sd

diffsynth/models/stable_diffusion_unet.py

@@ -0,0 +1,912 @@
+# Copyright 2025 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import math
+from typing import Optional
+
+
+# ===== Time Embedding =====
+
+class Timesteps(nn.Module):
+    def __init__(self, num_channels, flip_sin_to_cos=True, freq_shift=0):
+        super().__init__()
+        self.num_channels = num_channels
+        self.flip_sin_to_cos = flip_sin_to_cos
+        self.freq_shift = freq_shift
+
+    def forward(self, timesteps):
+        half_dim = self.num_channels // 2
+        exponent = -math.log(10000) * torch.arange(half_dim, dtype=torch.float32, device=timesteps.device)
+        exponent = exponent / half_dim + self.freq_shift
+        emb = torch.exp(exponent)
+        emb = timesteps[:, None].float() * emb[None, :]
+        sin_emb = torch.sin(emb)
+        cos_emb = torch.cos(emb)
+        if self.flip_sin_to_cos:
+            emb = torch.cat([cos_emb, sin_emb], dim=-1)
+        else:
+            emb = torch.cat([sin_emb, cos_emb], dim=-1)
+        return emb
+
+
+class TimestepEmbedding(nn.Module):
+    def __init__(self, in_channels, time_embed_dim, act_fn="silu", out_dim=None):
+        super().__init__()
+        self.linear_1 = nn.Linear(in_channels, time_embed_dim)
+        self.act = nn.SiLU() if act_fn == "silu" else nn.GELU()
+        out_dim = out_dim if out_dim is not None else time_embed_dim
+        self.linear_2 = nn.Linear(time_embed_dim, out_dim)
+
+    def forward(self, sample):
+        sample = self.linear_1(sample)
+        sample = self.act(sample)
+        sample = self.linear_2(sample)
+        return sample
+
+
+# ===== ResNet Blocks =====
+
+class ResnetBlock2D(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels=None,
+        conv_shortcut=False,
+        dropout=0.0,
+        temb_channels=512,
+        groups=32,
+        groups_out=None,
+        pre_norm=True,
+        eps=1e-6,
+        non_linearity="swish",
+        time_embedding_norm="default",
+        output_scale_factor=1.0,
+        use_in_shortcut=None,
+    ):
+        super().__init__()
+        self.pre_norm = pre_norm
+        self.time_embedding_norm = time_embedding_norm
+        self.output_scale_factor = output_scale_factor
+
+        if groups_out is None:
+            groups_out = groups
+
+        self.norm1 = nn.GroupNorm(num_groups=groups, num_channels=in_channels, eps=eps)
+        self.conv1 = nn.Conv2d(in_channels, out_channels or in_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 or in_channels)
+            elif self.time_embedding_norm == "scale_shift":
+                self.time_emb_proj = nn.Linear(temb_channels, 2 * (out_channels or in_channels))
+
+        self.norm2 = nn.GroupNorm(num_groups=groups_out, num_channels=out_channels or in_channels, eps=eps)
+        self.dropout = nn.Dropout(dropout)
+        self.conv2 = nn.Conv2d(out_channels or in_channels, out_channels or in_channels, kernel_size=3, stride=1, padding=1)
+
+        if non_linearity == "swish":
+            self.nonlinearity = nn.SiLU()
+        elif non_linearity == "silu":
+            self.nonlinearity = nn.SiLU()
+        elif non_linearity == "gelu":
+            self.nonlinearity = nn.GELU()
+        elif non_linearity == "relu":
+            self.nonlinearity = nn.ReLU()
+
+        self.use_conv_shortcut = conv_shortcut
+        self.conv_shortcut = None
+        if conv_shortcut:
+            self.conv_shortcut = nn.Conv2d(in_channels, out_channels or in_channels, kernel_size=1, stride=1, padding=0)
+        else:
+            self.conv_shortcut = nn.Conv2d(in_channels, out_channels or in_channels, kernel_size=1, stride=1, padding=0) if in_channels != (out_channels or in_channels) else None
+
+    def forward(self, input_tensor, temb=None):
+        hidden_states = input_tensor
+        hidden_states = self.norm1(hidden_states)
+        hidden_states = self.nonlinearity(hidden_states)
+        hidden_states = self.conv1(hidden_states)
+
+        if temb is not None:
+            temb = self.nonlinearity(temb)
+            temb = self.time_emb_proj(temb).unsqueeze(-1).unsqueeze(-1)
+
+        if temb is not None and self.time_embedding_norm == "default":
+            hidden_states = hidden_states + temb
+
+        hidden_states = self.norm2(hidden_states)
+
+        if temb is not None and self.time_embedding_norm == "scale_shift":
+            scale, shift = torch.chunk(temb, 2, dim=1)
+            hidden_states = hidden_states * (1 + scale) + shift
+
+        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
+
+
+# ===== Transformer Blocks =====
+
+class GEGLU(nn.Module):
+    def __init__(self, dim_in, dim_out):
+        super().__init__()
+        self.proj = nn.Linear(dim_in, dim_out * 2)
+
+    def forward(self, hidden_states):
+        hidden_states, gate = self.proj(hidden_states).chunk(2, dim=-1)
+        return hidden_states * F.gelu(gate)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, dim_out=None, dropout=0.0):
+        super().__init__()
+        self.net = nn.ModuleList([
+            GEGLU(dim, dim * 4),
+            nn.Dropout(dropout),
+            nn.Linear(dim * 4, dim if dim_out is None else dim_out),
+        ])
+
+    def forward(self, hidden_states):
+        for module in self.net:
+            hidden_states = module(hidden_states)
+        return hidden_states
+
+
+class Attention(nn.Module):
+    """Attention block matching diffusers checkpoint key format.
+    Keys: to_q.weight, to_k.weight, to_v.weight, to_out.0.weight, to_out.0.bias
+    """
+    def __init__(
+        self,
+        query_dim,
+        heads=8,
+        dim_head=64,
+        dropout=0.0,
+        bias=False,
+        upcast_attention=False,
+        cross_attention_dim=None,
+    ):
+        super().__init__()
+        inner_dim = dim_head * heads
+        self.heads = heads
+        self.inner_dim = inner_dim
+        self.cross_attention_dim = cross_attention_dim if cross_attention_dim is not None else query_dim
+
+        self.to_q = nn.Linear(query_dim, inner_dim, bias=bias)
+        self.to_k = nn.Linear(self.cross_attention_dim, inner_dim, bias=bias)
+        self.to_v = nn.Linear(self.cross_attention_dim, inner_dim, bias=bias)
+        self.to_out = nn.ModuleList([
+            nn.Linear(inner_dim, query_dim, bias=True),
+            nn.Dropout(dropout),
+        ])
+
+    def forward(self, hidden_states, encoder_hidden_states=None, attention_mask=None):
+        # Query
+        query = self.to_q(hidden_states)
+        batch_size, seq_len, _ = query.shape
+
+        # Key/Value
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        key = self.to_k(encoder_hidden_states)
+        value = self.to_v(encoder_hidden_states)
+
+        # Reshape for multi-head attention
+        head_dim = self.inner_dim // self.heads
+        query = query.view(batch_size, -1, self.heads, head_dim).transpose(1, 2)
+        key = key.view(batch_size, -1, self.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, self.heads, head_dim).transpose(1, 2)
+
+        # Scaled dot-product attention
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=None, dropout_p=0.0, is_causal=False
+        )
+
+        # Reshape back
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, self.inner_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        # Output projection
+        hidden_states = self.to_out[0](hidden_states)
+        hidden_states = self.to_out[1](hidden_states)
+
+        return hidden_states
+
+
+class BasicTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim,
+        n_heads,
+        d_head,
+        dropout=0.0,
+        cross_attention_dim=None,
+        upcast_attention=False,
+    ):
+        super().__init__()
+        self.norm1 = nn.LayerNorm(dim)
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=n_heads,
+            dim_head=d_head,
+            dropout=dropout,
+            bias=False,
+            upcast_attention=upcast_attention,
+        )
+        self.norm2 = nn.LayerNorm(dim)
+        self.attn2 = Attention(
+            query_dim=dim,
+            heads=n_heads,
+            dim_head=d_head,
+            dropout=dropout,
+            bias=False,
+            upcast_attention=upcast_attention,
+            cross_attention_dim=cross_attention_dim,
+        )
+        self.norm3 = nn.LayerNorm(dim)
+        self.ff = FeedForward(dim, dropout=dropout)
+
+    def forward(self, hidden_states, encoder_hidden_states=None, attention_mask=None):
+        # Self-attention
+        attn_output = self.attn1(self.norm1(hidden_states))
+        hidden_states = attn_output + hidden_states
+        # Cross-attention
+        attn_output = self.attn2(self.norm2(hidden_states), encoder_hidden_states=encoder_hidden_states)
+        hidden_states = attn_output + hidden_states
+        # Feed-forward
+        ff_output = self.ff(self.norm3(hidden_states))
+        hidden_states = ff_output + hidden_states
+        return hidden_states
+
+
+class Transformer2DModel(nn.Module):
+    """2D Transformer block wrapper matching diffusers checkpoint structure.
+    Keys: norm.weight/bias, proj_in.weight/bias, transformer_blocks.X.*, proj_out.weight/bias
+    """
+    def __init__(
+        self,
+        num_attention_heads=16,
+        attention_head_dim=64,
+        in_channels=320,
+        num_layers=1,
+        dropout=0.0,
+        norm_num_groups=32,
+        cross_attention_dim=768,
+        upcast_attention=False,
+    ):
+        super().__init__()
+        self.norm = nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6)
+        self.proj_in = nn.Conv2d(in_channels, num_attention_heads * attention_head_dim, kernel_size=1, bias=True)
+
+        self.transformer_blocks = nn.ModuleList([
+            BasicTransformerBlock(
+                dim=num_attention_heads * attention_head_dim,
+                n_heads=num_attention_heads,
+                d_head=attention_head_dim,
+                dropout=dropout,
+                cross_attention_dim=cross_attention_dim,
+                upcast_attention=upcast_attention,
+            )
+            for _ in range(num_layers)
+        ])
+
+        self.proj_out = nn.Conv2d(num_attention_heads * attention_head_dim, in_channels, kernel_size=1, bias=True)
+
+    def forward(self, hidden_states, encoder_hidden_states=None, attention_mask=None):
+        batch, channel, height, width = hidden_states.shape
+        residual = hidden_states
+
+        # Normalize and project to sequence
+        hidden_states = self.norm(hidden_states)
+        hidden_states = self.proj_in(hidden_states)
+        hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, -1, channel)
+
+        # Transformer blocks
+        for block in self.transformer_blocks:
+            hidden_states = block(hidden_states, encoder_hidden_states=encoder_hidden_states)
+
+        # Project back to 2D
+        hidden_states = hidden_states.reshape(batch, height, width, channel).permute(0, 3, 1, 2).contiguous()
+        hidden_states = self.proj_out(hidden_states)
+        hidden_states = hidden_states + residual
+        return hidden_states
+
+
+# ===== Down/Up Blocks =====
+
+class CrossAttnDownBlock2D(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        temb_channels=1280,
+        dropout=0.0,
+        num_layers=1,
+        transformer_layers_per_block=1,
+        resnet_eps=1e-6,
+        resnet_time_scale_shift="default",
+        resnet_act_fn="swish",
+        resnet_groups=32,
+        resnet_pre_norm=True,
+        cross_attention_dim=768,
+        attention_head_dim=1,
+        downsample=True,
+    ):
+        super().__init__()
+        self.has_cross_attention = True
+
+        resnets = []
+        attentions = []
+
+        for i in range(num_layers):
+            in_channels_i = in_channels if i == 0 else out_channels
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels_i,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=1.0,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+            attentions.append(
+                Transformer2DModel(
+                    num_attention_heads=attention_head_dim,
+                    attention_head_dim=out_channels // attention_head_dim,
+                    in_channels=out_channels,
+                    num_layers=transformer_layers_per_block,
+                    dropout=dropout,
+                    norm_num_groups=resnet_groups,
+                    cross_attention_dim=cross_attention_dim,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        if downsample:
+            self.downsamplers = nn.ModuleList([
+                Downsample2D(out_channels, out_channels, padding=1)
+            ])
+        else:
+            self.downsamplers = None
+
+    def forward(self, hidden_states, temb=None, encoder_hidden_states=None):
+        output_states = []
+
+        for resnet, attn in zip(self.resnets, self.attentions):
+            hidden_states = resnet(hidden_states, temb)
+            hidden_states = attn(hidden_states, encoder_hidden_states=encoder_hidden_states)
+            output_states.append(hidden_states)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states)
+            output_states.append(hidden_states)
+
+        return hidden_states, tuple(output_states)
+
+
+class DownBlock2D(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        temb_channels=1280,
+        dropout=0.0,
+        num_layers=1,
+        resnet_eps=1e-6,
+        resnet_time_scale_shift="default",
+        resnet_act_fn="swish",
+        resnet_groups=32,
+        resnet_pre_norm=True,
+        downsample=True,
+    ):
+        super().__init__()
+        self.has_cross_attention = False
+
+        resnets = []
+        for i in range(num_layers):
+            in_channels_i = in_channels if i == 0 else out_channels
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels_i,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=1.0,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+        self.resnets = nn.ModuleList(resnets)
+
+        if downsample:
+            self.downsamplers = nn.ModuleList([
+                Downsample2D(out_channels, out_channels, padding=1)
+            ])
+        else:
+            self.downsamplers = None
+
+    def forward(self, hidden_states, temb=None, encoder_hidden_states=None):
+        output_states = []
+        for resnet in self.resnets:
+            hidden_states = resnet(hidden_states, temb)
+            output_states.append(hidden_states)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states)
+            output_states.append(hidden_states)
+
+        return hidden_states, tuple(output_states)
+
+
+class CrossAttnUpBlock2D(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        prev_output_channel,
+        temb_channels=1280,
+        dropout=0.0,
+        num_layers=1,
+        transformer_layers_per_block=1,
+        resnet_eps=1e-6,
+        resnet_time_scale_shift="default",
+        resnet_act_fn="swish",
+        resnet_groups=32,
+        resnet_pre_norm=True,
+        cross_attention_dim=768,
+        attention_head_dim=1,
+        upsample=True,
+    ):
+        super().__init__()
+        self.has_cross_attention = True
+
+        resnets = []
+        attentions = []
+
+        for i in range(num_layers):
+            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+            resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=resnet_in_channels + res_skip_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=1.0,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+            attentions.append(
+                Transformer2DModel(
+                    num_attention_heads=attention_head_dim,
+                    attention_head_dim=out_channels // attention_head_dim,
+                    in_channels=out_channels,
+                    num_layers=transformer_layers_per_block,
+                    dropout=dropout,
+                    norm_num_groups=resnet_groups,
+                    cross_attention_dim=cross_attention_dim,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        if upsample:
+            self.upsamplers = nn.ModuleList([
+                Upsample2D(out_channels, out_channels)
+            ])
+        else:
+            self.upsamplers = None
+
+    def forward(self, hidden_states, res_hidden_states_tuple, temb=None, encoder_hidden_states=None, upsample_size=None):
+        for resnet, attn in zip(self.resnets, self.attentions):
+            # Pop res hidden states
+            res_hidden_states = res_hidden_states_tuple[-1]
+            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+            hidden_states = resnet(hidden_states, temb)
+            hidden_states = attn(hidden_states, encoder_hidden_states=encoder_hidden_states)
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(hidden_states, upsample_size=upsample_size)
+
+        return hidden_states
+
+
+class UpBlock2D(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        prev_output_channel,
+        temb_channels=1280,
+        dropout=0.0,
+        num_layers=1,
+        resnet_eps=1e-6,
+        resnet_time_scale_shift="default",
+        resnet_act_fn="swish",
+        resnet_groups=32,
+        resnet_pre_norm=True,
+        upsample=True,
+    ):
+        super().__init__()
+        self.has_cross_attention = False
+
+        resnets = []
+        for i in range(num_layers):
+            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+            resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=resnet_in_channels + res_skip_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=1.0,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+        self.resnets = nn.ModuleList(resnets)
+
+        if upsample:
+            self.upsamplers = nn.ModuleList([
+                Upsample2D(out_channels, out_channels)
+            ])
+        else:
+            self.upsamplers = None
+
+    def forward(self, hidden_states, res_hidden_states_tuple, temb=None, encoder_hidden_states=None, upsample_size=None):
+        for resnet in self.resnets:
+            res_hidden_states = res_hidden_states_tuple[-1]
+            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+            hidden_states = resnet(hidden_states, temb)
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(hidden_states, upsample_size=upsample_size)
+
+        return hidden_states
+
+
+# ===== UNet Mid Block =====
+
+class UNetMidBlock2DCrossAttn(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        temb_channels=1280,
+        dropout=0.0,
+        num_layers=1,
+        transformer_layers_per_block=1,
+        resnet_eps=1e-6,
+        resnet_time_scale_shift="default",
+        resnet_act_fn="swish",
+        resnet_groups=32,
+        resnet_pre_norm=True,
+        cross_attention_dim=768,
+        attention_head_dim=1,
+    ):
+        super().__init__()
+        resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+
+        # There is always at least one resnet
+        resnets = [
+            ResnetBlock2D(
+                in_channels=in_channels,
+                out_channels=in_channels,
+                temb_channels=temb_channels,
+                eps=resnet_eps,
+                groups=resnet_groups,
+                dropout=dropout,
+                time_embedding_norm=resnet_time_scale_shift,
+                non_linearity=resnet_act_fn,
+                output_scale_factor=1.0,
+                pre_norm=resnet_pre_norm,
+            )
+        ]
+        attentions = []
+
+        for _ in range(num_layers):
+            attentions.append(
+                Transformer2DModel(
+                    num_attention_heads=attention_head_dim,
+                    attention_head_dim=in_channels // attention_head_dim,
+                    in_channels=in_channels,
+                    num_layers=transformer_layers_per_block,
+                    dropout=dropout,
+                    norm_num_groups=resnet_groups,
+                    cross_attention_dim=cross_attention_dim,
+                )
+            )
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=in_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=1.0,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+    def forward(self, hidden_states, temb=None, encoder_hidden_states=None):
+        hidden_states = self.resnets[0](hidden_states, temb)
+        for attn, resnet in zip(self.attentions, self.resnets[1:]):
+            hidden_states = attn(hidden_states, encoder_hidden_states=encoder_hidden_states)
+            hidden_states = resnet(hidden_states, temb)
+        return hidden_states
+
+
+# ===== Downsample / Upsample =====
+
+class Downsample2D(nn.Module):
+    def __init__(self, in_channels, out_channels, padding=1):
+        super().__init__()
+        self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=2, padding=padding)
+        self.padding = padding
+
+    def forward(self, hidden_states):
+        if self.padding == 0:
+            hidden_states = F.pad(hidden_states, (0, 1, 0, 1), mode="constant", value=0)
+        return self.conv(hidden_states)
+
+
+class Upsample2D(nn.Module):
+    def __init__(self, in_channels, out_channels):
+        super().__init__()
+        self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1)
+
+    def forward(self, hidden_states, upsample_size=None):
+        if upsample_size is not None:
+            hidden_states = F.interpolate(hidden_states, size=upsample_size, mode="nearest")
+        else:
+            hidden_states = F.interpolate(hidden_states, scale_factor=2.0, mode="nearest")
+        return self.conv(hidden_states)
+
+
+# ===== UNet2DConditionModel =====
+
+class UNet2DConditionModel(nn.Module):
+    """Stable Diffusion UNet with cross-attention conditioning.
+    state_dict keys match the diffusers UNet2DConditionModel checkpoint format.
+    """
+    def __init__(
+        self,
+        sample_size=64,
+        in_channels=4,
+        out_channels=4,
+        down_block_types=("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D"),
+        up_block_types=("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"),
+        block_out_channels=(320, 640, 1280, 1280),
+        layers_per_block=2,
+        cross_attention_dim=768,
+        attention_head_dim=8,
+        norm_num_groups=32,
+        norm_eps=1e-5,
+        dropout=0.0,
+        act_fn="silu",
+        time_embedding_type="positional",
+        flip_sin_to_cos=True,
+        freq_shift=0,
+        time_embedding_dim=None,
+        resnet_time_scale_shift="default",
+        upcast_attention=False,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.sample_size = sample_size
+
+        # Time embedding
+        timestep_embedding_dim = time_embedding_dim or block_out_channels[0]
+        self.time_proj = Timesteps(timestep_embedding_dim, flip_sin_to_cos=flip_sin_to_cos, freq_shift=freq_shift)
+        time_embed_dim = block_out_channels[0] * 4
+        self.time_embedding = TimestepEmbedding(timestep_embedding_dim, time_embed_dim)
+
+        # Input
+        self.conv_in = nn.Conv2d(in_channels, block_out_channels[0], kernel_size=3, padding=1)
+
+        # Down blocks
+        self.down_blocks = nn.ModuleList()
+        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
+
+            if "CrossAttn" in down_block_type:
+                down_block = CrossAttnDownBlock2D(
+                    in_channels=input_channel,
+                    out_channels=output_channel,
+                    temb_channels=time_embed_dim,
+                    dropout=dropout,
+                    num_layers=layers_per_block,
+                    transformer_layers_per_block=1,
+                    resnet_eps=norm_eps,
+                    resnet_time_scale_shift=resnet_time_scale_shift,
+                    resnet_act_fn=act_fn,
+                    resnet_groups=norm_num_groups,
+                    cross_attention_dim=cross_attention_dim,
+                    attention_head_dim=attention_head_dim,
+                    downsample=not is_final_block,
+                )
+            else:
+                down_block = DownBlock2D(
+                    in_channels=input_channel,
+                    out_channels=output_channel,
+                    temb_channels=time_embed_dim,
+                    dropout=dropout,
+                    num_layers=layers_per_block,
+                    resnet_eps=norm_eps,
+                    resnet_time_scale_shift=resnet_time_scale_shift,
+                    resnet_act_fn=act_fn,
+                    resnet_groups=norm_num_groups,
+                    downsample=not is_final_block,
+                )
+            self.down_blocks.append(down_block)
+
+        # Mid block
+        self.mid_block = UNetMidBlock2DCrossAttn(
+            in_channels=block_out_channels[-1],
+            temb_channels=time_embed_dim,
+            dropout=dropout,
+            num_layers=1,
+            transformer_layers_per_block=1,
+            resnet_eps=norm_eps,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            resnet_act_fn=act_fn,
+            resnet_groups=norm_num_groups,
+            cross_attention_dim=cross_attention_dim,
+            attention_head_dim=attention_head_dim,
+        )
+
+        # Up blocks
+        self.up_blocks = nn.ModuleList()
+        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
+
+            # in_channels for up blocks: diffusers uses reversed_block_out_channels[min(i+1, len-1)]
+            input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
+
+            if "CrossAttn" in up_block_type:
+                up_block = CrossAttnUpBlock2D(
+                    in_channels=input_channel,
+                    out_channels=output_channel,
+                    prev_output_channel=prev_output_channel,
+                    temb_channels=time_embed_dim,
+                    dropout=dropout,
+                    num_layers=layers_per_block + 1,
+                    transformer_layers_per_block=1,
+                    resnet_eps=norm_eps,
+                    resnet_time_scale_shift=resnet_time_scale_shift,
+                    resnet_act_fn=act_fn,
+                    resnet_groups=norm_num_groups,
+                    cross_attention_dim=cross_attention_dim,
+                    attention_head_dim=attention_head_dim,
+                    upsample=not is_final_block,
+                )
+            else:
+                up_block = UpBlock2D(
+                    in_channels=input_channel,
+                    out_channels=output_channel,
+                    prev_output_channel=prev_output_channel,
+                    temb_channels=time_embed_dim,
+                    dropout=dropout,
+                    num_layers=layers_per_block + 1,
+                    resnet_eps=norm_eps,
+                    resnet_time_scale_shift=resnet_time_scale_shift,
+                    resnet_act_fn=act_fn,
+                    resnet_groups=norm_num_groups,
+                    upsample=not is_final_block,
+                )
+            self.up_blocks.append(up_block)
+
+        # Output
+        self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps)
+        self.conv_act = nn.SiLU()
+        self.conv_out = nn.Conv2d(block_out_channels[0], out_channels, kernel_size=3, padding=1)
+
+    def forward(self, sample, timestep, encoder_hidden_states, cross_attention_kwargs=None, timestep_cond=None, added_cond_kwargs=None, return_dict=True):
+        # 1. Time embedding
+        timesteps = timestep
+        if not torch.is_tensor(timesteps):
+            timesteps = torch.tensor([timesteps], dtype=torch.long, device=sample.device)
+        elif torch.is_tensor(timesteps) and len(timesteps.shape) == 0:
+            timesteps = timesteps[None].to(sample.device)
+
+        t_emb = self.time_proj(timesteps)
+        t_emb = t_emb.to(dtype=sample.dtype)
+        emb = self.time_embedding(t_emb)
+
+        # 2. Pre-process
+        sample = self.conv_in(sample)
+
+        # 3. Down
+        down_block_res_samples = (sample,)
+        for down_block in self.down_blocks:
+            sample, res_samples = down_block(
+                hidden_states=sample,
+                temb=emb,
+                encoder_hidden_states=encoder_hidden_states,
+            )
+            down_block_res_samples += res_samples
+
+        # 4. Mid
+        sample = self.mid_block(sample, emb, encoder_hidden_states=encoder_hidden_states)
+
+        # 5. Up
+        for up_block in self.up_blocks:
+            res_samples = down_block_res_samples[-len(up_block.resnets):]
+            down_block_res_samples = down_block_res_samples[:-len(up_block.resnets)]
+
+            upsample_size = down_block_res_samples[-1].shape[2:] if down_block_res_samples else None
+            sample = up_block(
+                hidden_states=sample,
+                temb=emb,
+                encoder_hidden_states=encoder_hidden_states,
+                res_hidden_states_tuple=res_samples,
+                upsample_size=upsample_size,
+            )
+
+        # 6. Post-process
+        sample = self.conv_norm_out(sample)
+        sample = self.conv_act(sample)
+        sample = self.conv_out(sample)
+
+        if not return_dict:
+            return (sample,)
+        return sample