the attention scores corresponding to "discard" tokens.
+            encoder_attention_mask ( `torch.Tensor`, *optional*):
+                Cross-attention mask applied to `encoder_hidden_states`. Two formats supported:
+
+                    * Mask `(batch, sequence_length)` True = keep, False = discard.
+                    * Bias `(batch, 1, sequence_length)` 0 = keep, -10000 = discard.
+
+                If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format
+                above. This bias will be added to the cross-attention scores.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+                tuple.
+
+        Returns:
+            If `return_dict` is True, an [`~models.transformers.transformer_2d.Transformer2DModelOutput`] is returned,
+            otherwise a `tuple` where the first element is the sample tensor.
+        """
+        if cross_attention_kwargs is not None:
+            if cross_attention_kwargs.get("scale", None) is not None:
+                logger.warning("Passing `scale` to `cross_attention_kwargs` is deprecated. `scale` will be ignored.")
+        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension.
+        #   we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward.
+        #   we can tell by counting dims; if ndim == 2: it's a mask rather than a bias.
+        # 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 and attention_mask.ndim == 2:
+            # 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(hidden_states.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 and encoder_attention_mask.ndim == 2:
+            encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # 1. Input
+        if self.is_input_continuous:
+            batch_size, _, height, width = hidden_states.shape
+            residual = hidden_states
+            hidden_states, inner_dim = self._operate_on_continuous_inputs(hidden_states)
+        elif self.is_input_vectorized:
+            hidden_states = self.latent_image_embedding(hidden_states)
+        elif self.is_input_patches:
+            height, width = hidden_states.shape[-2] // self.patch_size, hidden_states.shape[-1] // self.patch_size
+            hidden_states, encoder_hidden_states, timestep, embedded_timestep = self._operate_on_patched_inputs(
+                hidden_states, encoder_hidden_states, timestep, added_cond_kwargs
+            )
+        # print("Hidden States Shape: ", hidden_states.shape)
+        # hidden_states = rearrange(hidden_states, '(b t) l c ->  b (t l) c', t=6)
+        # print("Hidden States Shape: ", hidden_states.shape)
+        # 2. Blocks
+        for block in self.transformer_blocks:
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    timestep,
+                    cross_attention_kwargs,
+                    class_labels,
+                    **ckpt_kwargs,
+                )
+            else:
+                hidden_states = block(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_attention_mask=encoder_attention_mask,
+                    timestep=timestep,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    class_labels=class_labels,
+                )
+
+        # 3. Output
+        if self.is_input_continuous:
+            output = self._get_output_for_continuous_inputs(
+                hidden_states=hidden_states,
+                residual=residual,
+                batch_size=batch_size,
+                height=height,
+                width=width,
+                inner_dim=inner_dim,
+            )
+        elif self.is_input_vectorized:
+            output = self._get_output_for_vectorized_inputs(hidden_states)
+        elif self.is_input_patches:
+            output = self._get_output_for_patched_inputs(
+                hidden_states=hidden_states,
+                timestep=timestep,
+                class_labels=class_labels,
+                embedded_timestep=embedded_timestep,
+                height=height,
+                width=width,
+            )
+
+        if not return_dict:
+            return (output,)
+
+        return Transformer2DModelOutput(sample=output)
+
+    def _operate_on_continuous_inputs(self, hidden_states):
+        batch, _, height, width = hidden_states.shape
+        # sync_norm:
+        hidden_states = rearrange(hidden_states, "(b t) c h w -> b c (h w t)", b=1, h=height, w=width)
+        hidden_states = self.norm(hidden_states)
+        hidden_states = rearrange(hidden_states, "b c (h w t) -> (b t) c h w", b=1, h=height, w=width)
+        if not self.use_linear_projection:
+            hidden_states = self.proj_in(hidden_states)
+            inner_dim = hidden_states.shape[1]
+            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim)
+        else:
+            inner_dim = hidden_states.shape[1]
+            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim)
+            hidden_states = self.proj_in(hidden_states)
+        return hidden_states, inner_dim
+
+    def _operate_on_patched_inputs(self, hidden_states, encoder_hidden_states, timestep, added_cond_kwargs):
+        batch_size = hidden_states.shape[0]
+        hidden_states = self.pos_embed(hidden_states)
+        embedded_timestep = None
+
+        if self.adaln_single is not None:
+            if self.use_additional_conditions and added_cond_kwargs is None:
+                raise ValueError(
+                    "`added_cond_kwargs` cannot be None when using additional conditions for `adaln_single`."
+                )
+            timestep, embedded_timestep = self.adaln_single(
+                timestep, added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_states.dtype
+            )
+
+        if self.caption_projection is not None:
+            encoder_hidden_states = self.caption_projection(encoder_hidden_states)
+            encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1])
+
+        return hidden_states, encoder_hidden_states, timestep, embedded_timestep
+
+    def _get_output_for_continuous_inputs(self, hidden_states, residual, batch_size, height, width, inner_dim):
+        if not self.use_linear_projection:
+            hidden_states = (
+                hidden_states.reshape(batch_size, height, width, inner_dim).permute(0, 3, 1, 2).contiguous()
+            )
+            hidden_states = self.proj_out(hidden_states)
+        else:
+            hidden_states = self.proj_out(hidden_states)
+            hidden_states = (
+                hidden_states.reshape(batch_size, height, width, inner_dim).permute(0, 3, 1, 2).contiguous()
+            )
+
+        output = hidden_states + residual
+        return output
+
+    def _get_output_for_vectorized_inputs(self, hidden_states):
+        b, c, h, w = hidden_states.shape
+        # sync_norm:
+        hidden_states = rearrange(hidden_states, "(b t) c h w -> b c (h w t)", b=1, h=h, w=w)
+        hidden_states = self.norm_out(hidden_states)
+        hidden_states = rearrange(hidden_states, "b c (h w t) -> (b t) c h w", b=1, h=h, w=w)
+        logits = self.out(hidden_states)
+        # (batch, self.num_vector_embeds - 1, self.num_latent_pixels)
+        logits = logits.permute(0, 2, 1)
+        # log(p(x_0))
+        output = F.log_softmax(logits.double(), dim=1).float()
+        return output
+
+    def _get_output_for_patched_inputs(
+        self, hidden_states, timestep, class_labels, embedded_timestep, height=None, width=None
+    ):
+        if self.config.norm_type != "ada_norm_single":
+            conditioning = self.transformer_blocks[0].norm1.emb(
+                timestep, class_labels, hidden_dtype=hidden_states.dtype
+            )
+            shift, scale = self.proj_out_1(F.silu(conditioning)).chunk(2, dim=1)
+            # sync_norm:
+            b, c, h, w = hidden_states.shape
+            hidden_states = rearrange(hidden_states, "(b t) c h w -> b c (h w t)", b=1, h=h, w=w)
+            hidden_states = self.norm_out(hidden_states) * (1 + scale[:, None]) + shift[:, None]
+            hidden_states = rearrange(hidden_states, "b c (h w t) -> (b t) c h w", b=1, h=h, w=w)
+            hidden_states = self.proj_out_2(hidden_states)
+        elif self.config.norm_type == "ada_norm_single":
+            shift, scale = (self.scale_shift_table[None] + embedded_timestep[:, None]).chunk(2, dim=1)
+            # sync_norm:
+            b, c, h, w = hidden_states.shape
+            hidden_states = rearrange(hidden_states, "(b t) c h w -> b c (h w t)", b=1, h=h, w=w)
+            hidden_states = self.norm_out(hidden_states)
+            hidden_states = rearrange(hidden_states, "b c (h w t) -> (b t) c h w", b=1, h=h, w=w)
+            # Modulation
+            hidden_states = hidden_states * (1 + scale) + shift
+            hidden_states = self.proj_out(hidden_states)
+            hidden_states = hidden_states.squeeze(1)
+
+        # unpatchify
+        if self.adaln_single is None:
+            height = width = int(hidden_states.shape[1] ** 0.5)
+        hidden_states = hidden_states.reshape(
+            shape=(-1, height, width, self.patch_size, self.patch_size, self.out_channels)
+        )
+        hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states)
+        output = hidden_states.reshape(
+            shape=(-1, self.out_channels, height * self.patch_size, width * self.patch_size)
+        )
+        return output

Marigold/unet/unet_2d_condition.py

@@ -0,0 +1,1414 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.loaders import PeftAdapterMixin, UNet2DConditionLoadersMixin
+from diffusers.loaders.single_file_model import FromOriginalModelMixin
+from diffusers.utils import BaseOutput, logging
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.embeddings import (
+    GaussianFourierProjection,
+    GLIGENTextBoundingboxProjection,
+    ImageHintTimeEmbedding,
+    ImageProjection,
+    ImageTimeEmbedding,
+    TextImageProjection,
+    TextImageTimeEmbedding,
+    TextTimeEmbedding,
+    TimestepEmbedding,
+    Timesteps,
+)
+from diffusers.models.activations import get_activation
+from Marigold.unet.attention_processor import (
+    ADDED_KV_ATTENTION_PROCESSORS,
+    CROSS_ATTENTION_PROCESSORS,
+    Attention,
+    AttentionProcessor,
+    AttnAddedKVProcessor,
+    AttnProcessor,
+    FusedAttnProcessor2_0,
+)
+from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate, logging, scale_lora_layers, unscale_lora_layers
+from Marigold.unet.unet_2d_blocks import get_mid_block, get_down_block, get_up_block
+from einops import rearrange
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class UNet2DConditionOutput(BaseOutput):
+    """
+    The output of [`UNet2DConditionModel`].
+
+    Args:
+        sample (`torch.Tensor` 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.Tensor = None
+
+
+class UNet2DConditionModel(
+    ModelMixin, ConfigMixin, FromOriginalModelMixin, UNet2DConditionLoadersMixin, PeftAdapterMixin
+):
+    r"""
+    A conditional 2D UNet model that takes a noisy sample, conditional state, and 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.
+        flip_sin_to_cos (`bool`, *optional*, defaults to `True`):
+            Whether to flip the sin to cos in the time embedding.
+        freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding.
+        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 one of `UNetMidBlock2DCrossAttn`, `UNetMidBlock2D`, 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.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        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`, `Tuple[int]`, or `Tuple[Tuple]` , *optional*, defaults to 1):
+            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
+            [`~models.unets.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unets.unet_2d_blocks.CrossAttnUpBlock2D`],
+            [`~models.unets.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+        reverse_transformer_layers_per_block : (`Tuple[Tuple]`, *optional*, defaults to None):
+            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`], in the upsampling
+            blocks of the U-Net. Only relevant if `transformer_layers_per_block` is of type `Tuple[Tuple]` and for
+            [`~models.unets.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unets.unet_2d_blocks.CrossAttnUpBlock2D`],
+            [`~models.unets.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`
+        resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
+            for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`.
+        class_embed_type (`str`, *optional*, defaults to `None`):
+            The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
+            `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
+        addition_embed_type (`str`, *optional*, defaults to `None`):
+            Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
+            "text". "text" will use the `TextTimeEmbedding` layer.
+        addition_time_embed_dim: (`int`, *optional*, defaults to `None`):
+            Dimension for the timestep embeddings.
+        num_class_embeds (`int`, *optional*, defaults to `None`):
+            Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
+            class conditioning with `class_embed_type` equal to `None`.
+        time_embedding_type (`str`, *optional*, defaults to `positional`):
+            The type of position embedding to use for timesteps. Choose from `positional` or `fourier`.
+        time_embedding_dim (`int`, *optional*, defaults to `None`):
+            An optional override for the dimension of the projected time embedding.
+        time_embedding_act_fn (`str`, *optional*, defaults to `None`):
+            Optional activation function to use only once on the time embeddings before they are passed to the rest of
+            the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`.
+        timestep_post_act (`str`, *optional*, defaults to `None`):
+            The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`.
+        time_cond_proj_dim (`int`, *optional*, defaults to `None`):
+            The dimension of `cond_proj` layer in the timestep embedding.
+        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.
+        projection_class_embeddings_input_dim (`int`, *optional*): The dimension of the `class_labels` input when
+            `class_embed_type="projection"`. Required when `class_embed_type="projection"`.
+        class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time
+            embeddings with the class embeddings.
+        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
+    _no_split_modules = ["BasicTransformerBlock", "ResnetBlock2D", "CrossAttnUpBlock2D"]
+
+    @register_to_config
+    def __init__(
+        self,
+        sample_size: Optional[int] = None,
+        in_channels: int = 4,
+        out_channels: int = 4,
+        center_input_sample: bool = False,
+        flip_sin_to_cos: bool = True,
+        freq_shift: int = 0,
+        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,
+        dropout: float = 0.0,
+        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], Tuple[Tuple]] = 1,
+        reverse_transformer_layers_per_block: Optional[Tuple[Tuple[int]]] = None,
+        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,
+        class_embed_type: Optional[str] = None,
+        addition_embed_type: Optional[str] = None,
+        addition_time_embed_dim: Optional[int] = None,
+        num_class_embeds: Optional[int] = None,
+        upcast_attention: bool = False,
+        resnet_time_scale_shift: str = "default",
+        resnet_skip_time_act: bool = False,
+        resnet_out_scale_factor: float = 1.0,
+        time_embedding_type: str = "positional",
+        time_embedding_dim: Optional[int] = None,
+        time_embedding_act_fn: Optional[str] = None,
+        timestep_post_act: Optional[str] = None,
+        time_cond_proj_dim: Optional[int] = None,
+        conv_in_kernel: int = 3,
+        conv_out_kernel: int = 3,
+        projection_class_embeddings_input_dim: Optional[int] = None,
+        attention_type: str = "default",
+        class_embeddings_concat: bool = False,
+        mid_block_only_cross_attention: Optional[bool] = None,
+        cross_attention_norm: Optional[str] = None,
+        addition_embed_type_num_heads: int = 64,
+        use_RoPE: bool = False
+    ):
+        super().__init__()
+
+        self.sample_size = sample_size
+        self.use_RoPE = use_RoPE
+        self.gradient_checkpointing = False
+        
+        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
+        self._check_config(
+            down_block_types=down_block_types,
+            up_block_types=up_block_types,
+            only_cross_attention=only_cross_attention,
+            block_out_channels=block_out_channels,
+            layers_per_block=layers_per_block,
+            cross_attention_dim=cross_attention_dim,
+            transformer_layers_per_block=transformer_layers_per_block,
+            reverse_transformer_layers_per_block=reverse_transformer_layers_per_block,
+            attention_head_dim=attention_head_dim,
+            num_attention_heads=num_attention_heads,
+        )
+
+        # 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
+        )
+
+        # time
+        time_embed_dim, timestep_input_dim = self._set_time_proj(
+            time_embedding_type,
+            block_out_channels=block_out_channels,
+            flip_sin_to_cos=flip_sin_to_cos,
+            freq_shift=freq_shift,
+            time_embedding_dim=time_embedding_dim,
+        )
+
+        self.time_embedding = TimestepEmbedding(
+            timestep_input_dim,
+            time_embed_dim,
+            act_fn=act_fn,
+            post_act_fn=timestep_post_act,
+            cond_proj_dim=time_cond_proj_dim,
+        )
+
+        self._set_encoder_hid_proj(
+            encoder_hid_dim_type,
+            cross_attention_dim=cross_attention_dim,
+            encoder_hid_dim=encoder_hid_dim,
+        )
+
+        # class embedding
+        self._set_class_embedding(
+            class_embed_type,
+            act_fn=act_fn,
+            num_class_embeds=num_class_embeds,
+            projection_class_embeddings_input_dim=projection_class_embeddings_input_dim,
+            time_embed_dim=time_embed_dim,
+            timestep_input_dim=timestep_input_dim,
+        )
+
+        self._set_add_embedding(
+            addition_embed_type,
+            addition_embed_type_num_heads=addition_embed_type_num_heads,
+            addition_time_embed_dim=addition_time_embed_dim,
+            cross_attention_dim=cross_attention_dim,
+            encoder_hid_dim=encoder_hid_dim,
+            flip_sin_to_cos=flip_sin_to_cos,
+            freq_shift=freq_shift,
+            projection_class_embeddings_input_dim=projection_class_embeddings_input_dim,
+            time_embed_dim=time_embed_dim,
+        )
+
+        if time_embedding_act_fn is None:
+            self.time_embed_act = None
+        else:
+            self.time_embed_act = get_activation(time_embedding_act_fn)
+
+        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)
+
+        if class_embeddings_concat:
+            # The time embeddings are concatenated with the class embeddings. The dimension of the
+            # time embeddings passed to the down, middle, and up blocks is twice the dimension of the
+            # regular time embeddings
+            blocks_time_embed_dim = time_embed_dim * 2
+        else:
+            blocks_time_embed_dim = time_embed_dim
+
+        # 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,
+                attention_type=attention_type,
+                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,
+                dropout=dropout,
+                use_RoPE=use_RoPE,
+            )
+            self.down_blocks.append(down_block)
+
+        # mid
+        self.mid_block = get_mid_block(
+            mid_block_type,
+            temb_channels=blocks_time_embed_dim,
+            in_channels=block_out_channels[-1],
+            resnet_eps=norm_eps,
+            resnet_act_fn=act_fn,
+            resnet_groups=norm_num_groups,
+            output_scale_factor=mid_block_scale_factor,
+            transformer_layers_per_block=transformer_layers_per_block[-1],
+            num_attention_heads=num_attention_heads[-1],
+            cross_attention_dim=cross_attention_dim[-1],
+            dual_cross_attention=dual_cross_attention,
+            use_linear_projection=use_linear_projection,
+            mid_block_only_cross_attention=mid_block_only_cross_attention,
+            upcast_attention=upcast_attention,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            attention_type=attention_type,
+            resnet_skip_time_act=resnet_skip_time_act,
+            cross_attention_norm=cross_attention_norm,
+            attention_head_dim=attention_head_dim[-1],
+            dropout=dropout,
+            use_RoPE=use_RoPE,
+        )
+
+        # 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))
+            if reverse_transformer_layers_per_block is None
+            else reverse_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,
+                resolution_idx=i,
+                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,
+                attention_type=attention_type,
+                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,
+                dropout=dropout,
+                use_RoPE=use_RoPE,
+            )
+            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
+        )
+
+        self._set_pos_net_if_use_gligen(attention_type=attention_type, cross_attention_dim=cross_attention_dim)
+
+    def _check_config(
+        self,
+        down_block_types: Tuple[str],
+        up_block_types: Tuple[str],
+        only_cross_attention: Union[bool, Tuple[bool]],
+        block_out_channels: Tuple[int],
+        layers_per_block: Union[int, Tuple[int]],
+        cross_attention_dim: Union[int, Tuple[int]],
+        transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple[int]]],
+        reverse_transformer_layers_per_block: bool,
+        attention_head_dim: int,
+        num_attention_heads: Optional[Union[int, Tuple[int]]],
+    ):
+        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}."
+            )
+        if isinstance(transformer_layers_per_block, list) and reverse_transformer_layers_per_block is None:
+            for layer_number_per_block in transformer_layers_per_block:
+                if isinstance(layer_number_per_block, list):
+                    raise ValueError("Must provide 'reverse_transformer_layers_per_block` if using asymmetrical UNet.")
+
+    def _set_time_proj(
+        self,
+        time_embedding_type: str,
+        block_out_channels: int,
+        flip_sin_to_cos: bool,
+        freq_shift: float,
+        time_embedding_dim: int,
+    ) -> Tuple[int, int]:
+        if time_embedding_type == "fourier":
+            time_embed_dim = time_embedding_dim or block_out_channels[0] * 2
+            if time_embed_dim % 2 != 0:
+                raise ValueError(f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}.")
+            self.time_proj = GaussianFourierProjection(
+                time_embed_dim // 2, set_W_to_weight=False, log=False, flip_sin_to_cos=flip_sin_to_cos
+            )
+            timestep_input_dim = time_embed_dim
+        elif time_embedding_type == "positional":
+            time_embed_dim = time_embedding_dim or block_out_channels[0] * 4
+
+            self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
+            timestep_input_dim = block_out_channels[0]
+        else:
+            raise ValueError(
+                f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`."
+            )
+
+        return time_embed_dim, timestep_input_dim
+
+    def _set_encoder_hid_proj(
+        self,
+        encoder_hid_dim_type: Optional[str],
+        cross_attention_dim: Union[int, Tuple[int]],
+        encoder_hid_dim: Optional[int],
+    ):
+        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"` (Kandinsky 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
+
+    def _set_class_embedding(
+        self,
+        class_embed_type: Optional[str],
+        act_fn: str,
+        num_class_embeds: Optional[int],
+        projection_class_embeddings_input_dim: Optional[int],
+        time_embed_dim: int,
+        timestep_input_dim: int,
+    ):
+        if class_embed_type is None and num_class_embeds is not None:
+            self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
+        elif class_embed_type == "timestep":
+            self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim, act_fn=act_fn)
+        elif class_embed_type == "identity":
+            self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
+        elif class_embed_type == "projection":
+            if projection_class_embeddings_input_dim is None:
+                raise ValueError(
+                    "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
+                )
+            # The projection `class_embed_type` is the same as the timestep `class_embed_type` except
+            # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
+            # 2. it projects from an arbitrary input dimension.
+            #
+            # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
+            # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
+            # As a result, `TimestepEmbedding` can be passed arbitrary vectors.
+            self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+        elif class_embed_type == "simple_projection":
+            if projection_class_embeddings_input_dim is None:
+                raise ValueError(
+                    "`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set"
+                )
+            self.class_embedding = nn.Linear(projection_class_embeddings_input_dim, time_embed_dim)
+        else:
+            self.class_embedding = None
+
+    def _set_add_embedding(
+        self,
+        addition_embed_type: str,
+        addition_embed_type_num_heads: int,
+        addition_time_embed_dim: Optional[int],
+        flip_sin_to_cos: bool,
+        freq_shift: float,
+        cross_attention_dim: Optional[int],
+        encoder_hid_dim: Optional[int],
+        projection_class_embeddings_input_dim: Optional[int],
+        time_embed_dim: int,
+    ):
+        if addition_embed_type == "text":
+            if encoder_hid_dim is not None:
+                text_time_embedding_from_dim = encoder_hid_dim
+            else:
+                text_time_embedding_from_dim = cross_attention_dim
+
+            self.add_embedding = TextTimeEmbedding(
+                text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads
+            )
+        elif addition_embed_type == "text_image":
+            # text_embed_dim and image_embed_dim DON'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"` (Kandinsky 2.1)`
+            self.add_embedding = TextImageTimeEmbedding(
+                text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim
+            )
+        elif addition_embed_type == "text_time":
+            self.add_time_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift)
+            self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+        elif addition_embed_type == "image":
+            # Kandinsky 2.2
+            self.add_embedding = ImageTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
+        elif addition_embed_type == "image_hint":
+            # Kandinsky 2.2 ControlNet
+            self.add_embedding = ImageHintTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
+        elif addition_embed_type is not None:
+            raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.")
+
+    def _set_pos_net_if_use_gligen(self, attention_type: str, cross_attention_dim: int):
+        if attention_type in ["gated", "gated-text-image"]:
+            positive_len = 768
+            if isinstance(cross_attention_dim, int):
+                positive_len = cross_attention_dim
+            elif isinstance(cross_attention_dim, (list, tuple)):
+                positive_len = cross_attention_dim[0]
+
+            feature_type = "text-only" if attention_type == "gated" else "text-image"
+            self.position_net = GLIGENTextBoundingboxProjection(
+                positive_len=positive_len, out_dim=cross_attention_dim, feature_type=feature_type
+            )
+
+    @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, "get_processor"):
+                processors[f"{name}.processor"] = module.get_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.
+        """
+        if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnAddedKVProcessor()
+        elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnProcessor()
+        else:
+            raise ValueError(
+                f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+            )
+
+        self.set_attn_processor(processor)
+
+    def set_attention_slice(self, slice_size: Union[str, int, List[int]] = "auto"):
+        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):
+        self.gradient_checkpointing = True
+
+    def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
+        r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497.
+
+        The suffixes after the scaling factors represent the stage blocks where they are being applied.
+
+        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that
+        are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+        Args:
+            s1 (`float`):
+                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+                mitigate the "oversmoothing effect" in the enhanced denoising process.
+            s2 (`float`):
+                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+                mitigate the "oversmoothing effect" in the enhanced denoising process.
+            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+        """
+        for i, upsample_block in enumerate(self.up_blocks):
+            setattr(upsample_block, "s1", s1)
+            setattr(upsample_block, "s2", s2)
+            setattr(upsample_block, "b1", b1)
+            setattr(upsample_block, "b2", b2)
+
+    def disable_freeu(self):
+        """Disables the FreeU mechanism."""
+        freeu_keys = {"s1", "s2", "b1", "b2"}
+        for i, upsample_block in enumerate(self.up_blocks):
+            for k in freeu_keys:
+                if hasattr(upsample_block, k) or getattr(upsample_block, k, None) is not None:
+                    setattr(upsample_block, k, None)
+
+    def fuse_qkv_projections(self):
+        """
+        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value)
+        are fused. For cross-attention modules, key and value projection matrices are fused.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+        """
+        self.original_attn_processors = None
+
+        for _, attn_processor in self.attn_processors.items():
+            if "Added" in str(attn_processor.__class__.__name__):
+                raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
+
+        self.original_attn_processors = self.attn_processors
+
+        for module in self.modules():
+            if isinstance(module, Attention):
+                module.fuse_projections(fuse=True)
+
+        self.set_attn_processor(FusedAttnProcessor2_0())
+
+    def unfuse_qkv_projections(self):
+        """Disables the fused QKV projection if enabled.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        """
+        if self.original_attn_processors is not None:
+            self.set_attn_processor(self.original_attn_processors)
+
+    def get_time_embed(
+        self, sample: torch.Tensor, timestep: Union[torch.Tensor, float, int]
+    ) -> Optional[torch.Tensor]:
+        timesteps = timestep
+        if not torch.is_tensor(timesteps):
+            # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+            # This would be a good case for the `match` statement (Python 3.10+)
+            is_mps = sample.device.type == "mps"
+            if isinstance(timestep, float):
+                dtype = torch.float32 if is_mps else torch.float64
+            else:
+                dtype = torch.int32 if is_mps else torch.int64
+            timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+        elif len(timesteps.shape) == 0:
+            timesteps = timesteps[None].to(sample.device)
+
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        timesteps = timesteps.expand(sample.shape[0])
+
+        t_emb = self.time_proj(timesteps)
+        # `Timesteps` does not contain any weights and will always return f32 tensors
+        # but time_embedding might actually be running in fp16. so we need to cast here.
+        # there might be better ways to encapsulate this.
+        t_emb = t_emb.to(dtype=sample.dtype)
+        return t_emb
+
+    def get_class_embed(self, sample: torch.Tensor, class_labels: Optional[torch.Tensor]) -> Optional[torch.Tensor]:
+        class_emb = None
+        if self.class_embedding is not None:
+            if class_labels is None:
+                raise ValueError("class_labels should be provided when num_class_embeds > 0")
+
+            if self.config.class_embed_type == "timestep":
+                class_labels = self.time_proj(class_labels)
+
+                # `Timesteps` does not contain any weights and will always return f32 tensors
+                # there might be better ways to encapsulate this.
+                class_labels = class_labels.to(dtype=sample.dtype)
+
+            class_emb = self.class_embedding(class_labels).to(dtype=sample.dtype)
+        return class_emb
+
+    def get_aug_embed(
+        self, emb: torch.Tensor, encoder_hidden_states: torch.Tensor, added_cond_kwargs: Dict[str, Any]
+    ) -> Optional[torch.Tensor]:
+        aug_emb = None
+        if self.config.addition_embed_type == "text":
+            aug_emb = self.add_embedding(encoder_hidden_states)
+        elif self.config.addition_embed_type == "text_image":
+            # Kandinsky 2.1 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+                )
+
+            image_embs = added_cond_kwargs.get("image_embeds")
+            text_embs = added_cond_kwargs.get("text_embeds", encoder_hidden_states)
+            aug_emb = self.add_embedding(text_embs, image_embs)
+        elif self.config.addition_embed_type == "text_time":
+            # SDXL - style
+            if "text_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
+                )
+            text_embeds = added_cond_kwargs.get("text_embeds")
+            if "time_ids" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
+                )
+            time_ids = added_cond_kwargs.get("time_ids")
+            time_embeds = self.add_time_proj(time_ids.flatten())
+            time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
+            add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
+            add_embeds = add_embeds.to(emb.dtype)
+            aug_emb = self.add_embedding(add_embeds)
+        elif self.config.addition_embed_type == "image":
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+                )
+            image_embs = added_cond_kwargs.get("image_embeds")
+            aug_emb = self.add_embedding(image_embs)
+        elif self.config.addition_embed_type == "image_hint":
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs or "hint" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'image_hint' which requires the keyword arguments `image_embeds` and `hint` to be passed in `added_cond_kwargs`"
+                )
+            image_embs = added_cond_kwargs.get("image_embeds")
+            hint = added_cond_kwargs.get("hint")
+            aug_emb = self.add_embedding(image_embs, hint)
+        return aug_emb
+
+    def process_encoder_hidden_states(
+        self, encoder_hidden_states: torch.Tensor, added_cond_kwargs: Dict[str, Any]
+    ) -> torch.Tensor:
+        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":
+            # Kandinsky 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)
+        elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "ip_image_proj":
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'ip_image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+
+            if hasattr(self, "text_encoder_hid_proj") and self.text_encoder_hid_proj is not None:
+                encoder_hidden_states = self.text_encoder_hid_proj(encoder_hidden_states)
+
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            image_embeds = self.encoder_hid_proj(image_embeds)
+            encoder_hidden_states = (encoder_hidden_states, image_embeds)
+        return encoder_hidden_states
+
+    def forward(
+        self,
+        sample: torch.Tensor,
+        timestep: Union[torch.Tensor, float, int],
+        encoder_hidden_states: torch.Tensor,
+        class_labels: Optional[torch.Tensor] = None,
+        timestep_cond: Optional[torch.Tensor] = None,
+        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,
+        down_intrablock_additional_residuals: Optional[Tuple[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.Tensor`):
+                The noisy input tensor with the following shape `(batch, channel, height, width)`.
+            timestep (`torch.Tensor` or `float` or `int`): The number of timesteps to denoise an input.
+            encoder_hidden_states (`torch.Tensor`):
+                The encoder hidden states with shape `(batch, sequence_length, feature_dim)`.
+            class_labels (`torch.Tensor`, *optional*, defaults to `None`):
+                Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
+            timestep_cond: (`torch.Tensor`, *optional*, defaults to `None`):
+                Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed
+                through the `self.time_embedding` layer to obtain the timestep embeddings.
+            attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
+                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+                negative values to the attention scores corresponding to "discard" tokens.
+            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).
+            added_cond_kwargs: (`dict`, *optional*):
+                A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that
+                are passed along to the UNet blocks.
+            down_block_additional_residuals: (`tuple` of `torch.Tensor`, *optional*):
+                A tuple of tensors that if specified are added to the residuals of down unet blocks.
+            mid_block_additional_residual: (`torch.Tensor`, *optional*):
+                A tensor that if specified is added to the residual of the middle unet block.
+            down_intrablock_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
+                additional residuals to be added within UNet down blocks, for example from T2I-Adapter side model(s)
+            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.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+                tuple.
+
+        Returns:
+            [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
+                If `return_dict` is True, an [`~models.unets.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
+
+        # print("Sample shape: ", sample.shape)
+        # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
+        forward_upsample_size = False
+        upsample_size = None
+
+        for dim in sample.shape[-2:]:
+            if dim % default_overall_up_factor != 0:
+                # Forward upsample size to force interpolation output size.
+                forward_upsample_size = True
+                break
+
+        # 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
+        t_emb = self.get_time_embed(sample=sample, timestep=timestep)
+        emb = self.time_embedding(t_emb, timestep_cond)
+        aug_emb = None
+
+        class_emb = self.get_class_embed(sample=sample, class_labels=class_labels)
+        if class_emb is not None:
+            if self.config.class_embeddings_concat:
+                emb = torch.cat([emb, class_emb], dim=-1)
+            else:
+                emb = emb + class_emb
+
+        aug_emb = self.get_aug_embed(
+            emb=emb, encoder_hidden_states=encoder_hidden_states, added_cond_kwargs=added_cond_kwargs
+        )
+        if self.config.addition_embed_type == "image_hint":
+            aug_emb, hint = aug_emb
+            sample = torch.cat([sample, hint], dim=1)
+
+        emb = emb + aug_emb if aug_emb is not None else emb
+
+        if self.time_embed_act is not None:
+            emb = self.time_embed_act(emb)
+
+        encoder_hidden_states = self.process_encoder_hidden_states(
+            encoder_hidden_states=encoder_hidden_states, added_cond_kwargs=added_cond_kwargs
+        )
+
+        # 2. pre-process
+        if self.gradient_checkpointing:
+            sample = torch.utils.checkpoint.checkpoint(
+                self.conv_in,
+                sample,
+                use_reentrant=False,
+            )
+        else:
+            sample = self.conv_in(sample)
+
+        # 2.5 GLIGEN position net
+        if cross_attention_kwargs is not None and cross_attention_kwargs.get("gligen", None) is not None:
+            cross_attention_kwargs = cross_attention_kwargs.copy()
+            gligen_args = cross_attention_kwargs.pop("gligen")
+            cross_attention_kwargs["gligen"] = {"objs": self.position_net(**gligen_args)}
+
+        # 3. down
+        # we're popping the `scale` instead of getting it because otherwise `scale` will be propagated
+        # to the internal blocks and will raise deprecation warnings. this will be confusing for our users.
+        if cross_attention_kwargs is not None:
+            cross_attention_kwargs = cross_attention_kwargs.copy()
+            lora_scale = cross_attention_kwargs.pop("scale", 1.0)
+        else:
+            lora_scale = 1.0
+
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+
+        is_controlnet = mid_block_additional_residual is not None and down_block_additional_residuals is not None
+        # using new arg down_intrablock_additional_residuals for T2I-Adapters, to distinguish from controlnets
+        is_adapter = down_intrablock_additional_residuals is not None
+        # maintain backward compatibility for legacy usage, where
+        #       T2I-Adapter and ControlNet both use down_block_additional_residuals arg
+        #       but can only use one or the other
+        if not is_adapter and mid_block_additional_residual is None and down_block_additional_residuals is not None:
+            deprecate(
+                "T2I should not use down_block_additional_residuals",
+                "1.3.0",
+                "Passing intrablock residual connections with `down_block_additional_residuals` is deprecated \
+                    and will be removed in diffusers 1.3.0.  `down_block_additional_residuals` should only be used \
+                    for ControlNet. Please make sure use `down_intrablock_additional_residuals` instead. ",
+                standard_warn=False,
+            )
+            down_intrablock_additional_residuals = down_block_additional_residuals
+            is_adapter = True
+
+        down_block_res_samples = (sample,)
+        for i, downsample_block in enumerate(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_intrablock_additional_residuals) > 0:
+                    additional_residuals["additional_residuals"] = down_intrablock_additional_residuals.pop(0)
+
+                if self.gradient_checkpointing:
+                    sample, res_samples = torch.utils.checkpoint.checkpoint(
+                        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,
+                        use_reentrant=False,
+                    )
+                else:
+                    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:
+                if self.gradient_checkpointing:
+                    sample, res_samples = torch.utils.checkpoint.checkpoint(
+                        downsample_block,
+                        hidden_states=sample,
+                        temb=emb,
+                        use_reentrant=False,
+                    )
+                else:
+                    sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
+                if is_adapter and len(down_intrablock_additional_residuals) > 0:
+                    sample += down_intrablock_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:
+            # 
+            if hasattr(self.mid_block, "has_cross_attention") and self.mid_block.has_cross_attention:
+                if self.gradient_checkpointing:
+                    sample = torch.utils.checkpoint.checkpoint(
+                        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,
+                        use_reentrant=False,
+                    )
+                else:
+                    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,
+                    )
+            else:
+                if self.gradient_checkpointing:
+                    sample = torch.utils.checkpoint.checkpoint(
+                        self.mid_block,
+                        sample,
+                        emb,
+                        use_reentrant=False,
+                    )
+                else:
+                    sample = self.mid_block(sample, emb)
+
+            # To support T2I-Adapter-XL
+            if (
+                is_adapter
+                and len(down_intrablock_additional_residuals) > 0
+                and sample.shape == down_intrablock_additional_residuals[0].shape
+            ):
+                sample += down_intrablock_additional_residuals.pop(0)
+
+        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:
+                if self.gradient_checkpointing:
+                    sample = torch.utils.checkpoint.checkpoint(
+                        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,
+                        use_reentrant=False,
+                    )
+                else:
+                    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:
+                if self.gradient_checkpointing:
+                    sample = torch.utils.checkpoint.checkpoint(
+                        upsample_block,
+                        hidden_states=sample,
+                        temb=emb,
+                        res_hidden_states_tuple=res_samples,
+                        upsample_size=upsample_size,
+                        use_reentrant=False,
+                    )
+                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:
+            b, c, h, w = sample.shape
+            # sync_norm:
+            sample = rearrange(sample, '(b t) c h w -> b c (h w t)', t=6, h=h, w=w)
+            if self.gradient_checkpointing:
+                sample = torch.utils.checkpoint.checkpoint(
+                    self.conv_norm_out,
+                    sample,
+                    use_reentrant=False,
+                )
+            else:
+                sample = self.conv_norm_out(sample)
+            # sync_norm:
+            sample = rearrange(sample, 'b c (h w t) -> (b t) c h w', t=6, h=h, w=w)
+            
+            if self.gradient_checkpointing:
+                sample = torch.utils.checkpoint.checkpoint(
+                    self.conv_act,
+                    sample,
+                    use_reentrant=False,
+                )
+            else:
+                sample = self.conv_act(sample)
+        
+        if self.gradient_checkpointing:
+            sample = torch.utils.checkpoint.checkpoint(
+                self.conv_out,
+                sample,
+                use_reentrant=False,
+            )
+        else:
+            sample = self.conv_out(sample)
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return (sample,)
+
+        return UNet2DConditionOutput(sample=sample)

Marigold/vae/__init__.py

@@ -0,0 +1,7 @@
+from .vae import Encoder, Decoder, DecoderOutput, DiagonalGaussianDistribution
+
+__all__ = [
+    "AutoencoderKL",
+    "Encoder",
+    "Decoder"
+]

Marigold/vae/autoencoder_kl.py

@@ -0,0 +1,531 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Dict, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.loaders.single_file_model import FromOriginalModelMixin
+from diffusers.utils.accelerate_utils import apply_forward_hook
+import os
+import sys
+sys.path.append(os.path.abspath(os.path.join(os.getcwd(), "../unet")))
+from Marigold.unet.attention_processor import (
+    ADDED_KV_ATTENTION_PROCESSORS,
+    CROSS_ATTENTION_PROCESSORS,
+    Attention,
+    AttentionProcessor,
+    AttnAddedKVProcessor,
+    AttnProcessor,
+    FusedAttnProcessor2_0,
+)
+from diffusers.models.modeling_outputs import AutoencoderKLOutput
+from diffusers.models.modeling_utils import ModelMixin
+from Marigold.vae.vae import Decoder, DecoderOutput, DiagonalGaussianDistribution, Encoder
+
+
+class AutoencoderKL(ModelMixin, ConfigMixin, FromOriginalModelMixin):
+    r"""
+    A VAE model with KL loss for encoding images into latents and decoding latent representations into images.
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+    for all models (such as downloading or saving).
+
+    Parameters:
+        in_channels (int, *optional*, defaults to 3): Number of channels in the input image.
+        out_channels (int,  *optional*, defaults to 3): Number of channels in the output.
+        down_block_types (`Tuple[str]`, *optional*, defaults to `("DownEncoderBlock2D",)`):
+            Tuple of downsample block types.
+        up_block_types (`Tuple[str]`, *optional*, defaults to `("UpDecoderBlock2D",)`):
+            Tuple of upsample block types.
+        block_out_channels (`Tuple[int]`, *optional*, defaults to `(64,)`):
+            Tuple of block output channels.
+        act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
+        latent_channels (`int`, *optional*, defaults to 4): Number of channels in the latent space.
+        sample_size (`int`, *optional*, defaults to `32`): Sample input size.
+        scaling_factor (`float`, *optional*, defaults to 0.18215):
+            The component-wise standard deviation of the trained latent space computed using the first batch of the
+            training set. This is used to scale the latent space to have unit variance when training the diffusion
+            model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the
+            diffusion model. When decoding, the latents are scaled back to the original scale with the formula: `z = 1
+            / scaling_factor * z`. For more details, refer to sections 4.3.2 and D.1 of the [High-Resolution Image
+            Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) paper.
+        force_upcast (`bool`, *optional*, default to `True`):
+            If enabled it will force the VAE to run in float32 for high image resolution pipelines, such as SD-XL. VAE
+            can be fine-tuned / trained to a lower range without loosing too much precision in which case
+            `force_upcast` can be set to `False` - see: https://huggingface.co/madebyollin/sdxl-vae-fp16-fix
+        mid_block_add_attention (`bool`, *optional*, default to `True`):
+            If enabled, the mid_block of the Encoder and Decoder will have attention blocks. If set to false, the
+            mid_block will only have resnet blocks
+    """
+
+    _supports_gradient_checkpointing = True
+    _no_split_modules = ["BasicTransformerBlock", "ResnetBlock2D"]
+
+    @register_to_config
+    def __init__(
+        self,
+        in_channels: int = 3,
+        out_channels: int = 3,
+        down_block_types: Tuple[str] = ("DownEncoderBlock2D",),
+        up_block_types: Tuple[str] = ("UpDecoderBlock2D",),
+        block_out_channels: Tuple[int] = (64,),
+        layers_per_block: int = 1,
+        act_fn: str = "silu",
+        latent_channels: int = 4,
+        norm_num_groups: int = 32,
+        sample_size: int = 32,
+        scaling_factor: float = 0.18215,
+        shift_factor: Optional[float] = None,
+        latents_mean: Optional[Tuple[float]] = None,
+        latents_std: Optional[Tuple[float]] = None,
+        force_upcast: float = True,
+        use_quant_conv: bool = True,
+        use_post_quant_conv: bool = True,
+        mid_block_add_attention: bool = True,
+          # whether to use synchronized batch norm
+        use_RoPE: bool = False,  # whether to use RoPE positional encoding
+    ):
+        super().__init__()
+
+        # pass init params to Encoder
+        self.encoder = Encoder(
+            in_channels=in_channels,
+            out_channels=latent_channels,
+            down_block_types=down_block_types,
+            block_out_channels=block_out_channels,
+            layers_per_block=layers_per_block,
+            act_fn=act_fn,
+            norm_num_groups=norm_num_groups,
+            double_z=True,
+            mid_block_add_attention=mid_block_add_attention,
+            
+            use_RoPE=use_RoPE,
+        )
+
+        # pass init params to Decoder
+        self.decoder = Decoder(
+            in_channels=latent_channels,
+            out_channels=out_channels,
+            up_block_types=up_block_types,
+            block_out_channels=block_out_channels,
+            layers_per_block=layers_per_block,
+            norm_num_groups=norm_num_groups,
+            act_fn=act_fn,
+            mid_block_add_attention=mid_block_add_attention,
+            
+            use_RoPE=use_RoPE,
+        )
+
+        self.quant_conv = nn.Conv2d(2 * latent_channels, 2 * latent_channels, 1) if use_quant_conv else None
+        self.post_quant_conv = nn.Conv2d(latent_channels, latent_channels, 1) if use_post_quant_conv else None
+
+        self.use_slicing = False
+        self.use_tiling = False
+
+        # only relevant if vae tiling is enabled
+        self.tile_sample_min_size = self.config.sample_size
+        sample_size = (
+            self.config.sample_size[0]
+            if isinstance(self.config.sample_size, (list, tuple))
+            else self.config.sample_size
+        )
+        self.tile_latent_min_size = int(sample_size / (2 ** (len(self.config.block_out_channels) - 1)))
+        self.tile_overlap_factor = 0.25
+
+    def _set_gradient_checkpointing(self):
+            self.encoder.gradient_checkpointing = True
+            self.decoder.gradient_checkpointing = True
+
+    def enable_tiling(self, use_tiling: bool = True):
+        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.use_tiling = use_tiling
+
+    def disable_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_tiling` was previously enabled, this method will go back to computing
+        decoding in one step.
+        """
+        self.enable_tiling(False)
+
+    def enable_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.use_slicing = True
+
+    def disable_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_slicing` was previously enabled, this method will go back to computing
+        decoding in one step.
+        """
+        self.use_slicing = False
+
+    @property
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
+    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, "get_processor"):
+                processors[f"{name}.processor"] = module.get_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
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+    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)
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
+    def set_default_attn_processor(self):
+        """
+        Disables custom attention processors and sets the default attention implementation.
+        """
+        if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnAddedKVProcessor()
+        elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnProcessor()
+        else:
+            raise ValueError(
+                f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+            )
+
+        self.set_attn_processor(processor)
+
+    @apply_forward_hook
+    def encode(
+        self, x: torch.Tensor, deterministic: bool = False, return_dict: bool = True
+    ) -> Union[AutoencoderKLOutput, Tuple[DiagonalGaussianDistribution]]:
+        """
+        Encode a batch of images into latents.
+
+        Args:
+            x (`torch.Tensor`): Input batch of images.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple.
+
+        Returns:
+                The latent representations of the encoded images. If `return_dict` is True, a
+                [`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain `tuple` is returned.
+        """
+        if self.use_tiling and (x.shape[-1] > self.tile_sample_min_size or x.shape[-2] > self.tile_sample_min_size):
+            return self.tiled_encode(x, return_dict=return_dict)
+
+        if self.use_slicing and x.shape[0] > 1:
+            encoded_slices = [self.encoder(x_slice) for x_slice in x.split(1)]
+            h = torch.cat(encoded_slices)
+        else:
+            h = self.encoder(x)
+
+        if self.quant_conv is not None:
+            moments = self.quant_conv(h)
+        else:
+            moments = h
+
+        if deterministic:
+            latent, _ = torch.chunk(moments, 2, dim=1)
+            return latent
+
+        posterior = DiagonalGaussianDistribution(moments)
+
+        if not return_dict:
+            return (posterior,)
+
+        return AutoencoderKLOutput(latent_dist=posterior)
+
+    def _decode(self, z: torch.Tensor, deterministic: bool = False, return_dict: bool = True) -> Union[DecoderOutput, torch.Tensor]:
+        if self.use_tiling and (z.shape[-1] > self.tile_latent_min_size or z.shape[-2] > self.tile_latent_min_size):
+            return self.tiled_decode(z, return_dict=return_dict)
+
+        if self.post_quant_conv is not None:
+            z = self.post_quant_conv(z)
+
+        dec = self.decoder(z)
+
+        if deterministic:
+            return dec
+        
+        if not return_dict:
+            return (dec,)
+
+        return DecoderOutput(sample=dec)
+
+    @apply_forward_hook
+    def decode(
+        self, z: torch.FloatTensor, deterministic: bool = False, return_dict: bool = True, generator=None
+    ) -> Union[DecoderOutput, torch.FloatTensor]:
+        """
+        Decode a batch of images.
+
+        Args:
+            z (`torch.Tensor`): Input batch of latent vectors.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether to return a [`~models.vae.DecoderOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.vae.DecoderOutput`] or `tuple`:
+                If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is
+                returned.
+
+        """
+        if self.use_slicing and z.shape[0] > 1:
+            decoded_slices = [self._decode(z_slice).sample for z_slice in z.split(1)]
+            decoded = torch.cat(decoded_slices)
+        else:
+            decoded = self._decode(z, deterministic)
+        
+        if deterministic:
+            return decoded
+        else:
+            decoded = decoded.sample
+
+        if not return_dict:
+            return (decoded,)
+
+        return DecoderOutput(sample=decoded)
+
+    def blend_v(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor:
+        blend_extent = min(a.shape[2], b.shape[2], blend_extent)
+        for y in range(blend_extent):
+            b[:, :, y, :] = a[:, :, -blend_extent + y, :] * (1 - y / blend_extent) + b[:, :, y, :] * (y / blend_extent)
+        return b
+
+    def blend_h(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor:
+        blend_extent = min(a.shape[3], b.shape[3], blend_extent)
+        for x in range(blend_extent):
+            b[:, :, :, x] = a[:, :, :, -blend_extent + x] * (1 - x / blend_extent) + b[:, :, :, x] * (x / blend_extent)
+        return b
+
+    def tiled_encode(self, x: torch.Tensor, return_dict: bool = True) -> AutoencoderKLOutput:
+        r"""Encode a batch of images using a tiled encoder.
+
+        When this option is enabled, the VAE will split the input tensor into tiles to compute encoding in several
+        steps. This is useful to keep memory use constant regardless of image size. The end result of tiled encoding is
+        different from non-tiled encoding because each tile uses a different encoder. To avoid tiling artifacts, the
+        tiles overlap and are blended together to form a smooth output. You may still see tile-sized changes in the
+        output, but they should be much less noticeable.
+
+        Args:
+            x (`torch.Tensor`): Input batch of images.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.autoencoder_kl.AutoencoderKLOutput`] or `tuple`:
+                If return_dict is True, a [`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain
+                `tuple` is returned.
+        """
+        overlap_size = int(self.tile_sample_min_size * (1 - self.tile_overlap_factor))
+        blend_extent = int(self.tile_latent_min_size * self.tile_overlap_factor)
+        row_limit = self.tile_latent_min_size - blend_extent
+
+        # Split the image into 512x512 tiles and encode them separately.
+        rows = []
+        for i in range(0, x.shape[2], overlap_size):
+            row = []
+            for j in range(0, x.shape[3], overlap_size):
+                tile = x[:, :, i : i + self.tile_sample_min_size, j : j + self.tile_sample_min_size]
+                tile = self.encoder(tile)
+                if self.config.use_quant_conv:
+                    tile = self.quant_conv(tile)
+                row.append(tile)
+            rows.append(row)
+        result_rows = []
+        for i, row in enumerate(rows):
+            result_row = []
+            for j, tile in enumerate(row):
+                # blend the above tile and the left tile
+                # to the current tile and add the current tile to the result row
+                if i > 0:
+                    tile = self.blend_v(rows[i - 1][j], tile, blend_extent)
+                if j > 0:
+                    tile = self.blend_h(row[j - 1], tile, blend_extent)
+                result_row.append(tile[:, :, :row_limit, :row_limit])
+            result_rows.append(torch.cat(result_row, dim=3))
+
+        moments = torch.cat(result_rows, dim=2)
+        posterior = DiagonalGaussianDistribution(moments)
+
+        if not return_dict:
+            return (posterior,)
+
+        return AutoencoderKLOutput(latent_dist=posterior)
+
+    def tiled_decode(self, z: torch.Tensor, return_dict: bool = True) -> Union[DecoderOutput, torch.Tensor]:
+        r"""
+        Decode a batch of images using a tiled decoder.
+
+        Args:
+            z (`torch.Tensor`): Input batch of latent vectors.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.vae.DecoderOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.vae.DecoderOutput`] or `tuple`:
+                If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is
+                returned.
+        """
+        overlap_size = int(self.tile_latent_min_size * (1 - self.tile_overlap_factor))
+        blend_extent = int(self.tile_sample_min_size * self.tile_overlap_factor)
+        row_limit = self.tile_sample_min_size - blend_extent
+
+        # Split z into overlapping 64x64 tiles and decode them separately.
+        # The tiles have an overlap to avoid seams between tiles.
+        rows = []
+        for i in range(0, z.shape[2], overlap_size):
+            row = []
+            for j in range(0, z.shape[3], overlap_size):
+                tile = z[:, :, i : i + self.tile_latent_min_size, j : j + self.tile_latent_min_size]
+                if self.config.use_post_quant_conv:
+                    tile = self.post_quant_conv(tile)
+                decoded = self.decoder(tile)
+                row.append(decoded)
+            rows.append(row)
+        result_rows = []
+        for i, row in enumerate(rows):
+            result_row = []
+            for j, tile in enumerate(row):
+                # blend the above tile and the left tile
+                # to the current tile and add the current tile to the result row
+                if i > 0:
+                    tile = self.blend_v(rows[i - 1][j], tile, blend_extent)
+                if j > 0:
+                    tile = self.blend_h(row[j - 1], tile, blend_extent)
+                result_row.append(tile[:, :, :row_limit, :row_limit])
+            result_rows.append(torch.cat(result_row, dim=3))
+
+        dec = torch.cat(result_rows, dim=2)
+        if not return_dict:
+            return (dec,)
+
+        return DecoderOutput(sample=dec)
+
+    def forward(
+        self,
+        sample: torch.Tensor,
+        sample_posterior: bool = False,
+        return_dict: bool = True,
+        generator: Optional[torch.Generator] = None,
+    ) -> Union[DecoderOutput, torch.Tensor]:
+        r"""
+        Args:
+            sample (`torch.Tensor`): Input sample.
+            sample_posterior (`bool`, *optional*, defaults to `False`):
+                Whether to sample from the posterior.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`DecoderOutput`] instead of a plain tuple.
+        """
+        x = sample
+        posterior = self.encode(x).latent_dist
+        if sample_posterior:
+            z = posterior.sample(generator=generator)
+        else:
+            z = posterior.mode()
+        dec = self.decode(z).sample
+
+        if not return_dict:
+            return (dec,)
+
+        return DecoderOutput(sample=dec)
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections
+    def fuse_qkv_projections(self):
+        """
+        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value)
+        are fused. For cross-attention modules, key and value projection matrices are fused.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+        """
+        self.original_attn_processors = None
+
+        for _, attn_processor in self.attn_processors.items():
+            if "Added" in str(attn_processor.__class__.__name__):
+                raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
+
+        self.original_attn_processors = self.attn_processors
+
+        for module in self.modules():
+            if isinstance(module, Attention):
+                module.fuse_projections(fuse=True)
+
+        self.set_attn_processor(FusedAttnProcessor2_0())
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
+    def unfuse_qkv_projections(self):
+        """Disables the fused QKV projection if enabled.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        """
+        if self.original_attn_processors is not None:
+            self.set_attn_processor(self.original_attn_processors)

Marigold/vae/vae.py

@@ -0,0 +1,1015 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from dataclasses import dataclass
+from typing import Optional, Tuple
+
+import numpy as np
+import torch
+import torch.nn as nn
+
+from diffusers.utils import BaseOutput, is_torch_version
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.models.activations import get_activation
+import sys
+import os
+sys.path.append(os.path.abspath(os.path.join(os.getcwd(), "../unet")))
+
+from Marigold.unet.attention_processor import SpatialNorm
+
+from Marigold.unet.unet_2d_blocks import (
+    AutoencoderTinyBlock,
+    UNetMidBlock2D,
+    get_down_block,
+    get_up_block,
+)
+from einops import rearrange
+
+@dataclass
+class DecoderOutput(BaseOutput):
+    r"""
+    Output of decoding method.
+
+    Args:
+        sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)`):
+            The decoded output sample from the last layer of the model.
+    """
+
+    sample: torch.Tensor
+    commit_loss: Optional[torch.FloatTensor] = None
+
+
+class Encoder(nn.Module):
+    r"""
+    The `Encoder` layer of a variational autoencoder that encodes its input into a latent representation.
+
+    Args:
+        in_channels (`int`, *optional*, defaults to 3):
+            The number of input channels.
+        out_channels (`int`, *optional*, defaults to 3):
+            The number of output channels.
+        down_block_types (`Tuple[str, ...]`, *optional*, defaults to `("DownEncoderBlock2D",)`):
+            The types of down blocks to use. See `~diffusers.models.unet_2d_blocks.get_down_block` for available
+            options.
+        block_out_channels (`Tuple[int, ...]`, *optional*, defaults to `(64,)`):
+            The number of output channels for each block.
+        layers_per_block (`int`, *optional*, defaults to 2):
+            The number of layers per block.
+        norm_num_groups (`int`, *optional*, defaults to 32):
+            The number of groups for normalization.
+        act_fn (`str`, *optional*, defaults to `"silu"`):
+            The activation function to use. See `~diffusers.models.activations.get_activation` for available options.
+        double_z (`bool`, *optional*, defaults to `True`):
+            Whether to double the number of output channels for the last block.
+    """
+
+    def __init__(
+        self,
+        in_channels: int = 3,
+        out_channels: int = 3,
+        down_block_types: Tuple[str, ...] = ("DownEncoderBlock2D",),
+        block_out_channels: Tuple[int, ...] = (64,),
+        layers_per_block: int = 2,
+        norm_num_groups: int = 32,
+        act_fn: str = "silu",
+        double_z: bool = True,
+        mid_block_add_attention=True,
+        
+        padding_mode: str = "zeros",
+        use_RoPE: bool = False,  # whether to use RoPE positional encoding
+    ):
+        super().__init__()
+        self.layers_per_block = layers_per_block
+        
+        self.conv_in = nn.Conv2d(
+            in_channels,
+            block_out_channels[0],
+            kernel_size=3,
+            stride=1,
+            padding=1,
+            padding_mode=padding_mode,
+        )
+
+        self.down_blocks = nn.ModuleList([])
+
+        # 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=self.layers_per_block,
+                in_channels=input_channel,
+                out_channels=output_channel,
+                add_downsample=not is_final_block,
+                resnet_eps=1e-6,
+                downsample_padding=0,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                attention_head_dim=output_channel,
+                temb_channels=None,
+                
+            )
+            self.down_blocks.append(down_block)
+
+        # mid
+        self.mid_block = UNetMidBlock2D(
+            in_channels=block_out_channels[-1],
+            resnet_eps=1e-6,
+            resnet_act_fn=act_fn,
+            output_scale_factor=1,
+            resnet_time_scale_shift="default",
+            attention_head_dim=block_out_channels[-1],
+            resnet_groups=norm_num_groups,
+            temb_channels=None,
+            add_attention=mid_block_add_attention,
+            
+            use_RoPE=use_RoPE,
+        )
+
+        # out
+        self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[-1], num_groups=norm_num_groups, eps=1e-6)
+        self.conv_act = nn.SiLU()
+
+        conv_out_channels = 2 * out_channels if double_z else out_channels
+        self.conv_out = nn.Conv2d(block_out_channels[-1], conv_out_channels, 3, padding=1, padding_mode=padding_mode)
+
+        self.gradient_checkpointing = False
+
+    def forward(self, sample: torch.Tensor) -> torch.Tensor:
+        r"""The forward method of the `Encoder` class."""
+
+        sample = self.conv_in(sample)
+
+        if self.training and self.gradient_checkpointing:
+
+            def create_custom_forward(module):
+                def custom_forward(*inputs):
+                    return module(*inputs)
+
+                return custom_forward
+
+            # down
+            if is_torch_version(">=", "1.11.0"):
+                for down_block in self.down_blocks:
+                    sample = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(down_block), sample, use_reentrant=False
+                    )
+                # middle
+                sample = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(self.mid_block), sample, use_reentrant=False
+                )
+            else:
+                for down_block in self.down_blocks:
+                    sample = torch.utils.checkpoint.checkpoint(create_custom_forward(down_block), sample)
+                # middle
+                sample = torch.utils.checkpoint.checkpoint(create_custom_forward(self.mid_block), sample)
+
+        else:
+            # down
+            for down_block in self.down_blocks:
+                sample = down_block(sample)
+
+            # middle
+            sample = self.mid_block(sample)
+
+        b, c, h, w = sample.shape
+        # post-process
+        # sync_norm:
+        sample = rearrange(sample, '(b t) c h w -> b c (h w t)', t=6, h=h, w=w)
+
+        sample = self.conv_norm_out(sample)
+
+        # sync_norm:
+        sample = rearrange(sample, 'b c (h w t) -> (b t) c h w', t=6, h=h, w=w)
+        
+        sample = self.conv_act(sample)
+        sample = self.conv_out(sample)
+
+        return sample
+
+
+class Decoder(nn.Module):
+    r"""
+    The `Decoder` layer of a variational autoencoder that decodes its latent representation into an output sample.
+
+    Args:
+        in_channels (`int`, *optional*, defaults to 3):
+            The number of input channels.
+        out_channels (`int`, *optional*, defaults to 3):
+            The number of output channels.
+        up_block_types (`Tuple[str, ...]`, *optional*, defaults to `("UpDecoderBlock2D",)`):
+            The types of up blocks to use. See `~diffusers.models.unet_2d_blocks.get_up_block` for available options.
+        block_out_channels (`Tuple[int, ...]`, *optional*, defaults to `(64,)`):
+            The number of output channels for each block.
+        layers_per_block (`int`, *optional*, defaults to 2):
+            The number of layers per block.
+        norm_num_groups (`int`, *optional*, defaults to 32):
+            The number of groups for normalization.
+        act_fn (`str`, *optional*, defaults to `"silu"`):
+            The activation function to use. See `~diffusers.models.activations.get_activation` for available options.
+        norm_type (`str`, *optional*, defaults to `"group"`):
+            The normalization type to use. Can be either `"group"` or `"spatial"`.
+    """
+
+    def __init__(
+        self,
+        in_channels: int = 3,
+        out_channels: int = 3,
+        up_block_types: Tuple[str, ...] = ("UpDecoderBlock2D",),
+        block_out_channels: Tuple[int, ...] = (64,),
+        layers_per_block: int = 2,
+        norm_num_groups: int = 32,
+        act_fn: str = "silu",
+        norm_type: str = "group",  # group, spatial
+        mid_block_add_attention=True,
+        
+        use_RoPE: bool = False,  # whether to use RoPE positional encoding
+    ):
+        super().__init__()
+        self.layers_per_block = layers_per_block
+        
+        self.conv_in = nn.Conv2d(
+            in_channels,
+            block_out_channels[-1],
+            kernel_size=3,
+            stride=1,
+            padding=1,
+        )
+
+        self.up_blocks = nn.ModuleList([])
+
+        temb_channels = in_channels if norm_type == "spatial" else None
+
+        # mid
+        self.mid_block = UNetMidBlock2D(
+            in_channels=block_out_channels[-1],
+            resnet_eps=1e-6,
+            resnet_act_fn=act_fn,
+            output_scale_factor=1,
+            resnet_time_scale_shift="default" if norm_type == "group" else norm_type,
+            attention_head_dim=block_out_channels[-1],
+            resnet_groups=norm_num_groups,
+            temb_channels=temb_channels,
+            add_attention=mid_block_add_attention,
+            
+            use_RoPE=use_RoPE,
+        )
+
+        # up
+        reversed_block_out_channels = list(reversed(block_out_channels))
+        output_channel = reversed_block_out_channels[0]
+        for i, up_block_type in enumerate(up_block_types):
+            prev_output_channel = output_channel
+            output_channel = reversed_block_out_channels[i]
+
+            is_final_block = i == len(block_out_channels) - 1
+
+            up_block = get_up_block(
+                up_block_type,
+                num_layers=self.layers_per_block + 1,
+                in_channels=prev_output_channel,
+                out_channels=output_channel,
+                prev_output_channel=None,
+                add_upsample=not is_final_block,
+                resnet_eps=1e-6,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                attention_head_dim=output_channel,
+                temb_channels=temb_channels,
+                resnet_time_scale_shift=norm_type,
+                
+            )
+            self.up_blocks.append(up_block)
+            prev_output_channel = output_channel
+
+        # out
+        if norm_type == "spatial":
+            self.conv_norm_out = SpatialNorm(block_out_channels[0], temb_channels)
+        else:
+            self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=1e-6)
+        self.conv_act = nn.SiLU()
+        self.conv_out = nn.Conv2d(block_out_channels[0], out_channels, 3, padding=1)
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        sample: torch.Tensor,
+        latent_embeds: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        r"""The forward method of the `Decoder` class."""
+
+        sample = self.conv_in(sample)
+        
+
+        upscale_dtype = next(iter(self.up_blocks.parameters())).dtype
+        if self.gradient_checkpointing:
+
+            def create_custom_forward(module):
+                def custom_forward(*inputs):
+                    return module(*inputs)
+
+                return custom_forward
+
+            if is_torch_version(">=", "1.11.0"):
+                # middle
+                sample = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(self.mid_block),
+                    sample,
+                    latent_embeds,
+                    use_reentrant=False,
+                )
+                sample = sample.to(upscale_dtype)
+
+                # up
+                for up_block in self.up_blocks:
+                    sample = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(up_block),
+                        sample,
+                        latent_embeds,
+                        use_reentrant=False,
+                    )
+            else:
+                # middle
+                sample = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(self.mid_block), sample, latent_embeds
+                )
+                sample = sample.to(upscale_dtype)
+
+                # up
+                for up_block in self.up_blocks:
+                    sample = torch.utils.checkpoint.checkpoint(create_custom_forward(up_block), sample, latent_embeds)
+        else:
+            # middle
+            sample = self.mid_block(sample, latent_embeds)
+            sample = sample.to(upscale_dtype)
+
+            # up
+            for up_block in self.up_blocks:
+                sample = up_block(sample, latent_embeds)
+        
+        b, c, h, w = sample.shape
+        # sync_norm:
+        sample = rearrange(sample, '(b t) c h w -> b c (h w t)', t=6, h=h, w=w)
+
+        # post-process
+        if latent_embeds is None:
+            sample = self.conv_norm_out(sample)
+        else:
+            sample = self.conv_norm_out(sample, latent_embeds)
+        
+        # sync_norm:
+        sample = rearrange(sample, 'b c (h w t) -> (b t) c h w', t=6, h=h, w=w)
+
+        sample = self.conv_act(sample)
+        sample = self.conv_out(sample)
+
+        return sample
+
+
+class UpSample(nn.Module):
+    r"""
+    The `UpSample` layer of a variational autoencoder that upsamples its input.
+
+    Args:
+        in_channels (`int`, *optional*, defaults to 3):
+            The number of input channels.
+        out_channels (`int`, *optional*, defaults to 3):
+            The number of output channels.
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+    ) -> None:
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.deconv = nn.ConvTranspose2d(in_channels, out_channels, kernel_size=4, stride=2, padding=1)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        r"""The forward method of the `UpSample` class."""
+        x = torch.relu(x)
+        x = self.deconv(x)
+        return x
+
+
+class MaskConditionEncoder(nn.Module):
+    """
+    used in AsymmetricAutoencoderKL
+    """
+
+    def __init__(
+        self,
+        in_ch: int,
+        out_ch: int = 192,
+        res_ch: int = 768,
+        stride: int = 16,
+    ) -> None:
+        super().__init__()
+
+        channels = []
+        while stride > 1:
+            stride = stride // 2
+            in_ch_ = out_ch * 2
+            if out_ch > res_ch:
+                out_ch = res_ch
+            if stride == 1:
+                in_ch_ = res_ch
+            channels.append((in_ch_, out_ch))
+            out_ch *= 2
+
+        out_channels = []
+        for _in_ch, _out_ch in channels:
+            out_channels.append(_out_ch)
+        out_channels.append(channels[-1][0])
+
+        layers = []
+        in_ch_ = in_ch
+        for l in range(len(out_channels)):
+            out_ch_ = out_channels[l]
+            if l == 0 or l == 1:
+                layers.append(nn.Conv2d(in_ch_, out_ch_, kernel_size=3, stride=1, padding=1))
+            else:
+                layers.append(nn.Conv2d(in_ch_, out_ch_, kernel_size=4, stride=2, padding=1))
+            in_ch_ = out_ch_
+
+        self.layers = nn.Sequential(*layers)
+
+    def forward(self, x: torch.Tensor, mask=None) -> torch.Tensor:
+        r"""The forward method of the `MaskConditionEncoder` class."""
+        out = {}
+        for l in range(len(self.layers)):
+            layer = self.layers[l]
+            x = layer(x)
+            out[str(tuple(x.shape))] = x
+            x = torch.relu(x)
+        return out
+
+
+class MaskConditionDecoder(nn.Module):
+    r"""The `MaskConditionDecoder` should be used in combination with [`AsymmetricAutoencoderKL`] to enhance the model's
+    decoder with a conditioner on the mask and masked image.
+
+    Args:
+        in_channels (`int`, *optional*, defaults to 3):
+            The number of input channels.
+        out_channels (`int`, *optional*, defaults to 3):
+            The number of output channels.
+        up_block_types (`Tuple[str, ...]`, *optional*, defaults to `("UpDecoderBlock2D",)`):
+            The types of up blocks to use. See `~diffusers.models.unet_2d_blocks.get_up_block` for available options.
+        block_out_channels (`Tuple[int, ...]`, *optional*, defaults to `(64,)`):
+            The number of output channels for each block.
+        layers_per_block (`int`, *optional*, defaults to 2):
+            The number of layers per block.
+        norm_num_groups (`int`, *optional*, defaults to 32):
+            The number of groups for normalization.
+        act_fn (`str`, *optional*, defaults to `"silu"`):
+            The activation function to use. See `~diffusers.models.activations.get_activation` for available options.
+        norm_type (`str`, *optional*, defaults to `"group"`):
+            The normalization type to use. Can be either `"group"` or `"spatial"`.
+    """
+
+    def __init__(
+        self,
+        in_channels: int = 3,
+        out_channels: int = 3,
+        up_block_types: Tuple[str, ...] = ("UpDecoderBlock2D",),
+        block_out_channels: Tuple[int, ...] = (64,),
+        layers_per_block: int = 2,
+        norm_num_groups: int = 32,
+        act_fn: str = "silu",
+        norm_type: str = "group",  # group, spatial
+    ):
+        super().__init__()
+        self.layers_per_block = layers_per_block
+
+        self.conv_in = nn.Conv2d(
+            in_channels,
+            block_out_channels[-1],
+            kernel_size=3,
+            stride=1,
+            padding=1,
+        )
+
+        self.up_blocks = nn.ModuleList([])
+
+        temb_channels = in_channels if norm_type == "spatial" else None
+
+        # mid
+        self.mid_block = UNetMidBlock2D(
+            in_channels=block_out_channels[-1],
+            resnet_eps=1e-6,
+            resnet_act_fn=act_fn,
+            output_scale_factor=1,
+            resnet_time_scale_shift="default" if norm_type == "group" else norm_type,
+            attention_head_dim=block_out_channels[-1],
+            resnet_groups=norm_num_groups,
+            temb_channels=temb_channels,
+        )
+
+        # up
+        reversed_block_out_channels = list(reversed(block_out_channels))
+        output_channel = reversed_block_out_channels[0]
+        for i, up_block_type in enumerate(up_block_types):
+            prev_output_channel = output_channel
+            output_channel = reversed_block_out_channels[i]
+
+            is_final_block = i == len(block_out_channels) - 1
+
+            up_block = get_up_block(
+                up_block_type,
+                num_layers=self.layers_per_block + 1,
+                in_channels=prev_output_channel,
+                out_channels=output_channel,
+                prev_output_channel=None,
+                add_upsample=not is_final_block,
+                resnet_eps=1e-6,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                attention_head_dim=output_channel,
+                temb_channels=temb_channels,
+                resnet_time_scale_shift=norm_type,
+            )
+            self.up_blocks.append(up_block)
+            prev_output_channel = output_channel
+
+        # condition encoder
+        self.condition_encoder = MaskConditionEncoder(
+            in_ch=out_channels,
+            out_ch=block_out_channels[0],
+            res_ch=block_out_channels[-1],
+        )
+
+        # out
+        if norm_type == "spatial":
+            self.conv_norm_out = SpatialNorm(block_out_channels[0], temb_channels)
+        else:
+            self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=1e-6)
+        self.conv_act = nn.SiLU()
+        self.conv_out = nn.Conv2d(block_out_channels[0], out_channels, 3, padding=1)
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        z: torch.Tensor,
+        image: Optional[torch.Tensor] = None,
+        mask: Optional[torch.Tensor] = None,
+        latent_embeds: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        r"""The forward method of the `MaskConditionDecoder` class."""
+        sample = z
+        sample = self.conv_in(sample)
+
+        upscale_dtype = next(iter(self.up_blocks.parameters())).dtype
+        if self.training and self.gradient_checkpointing:
+
+            def create_custom_forward(module):
+                def custom_forward(*inputs):
+                    return module(*inputs)
+
+                return custom_forward
+
+            if is_torch_version(">=", "1.11.0"):
+                # middle
+                sample = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(self.mid_block),
+                    sample,
+                    latent_embeds,
+                    use_reentrant=False,
+                )
+                sample = sample.to(upscale_dtype)
+
+                # condition encoder
+                if image is not None and mask is not None:
+                    masked_image = (1 - mask) * image
+                    im_x = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(self.condition_encoder),
+                        masked_image,
+                        mask,
+                        use_reentrant=False,
+                    )
+
+                # up
+                for up_block in self.up_blocks:
+                    if image is not None and mask is not None:
+                        sample_ = im_x[str(tuple(sample.shape))]
+                        mask_ = nn.functional.interpolate(mask, size=sample.shape[-2:], mode="nearest")
+                        sample = sample * mask_ + sample_ * (1 - mask_)
+                    sample = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(up_block),
+                        sample,
+                        latent_embeds,
+                        use_reentrant=False,
+                    )
+                if image is not None and mask is not None:
+                    sample = sample * mask + im_x[str(tuple(sample.shape))] * (1 - mask)
+            else:
+                # middle
+                sample = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(self.mid_block), sample, latent_embeds
+                )
+                sample = sample.to(upscale_dtype)
+
+                # condition encoder
+                if image is not None and mask is not None:
+                    masked_image = (1 - mask) * image
+                    im_x = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(self.condition_encoder),
+                        masked_image,
+                        mask,
+                    )
+
+                # up
+                for up_block in self.up_blocks:
+                    if image is not None and mask is not None:
+                        sample_ = im_x[str(tuple(sample.shape))]
+                        mask_ = nn.functional.interpolate(mask, size=sample.shape[-2:], mode="nearest")
+                        sample = sample * mask_ + sample_ * (1 - mask_)
+                    sample = torch.utils.checkpoint.checkpoint(create_custom_forward(up_block), sample, latent_embeds)
+                if image is not None and mask is not None:
+                    sample = sample * mask + im_x[str(tuple(sample.shape))] * (1 - mask)
+        else:
+            # middle
+            sample = self.mid_block(sample, latent_embeds)
+            sample = sample.to(upscale_dtype)
+
+            # condition encoder
+            if image is not None and mask is not None:
+                masked_image = (1 - mask) * image
+                im_x = self.condition_encoder(masked_image, mask)
+
+            # up
+            for up_block in self.up_blocks:
+                if image is not None and mask is not None:
+                    sample_ = im_x[str(tuple(sample.shape))]
+                    mask_ = nn.functional.interpolate(mask, size=sample.shape[-2:], mode="nearest")
+                    sample = sample * mask_ + sample_ * (1 - mask_)
+                sample = up_block(sample, latent_embeds)
+            if image is not None and mask is not None:
+                sample = sample * mask + im_x[str(tuple(sample.shape))] * (1 - mask)
+
+        # post-process
+        if latent_embeds is None:
+            sample = self.conv_norm_out(sample)
+        else:
+            sample = self.conv_norm_out(sample, latent_embeds)
+        sample = self.conv_act(sample)
+        sample = self.conv_out(sample)
+
+        return sample
+
+
+class VectorQuantizer(nn.Module):
+    """
+    Improved version over VectorQuantizer, can be used as a drop-in replacement. Mostly avoids costly matrix
+    multiplications and allows for post-hoc remapping of indices.
+    """
+
+    # NOTE: due to a bug the beta term was applied to the wrong term. for
+    # backwards compatibility we use the buggy version by default, but you can
+    # specify legacy=False to fix it.
+    def __init__(
+        self,
+        n_e: int,
+        vq_embed_dim: int,
+        beta: float,
+        remap=None,
+        unknown_index: str = "random",
+        sane_index_shape: bool = False,
+        legacy: bool = True,
+    ):
+        super().__init__()
+        self.n_e = n_e
+        self.vq_embed_dim = vq_embed_dim
+        self.beta = beta
+        self.legacy = legacy
+
+        self.embedding = nn.Embedding(self.n_e, self.vq_embed_dim)
+        self.embedding.weight.data.uniform_(-1.0 / self.n_e, 1.0 / self.n_e)
+
+        self.remap = remap
+        if self.remap is not None:
+            self.register_buffer("used", torch.tensor(np.load(self.remap)))
+            self.used: torch.Tensor
+            self.re_embed = self.used.shape[0]
+            self.unknown_index = unknown_index  # "random" or "extra" or integer
+            if self.unknown_index == "extra":
+                self.unknown_index = self.re_embed
+                self.re_embed = self.re_embed + 1
+            print(
+                f"Remapping {self.n_e} indices to {self.re_embed} indices. "
+                f"Using {self.unknown_index} for unknown indices."
+            )
+        else:
+            self.re_embed = n_e
+
+        self.sane_index_shape = sane_index_shape
+
+    def remap_to_used(self, inds: torch.LongTensor) -> torch.LongTensor:
+        ishape = inds.shape
+        assert len(ishape) > 1
+        inds = inds.reshape(ishape[0], -1)
+        used = self.used.to(inds)
+        match = (inds[:, :, None] == used[None, None, ...]).long()
+        new = match.argmax(-1)
+        unknown = match.sum(2) < 1
+        if self.unknown_index == "random":
+            new[unknown] = torch.randint(0, self.re_embed, size=new[unknown].shape).to(device=new.device)
+        else:
+            new[unknown] = self.unknown_index
+        return new.reshape(ishape)
+
+    def unmap_to_all(self, inds: torch.LongTensor) -> torch.LongTensor:
+        ishape = inds.shape
+        assert len(ishape) > 1
+        inds = inds.reshape(ishape[0], -1)
+        used = self.used.to(inds)
+        if self.re_embed > self.used.shape[0]:  # extra token
+            inds[inds >= self.used.shape[0]] = 0  # simply set to zero
+        back = torch.gather(used[None, :][inds.shape[0] * [0], :], 1, inds)
+        return back.reshape(ishape)
+
+    def forward(self, z: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor, Tuple]:
+        # reshape z -> (batch, height, width, channel) and flatten
+        z = z.permute(0, 2, 3, 1).contiguous()
+        z_flattened = z.view(-1, self.vq_embed_dim)
+
+        # distances from z to embeddings e_j (z - e)^2 = z^2 + e^2 - 2 e * z
+        min_encoding_indices = torch.argmin(torch.cdist(z_flattened, self.embedding.weight), dim=1)
+
+        z_q = self.embedding(min_encoding_indices).view(z.shape)
+        perplexity = None
+        min_encodings = None
+
+        # compute loss for embedding
+        if not self.legacy:
+            loss = self.beta * torch.mean((z_q.detach() - z) ** 2) + torch.mean((z_q - z.detach()) ** 2)
+        else:
+            loss = torch.mean((z_q.detach() - z) ** 2) + self.beta * torch.mean((z_q - z.detach()) ** 2)
+
+        # preserve gradients
+        z_q: torch.Tensor = z + (z_q - z).detach()
+
+        # reshape back to match original input shape
+        z_q = z_q.permute(0, 3, 1, 2).contiguous()
+
+        if self.remap is not None:
+            min_encoding_indices = min_encoding_indices.reshape(z.shape[0], -1)  # add batch axis
+            min_encoding_indices = self.remap_to_used(min_encoding_indices)
+            min_encoding_indices = min_encoding_indices.reshape(-1, 1)  # flatten
+
+        if self.sane_index_shape:
+            min_encoding_indices = min_encoding_indices.reshape(z_q.shape[0], z_q.shape[2], z_q.shape[3])
+
+        return z_q, loss, (perplexity, min_encodings, min_encoding_indices)
+
+    def get_codebook_entry(self, indices: torch.LongTensor, shape: Tuple[int, ...]) -> torch.Tensor:
+        # shape specifying (batch, height, width, channel)
+        if self.remap is not None:
+            indices = indices.reshape(shape[0], -1)  # add batch axis
+            indices = self.unmap_to_all(indices)
+            indices = indices.reshape(-1)  # flatten again
+
+        # get quantized latent vectors
+        z_q: torch.Tensor = self.embedding(indices)
+
+        if shape is not None:
+            z_q = z_q.view(shape)
+            # reshape back to match original input shape
+            z_q = z_q.permute(0, 3, 1, 2).contiguous()
+
+        return z_q
+
+
+class DiagonalGaussianDistribution(object):
+    def __init__(self, parameters: torch.Tensor, deterministic: bool = False):
+        self.parameters = parameters
+        self.mean, self.logvar = torch.chunk(parameters, 2, dim=1)
+        self.logvar = torch.clamp(self.logvar, -30.0, 20.0)
+        self.deterministic = deterministic
+        self.std = torch.exp(0.5 * self.logvar)
+        self.var = torch.exp(self.logvar)
+        if self.deterministic:
+            self.var = self.std = torch.zeros_like(
+                self.mean, device=self.parameters.device, dtype=self.parameters.dtype
+            )
+
+    def sample(self, generator: Optional[torch.Generator] = None) -> torch.Tensor:
+        # make sure sample is on the same device as the parameters and has same dtype
+        sample = randn_tensor(
+            self.mean.shape,
+            generator=generator,
+            device=self.parameters.device,
+            dtype=self.parameters.dtype,
+        )
+        x = self.mean + self.std * sample
+        return x
+
+    def kl(self, other: "DiagonalGaussianDistribution" = None) -> torch.Tensor:
+        if self.deterministic:
+            return torch.Tensor([0.0])
+        else:
+            if other is None:
+                return 0.5 * torch.sum(
+                    torch.pow(self.mean, 2) + self.var - 1.0 - self.logvar,
+                    dim=[1, 2, 3],
+                )
+            else:
+                return 0.5 * torch.sum(
+                    torch.pow(self.mean - other.mean, 2) / other.var
+                    + self.var / other.var
+                    - 1.0
+                    - self.logvar
+                    + other.logvar,
+                    dim=[1, 2, 3],
+                )
+
+    def nll(self, sample: torch.Tensor, dims: Tuple[int, ...] = [1, 2, 3]) -> torch.Tensor:
+        if self.deterministic:
+            return torch.Tensor([0.0])
+        logtwopi = np.log(2.0 * np.pi)
+        return 0.5 * torch.sum(
+            logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var,
+            dim=dims,
+        )
+
+    def mode(self) -> torch.Tensor:
+        return self.mean
+
+
+class EncoderTiny(nn.Module):
+    r"""
+    The `EncoderTiny` layer is a simpler version of the `Encoder` layer.
+
+    Args:
+        in_channels (`int`):
+            The number of input channels.
+        out_channels (`int`):
+            The number of output channels.
+        num_blocks (`Tuple[int, ...]`):
+            Each value of the tuple represents a Conv2d layer followed by `value` number of `AutoencoderTinyBlock`'s to
+            use.
+        block_out_channels (`Tuple[int, ...]`):
+            The number of output channels for each block.
+        act_fn (`str`):
+            The activation function to use. See `~diffusers.models.activations.get_activation` for available options.
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        num_blocks: Tuple[int, ...],
+        block_out_channels: Tuple[int, ...],
+        act_fn: str,
+    ):
+        super().__init__()
+        layers = []
+        for i, num_block in enumerate(num_blocks):
+            num_channels = block_out_channels[i]
+
+            if i == 0:
+                layers.append(nn.Conv2d(in_channels, num_channels, kernel_size=3, padding=1))
+            else:
+                layers.append(
+                    nn.Conv2d(
+                        num_channels,
+                        num_channels,
+                        kernel_size=3,
+                        padding=1,
+                        stride=2,
+                        bias=False,
+                    )
+                )
+
+            for _ in range(num_block):
+                layers.append(AutoencoderTinyBlock(num_channels, num_channels, act_fn))
+
+        layers.append(nn.Conv2d(block_out_channels[-1], out_channels, kernel_size=3, padding=1))
+
+        self.layers = nn.Sequential(*layers)
+        self.gradient_checkpointing = False
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        r"""The forward method of the `EncoderTiny` class."""
+        if self.training and self.gradient_checkpointing:
+
+            def create_custom_forward(module):
+                def custom_forward(*inputs):
+                    return module(*inputs)
+
+                return custom_forward
+
+            if is_torch_version(">=", "1.11.0"):
+                x = torch.utils.checkpoint.checkpoint(create_custom_forward(self.layers), x, use_reentrant=False)
+            else:
+                x = torch.utils.checkpoint.checkpoint(create_custom_forward(self.layers), x)
+
+        else:
+            # scale image from [-1, 1] to [0, 1] to match TAESD convention
+            x = self.layers(x.add(1).div(2))
+
+        return x
+
+
+class DecoderTiny(nn.Module):
+    r"""
+    The `DecoderTiny` layer is a simpler version of the `Decoder` layer.
+
+    Args:
+        in_channels (`int`):
+            The number of input channels.
+        out_channels (`int`):
+            The number of output channels.
+        num_blocks (`Tuple[int, ...]`):
+            Each value of the tuple represents a Conv2d layer followed by `value` number of `AutoencoderTinyBlock`'s to
+            use.
+        block_out_channels (`Tuple[int, ...]`):
+            The number of output channels for each block.
+        upsampling_scaling_factor (`int`):
+            The scaling factor to use for upsampling.
+        act_fn (`str`):
+            The activation function to use. See `~diffusers.models.activations.get_activation` for available options.
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        num_blocks: Tuple[int, ...],
+        block_out_channels: Tuple[int, ...],
+        upsampling_scaling_factor: int,
+        act_fn: str,
+        upsample_fn: str,
+    ):
+        super().__init__()
+
+        layers = [
+            nn.Conv2d(in_channels, block_out_channels[0], kernel_size=3, padding=1),
+            get_activation(act_fn),
+        ]
+
+        for i, num_block in enumerate(num_blocks):
+            is_final_block = i == (len(num_blocks) - 1)
+            num_channels = block_out_channels[i]
+
+            for _ in range(num_block):
+                layers.append(AutoencoderTinyBlock(num_channels, num_channels, act_fn))
+
+            if not is_final_block:
+                layers.append(nn.Upsample(scale_factor=upsampling_scaling_factor, mode=upsample_fn))
+
+            conv_out_channel = num_channels if not is_final_block else out_channels
+            layers.append(
+                nn.Conv2d(
+                    num_channels,
+                    conv_out_channel,
+                    kernel_size=3,
+                    padding=1,
+                    bias=is_final_block,
+                )
+            )
+
+        self.layers = nn.Sequential(*layers)
+        self.gradient_checkpointing = False
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        r"""The forward method of the `DecoderTiny` class."""
+        # Clamp.
+        x = torch.tanh(x / 3) * 3
+
+        if self.training and self.gradient_checkpointing:
+
+            def create_custom_forward(module):
+                def custom_forward(*inputs):
+                    return module(*inputs)
+
+                return custom_forward
+
+            if is_torch_version(">=", "1.11.0"):
+                x = torch.utils.checkpoint.checkpoint(create_custom_forward(self.layers), x, use_reentrant=False)
+            else:
+                x = torch.utils.checkpoint.checkpoint(create_custom_forward(self.layers), x)
+
+        else:
+            x = self.layers(x)
+
+        # scale image from [0, 1] to [-1, 1] to match diffusers convention
+        return x.mul(2).sub(1)