DiffSynth-Studio/diffsynth/models/qwen_image_vae.py

import torch
from typing import List, Optional, Tuple, Union
from torch import nn


CACHE_T = 2

class QwenImageCausalConv3d(torch.nn.Conv3d):
    r"""
    A custom 3D causal convolution layer with feature caching support.

    This layer extends the standard Conv3D layer by ensuring causality in the time dimension and handling feature
    caching for efficient inference.

    Args:
        in_channels (int): Number of channels in the input image
        out_channels (int): Number of channels produced by the convolution
        kernel_size (int or tuple): Size of the convolving kernel
        stride (int or tuple, optional): Stride of the convolution. Default: 1
        padding (int or tuple, optional): Zero-padding added to all three sides of the input. Default: 0
    """

    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        kernel_size: Union[int, Tuple[int, int, int]],
        stride: Union[int, Tuple[int, int, int]] = 1,
        padding: Union[int, Tuple[int, int, int]] = 0,
    ) -> None:
        super().__init__(
            in_channels=in_channels,
            out_channels=out_channels,
            kernel_size=kernel_size,
            stride=stride,
            padding=padding,
        )

        # Set up causal padding
        self._padding = (self.padding[2], self.padding[2], self.padding[1], self.padding[1], 2 * self.padding[0], 0)
        self.padding = (0, 0, 0)

    def forward(self, x, cache_x=None):
        padding = list(self._padding)
        if cache_x is not None and self._padding[4] > 0:
            cache_x = cache_x.to(x.device)
            x = torch.cat([cache_x, x], dim=2)
            padding[4] -= cache_x.shape[2]
        x = torch.nn.functional.pad(x, padding)
        return super().forward(x)



class QwenImageRMS_norm(nn.Module):
    r"""
    A custom RMS normalization layer.

    Args:
        dim (int): The number of dimensions to normalize over.
        channel_first (bool, optional): Whether the input tensor has channels as the first dimension.
            Default is True.
        images (bool, optional): Whether the input represents image data. Default is True.
        bias (bool, optional): Whether to include a learnable bias term. Default is False.
    """

    def __init__(self, dim: int, channel_first: bool = True, images: bool = True, bias: bool = False) -> None:
        super().__init__()
        broadcastable_dims = (1, 1, 1) if not images else (1, 1)
        shape = (dim, *broadcastable_dims) if channel_first else (dim,)

        self.channel_first = channel_first
        self.scale = dim**0.5
        self.gamma = nn.Parameter(torch.ones(shape))
        self.bias = nn.Parameter(torch.zeros(shape)) if bias else 0.0

    def forward(self, x):
        return torch.nn.functional.normalize(x, dim=(1 if self.channel_first else -1)) * self.scale * self.gamma + self.bias



class QwenImageResidualBlock(nn.Module):
    r"""
    A custom residual block module.

    Args:
        in_dim (int): Number of input channels.
        out_dim (int): Number of output channels.
        dropout (float, optional): Dropout rate for the dropout layer. Default is 0.0.
        non_linearity (str, optional): Type of non-linearity to use. Default is "silu".
    """

    def __init__(
        self,
        in_dim: int,
        out_dim: int,
        dropout: float = 0.0,
        non_linearity: str = "silu",
    ) -> None:
        super().__init__()
        self.in_dim = in_dim
        self.out_dim = out_dim
        self.nonlinearity = torch.nn.SiLU()

        # layers
        self.norm1 = QwenImageRMS_norm(in_dim, images=False)
        self.conv1 = QwenImageCausalConv3d(in_dim, out_dim, 3, padding=1)
        self.norm2 = QwenImageRMS_norm(out_dim, images=False)
        self.dropout = nn.Dropout(dropout)
        self.conv2 = QwenImageCausalConv3d(out_dim, out_dim, 3, padding=1)
        self.conv_shortcut = QwenImageCausalConv3d(in_dim, out_dim, 1) if in_dim != out_dim else nn.Identity()

    def forward(self, x, feat_cache=None, feat_idx=[0]):
        # Apply shortcut connection
        h = self.conv_shortcut(x)

        # First normalization and activation
        x = self.norm1(x)
        x = self.nonlinearity(x)

        if feat_cache is not None:
            idx = feat_idx[0]
            cache_x = x[:, :, -CACHE_T:, :, :].clone()
            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
                cache_x = torch.cat([feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x], dim=2)

            x = self.conv1(x, feat_cache[idx])
            feat_cache[idx] = cache_x
            feat_idx[0] += 1
        else:
            x = self.conv1(x)

        # Second normalization and activation
        x = self.norm2(x)
        x = self.nonlinearity(x)

        # Dropout
        x = self.dropout(x)

        if feat_cache is not None:
            idx = feat_idx[0]
            cache_x = x[:, :, -CACHE_T:, :, :].clone()
            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
                cache_x = torch.cat([feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x], dim=2)

            x = self.conv2(x, feat_cache[idx])
            feat_cache[idx] = cache_x
            feat_idx[0] += 1
        else:
            x = self.conv2(x)

        # Add residual connection
        return x + h



class QwenImageAttentionBlock(nn.Module):
    r"""
    Causal self-attention with a single head.

    Args:
        dim (int): The number of channels in the input tensor.
    """

    def __init__(self, dim):
        super().__init__()
        self.dim = dim

        # layers
        self.norm = QwenImageRMS_norm(dim)
        self.to_qkv = nn.Conv2d(dim, dim * 3, 1)
        self.proj = nn.Conv2d(dim, dim, 1)

    def forward(self, x):
        identity = x
        batch_size, channels, time, height, width = x.size()

        x = x.permute(0, 2, 1, 3, 4).reshape(batch_size * time, channels, height, width)
        x = self.norm(x)

        # compute query, key, value
        qkv = self.to_qkv(x)
        qkv = qkv.reshape(batch_size * time, 1, channels * 3, -1)
        qkv = qkv.permute(0, 1, 3, 2).contiguous()
        q, k, v = qkv.chunk(3, dim=-1)

        # apply attention
        x = torch.nn.functional.scaled_dot_product_attention(q, k, v)

        x = x.squeeze(1).permute(0, 2, 1).reshape(batch_size * time, channels, height, width)

        # output projection
        x = self.proj(x)

        # Reshape back: [(b*t), c, h, w] -> [b, c, t, h, w]
        x = x.view(batch_size, time, channels, height, width)
        x = x.permute(0, 2, 1, 3, 4)

        return x + identity



class QwenImageUpsample(nn.Upsample):
    r"""
    Perform upsampling while ensuring the output tensor has the same data type as the input.

    Args:
        x (torch.Tensor): Input tensor to be upsampled.

    Returns:
        torch.Tensor: Upsampled tensor with the same data type as the input.
    """

    def forward(self, x):
        return super().forward(x.float()).type_as(x)



class QwenImageResample(nn.Module):
    r"""
    A custom resampling module for 2D and 3D data.

    Args:
        dim (int): The number of input/output channels.
        mode (str): The resampling mode. Must be one of:
            - 'none': No resampling (identity operation).
            - 'upsample2d': 2D upsampling with nearest-exact interpolation and convolution.
            - 'upsample3d': 3D upsampling with nearest-exact interpolation, convolution, and causal 3D convolution.
            - 'downsample2d': 2D downsampling with zero-padding and convolution.
            - 'downsample3d': 3D downsampling with zero-padding, convolution, and causal 3D convolution.
    """

    def __init__(self, dim: int, mode: str) -> None:
        super().__init__()
        self.dim = dim
        self.mode = mode

        # layers
        if mode == "upsample2d":
            self.resample = nn.Sequential(
                QwenImageUpsample(scale_factor=(2.0, 2.0), mode="nearest-exact"), nn.Conv2d(dim, dim // 2, 3, padding=1)
            )
        elif mode == "upsample3d":
            self.resample = nn.Sequential(
                QwenImageUpsample(scale_factor=(2.0, 2.0), mode="nearest-exact"), nn.Conv2d(dim, dim // 2, 3, padding=1)
            )
            self.time_conv = QwenImageCausalConv3d(dim, dim * 2, (3, 1, 1), padding=(1, 0, 0))

        elif mode == "downsample2d":
            self.resample = nn.Sequential(nn.ZeroPad2d((0, 1, 0, 1)), nn.Conv2d(dim, dim, 3, stride=(2, 2)))
        elif mode == "downsample3d":
            self.resample = nn.Sequential(nn.ZeroPad2d((0, 1, 0, 1)), nn.Conv2d(dim, dim, 3, stride=(2, 2)))
            self.time_conv = QwenImageCausalConv3d(dim, dim, (3, 1, 1), stride=(2, 1, 1), padding=(0, 0, 0))

        else:
            self.resample = nn.Identity()

    def forward(self, x, feat_cache=None, feat_idx=[0]):
        b, c, t, h, w = x.size()
        if self.mode == "upsample3d":
            if feat_cache is not None:
                idx = feat_idx[0]
                if feat_cache[idx] is None:
                    feat_cache[idx] = "Rep"
                    feat_idx[0] += 1
                else:
                    cache_x = x[:, :, -CACHE_T:, :, :].clone()
                    if cache_x.shape[2] < 2 and feat_cache[idx] is not None and feat_cache[idx] != "Rep":
                        # cache last frame of last two chunk
                        cache_x = torch.cat(
                            [feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x], dim=2
                        )
                    if cache_x.shape[2] < 2 and feat_cache[idx] is not None and feat_cache[idx] == "Rep":
                        cache_x = torch.cat([torch.zeros_like(cache_x).to(cache_x.device), cache_x], dim=2)
                    if feat_cache[idx] == "Rep":
                        x = self.time_conv(x)
                    else:
                        x = self.time_conv(x, feat_cache[idx])
                    feat_cache[idx] = cache_x
                    feat_idx[0] += 1

                    x = x.reshape(b, 2, c, t, h, w)
                    x = torch.stack((x[:, 0, :, :, :, :], x[:, 1, :, :, :, :]), 3)
                    x = x.reshape(b, c, t * 2, h, w)
        t = x.shape[2]
        x = x.permute(0, 2, 1, 3, 4).reshape(b * t, c, h, w)
        x = self.resample(x)
        x = x.view(b, t, x.size(1), x.size(2), x.size(3)).permute(0, 2, 1, 3, 4)

        if self.mode == "downsample3d":
            if feat_cache is not None:
                idx = feat_idx[0]
                if feat_cache[idx] is None:
                    feat_cache[idx] = x.clone()
                    feat_idx[0] += 1
                else:
                    cache_x = x[:, :, -1:, :, :].clone()
                    x = self.time_conv(torch.cat([feat_cache[idx][:, :, -1:, :, :], x], 2))
                    feat_cache[idx] = cache_x
                    feat_idx[0] += 1
        return x



class QwenImageMidBlock(nn.Module):
    """
    Middle block for WanVAE encoder and decoder.

    Args:
        dim (int): Number of input/output channels.
        dropout (float): Dropout rate.
        non_linearity (str): Type of non-linearity to use.
    """

    def __init__(self, dim: int, dropout: float = 0.0, non_linearity: str = "silu", num_layers: int = 1):
        super().__init__()
        self.dim = dim

        # Create the components
        resnets = [QwenImageResidualBlock(dim, dim, dropout, non_linearity)]
        attentions = []
        for _ in range(num_layers):
            attentions.append(QwenImageAttentionBlock(dim))
            resnets.append(QwenImageResidualBlock(dim, dim, dropout, non_linearity))
        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

        self.gradient_checkpointing = False

    def forward(self, x, feat_cache=None, feat_idx=[0]):
        # First residual block
        x = self.resnets[0](x, feat_cache, feat_idx)

        # Process through attention and residual blocks
        for attn, resnet in zip(self.attentions, self.resnets[1:]):
            if attn is not None:
                x = attn(x)

            x = resnet(x, feat_cache, feat_idx)

        return x



class QwenImageEncoder3d(nn.Module):
    r"""
    A 3D encoder module.

    Args:
        dim (int): The base number of channels in the first layer.
        z_dim (int): The dimensionality of the latent space.
        dim_mult (list of int): Multipliers for the number of channels in each block.
        num_res_blocks (int): Number of residual blocks in each block.
        attn_scales (list of float): Scales at which to apply attention mechanisms.
        temperal_downsample (list of bool): Whether to downsample temporally in each block.
        dropout (float): Dropout rate for the dropout layers.
        non_linearity (str): Type of non-linearity to use.
    """

    def __init__(
        self,
        dim=128,
        z_dim=4,
        dim_mult=[1, 2, 4, 4],
        num_res_blocks=2,
        attn_scales=[],
        temperal_downsample=[True, True, False],
        dropout=0.0,
        non_linearity: str = "silu",
        image_channels=3
    ):
        super().__init__()
        self.dim = dim
        self.z_dim = z_dim
        self.dim_mult = dim_mult
        self.num_res_blocks = num_res_blocks
        self.attn_scales = attn_scales
        self.temperal_downsample = temperal_downsample
        self.nonlinearity = torch.nn.SiLU()

        # dimensions
        dims = [dim * u for u in [1] + dim_mult]
        scale = 1.0

        # init block
        self.conv_in = QwenImageCausalConv3d(image_channels, dims[0], 3, padding=1)

        # downsample blocks
        self.down_blocks = torch.nn.ModuleList([])
        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
            # residual (+attention) blocks
            for _ in range(num_res_blocks):
                self.down_blocks.append(QwenImageResidualBlock(in_dim, out_dim, dropout))
                if scale in attn_scales:
                    self.down_blocks.append(QwenImageAttentionBlock(out_dim))
                in_dim = out_dim

            # downsample block
            if i != len(dim_mult) - 1:
                mode = "downsample3d" if temperal_downsample[i] else "downsample2d"
                self.down_blocks.append(QwenImageResample(out_dim, mode=mode))
                scale /= 2.0

        # middle blocks
        self.mid_block = QwenImageMidBlock(out_dim, dropout, non_linearity, num_layers=1)

        # output blocks
        self.norm_out = QwenImageRMS_norm(out_dim, images=False)
        self.conv_out = QwenImageCausalConv3d(out_dim, z_dim, 3, padding=1)

        self.gradient_checkpointing = False

    def forward(self, x, feat_cache=None, feat_idx=[0]):
        if feat_cache is not None:
            idx = feat_idx[0]
            cache_x = x[:, :, -CACHE_T:, :, :].clone()
            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
                # cache last frame of last two chunk
                cache_x = torch.cat([feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x], dim=2)
            x = self.conv_in(x, feat_cache[idx])
            feat_cache[idx] = cache_x
            feat_idx[0] += 1
        else:
            x = self.conv_in(x)

        ## downsamples
        for layer in self.down_blocks:
            if feat_cache is not None:
                x = layer(x, feat_cache, feat_idx)
            else:
                x = layer(x)

        ## middle
        x = self.mid_block(x, feat_cache, feat_idx)

        ## head
        x = self.norm_out(x)
        x = self.nonlinearity(x)
        if feat_cache is not None:
            idx = feat_idx[0]
            cache_x = x[:, :, -CACHE_T:, :, :].clone()
            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
                # cache last frame of last two chunk
                cache_x = torch.cat([feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x], dim=2)
            x = self.conv_out(x, feat_cache[idx])
            feat_cache[idx] = cache_x
            feat_idx[0] += 1
        else:
            x = self.conv_out(x)
        return x



class QwenImageUpBlock(nn.Module):
    """
    A block that handles upsampling for the WanVAE decoder.

    Args:
        in_dim (int): Input dimension
        out_dim (int): Output dimension
        num_res_blocks (int): Number of residual blocks
        dropout (float): Dropout rate
        upsample_mode (str, optional): Mode for upsampling ('upsample2d' or 'upsample3d')
        non_linearity (str): Type of non-linearity to use
    """

    def __init__(
        self,
        in_dim: int,
        out_dim: int,
        num_res_blocks: int,
        dropout: float = 0.0,
        upsample_mode: Optional[str] = None,
        non_linearity: str = "silu",
    ):
        super().__init__()
        self.in_dim = in_dim
        self.out_dim = out_dim

        # Create layers list
        resnets = []
        # Add residual blocks and attention if needed
        current_dim = in_dim
        for _ in range(num_res_blocks + 1):
            resnets.append(QwenImageResidualBlock(current_dim, out_dim, dropout, non_linearity))
            current_dim = out_dim

        self.resnets = nn.ModuleList(resnets)

        # Add upsampling layer if needed
        self.upsamplers = None
        if upsample_mode is not None:
            self.upsamplers = nn.ModuleList([QwenImageResample(out_dim, mode=upsample_mode)])

        self.gradient_checkpointing = False

    def forward(self, x, feat_cache=None, feat_idx=[0]):
        """
        Forward pass through the upsampling block.

        Args:
            x (torch.Tensor): Input tensor
            feat_cache (list, optional): Feature cache for causal convolutions
            feat_idx (list, optional): Feature index for cache management

        Returns:
            torch.Tensor: Output tensor
        """
        for resnet in self.resnets:
            if feat_cache is not None:
                x = resnet(x, feat_cache, feat_idx)
            else:
                x = resnet(x)

        if self.upsamplers is not None:
            if feat_cache is not None:
                x = self.upsamplers[0](x, feat_cache, feat_idx)
            else:
                x = self.upsamplers[0](x)
        return x



class QwenImageDecoder3d(nn.Module):
    r"""
    A 3D decoder module.

    Args:
        dim (int): The base number of channels in the first layer.
        z_dim (int): The dimensionality of the latent space.
        dim_mult (list of int): Multipliers for the number of channels in each block.
        num_res_blocks (int): Number of residual blocks in each block.
        attn_scales (list of float): Scales at which to apply attention mechanisms.
        temperal_upsample (list of bool): Whether to upsample temporally in each block.
        dropout (float): Dropout rate for the dropout layers.
        non_linearity (str): Type of non-linearity to use.
    """

    def __init__(
        self,
        dim=128,
        z_dim=4,
        dim_mult=[1, 2, 4, 4],
        num_res_blocks=2,
        attn_scales=[],
        temperal_upsample=[False, True, True],
        dropout=0.0,
        non_linearity: str = "silu",
        image_channels=3,
    ):
        super().__init__()
        self.dim = dim
        self.z_dim = z_dim
        self.dim_mult = dim_mult
        self.num_res_blocks = num_res_blocks
        self.attn_scales = attn_scales
        self.temperal_upsample = temperal_upsample

        self.nonlinearity = torch.nn.SiLU()

        # dimensions
        dims = [dim * u for u in [dim_mult[-1]] + dim_mult[::-1]]
        scale = 1.0 / 2 ** (len(dim_mult) - 2)

        # init block
        self.conv_in = QwenImageCausalConv3d(z_dim, dims[0], 3, padding=1)

        # middle blocks
        self.mid_block = QwenImageMidBlock(dims[0], dropout, non_linearity, num_layers=1)

        # upsample blocks
        self.up_blocks = nn.ModuleList([])
        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
            # residual (+attention) blocks
            if i > 0:
                in_dim = in_dim // 2

            # Determine if we need upsampling
            upsample_mode = None
            if i != len(dim_mult) - 1:
                upsample_mode = "upsample3d" if temperal_upsample[i] else "upsample2d"

            # Create and add the upsampling block
            up_block = QwenImageUpBlock(
                in_dim=in_dim,
                out_dim=out_dim,
                num_res_blocks=num_res_blocks,
                dropout=dropout,
                upsample_mode=upsample_mode,
                non_linearity=non_linearity,
            )
            self.up_blocks.append(up_block)

            # Update scale for next iteration
            if upsample_mode is not None:
                scale *= 2.0

        # output blocks
        self.norm_out = QwenImageRMS_norm(out_dim, images=False)
        self.conv_out = QwenImageCausalConv3d(out_dim, image_channels, 3, padding=1)

        self.gradient_checkpointing = False

    def forward(self, x, feat_cache=None, feat_idx=[0]):
        ## conv1
        if feat_cache is not None:
            idx = feat_idx[0]
            cache_x = x[:, :, -CACHE_T:, :, :].clone()
            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
                # cache last frame of last two chunk
                cache_x = torch.cat([feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x], dim=2)
            x = self.conv_in(x, feat_cache[idx])
            feat_cache[idx] = cache_x
            feat_idx[0] += 1
        else:
            x = self.conv_in(x)

        ## middle
        x = self.mid_block(x, feat_cache, feat_idx)

        ## upsamples
        for up_block in self.up_blocks:
            x = up_block(x, feat_cache, feat_idx)

        ## head
        x = self.norm_out(x)
        x = self.nonlinearity(x)
        if feat_cache is not None:
            idx = feat_idx[0]
            cache_x = x[:, :, -CACHE_T:, :, :].clone()
            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
                # cache last frame of last two chunk
                cache_x = torch.cat([feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x], dim=2)
            x = self.conv_out(x, feat_cache[idx])
            feat_cache[idx] = cache_x
            feat_idx[0] += 1
        else:
            x = self.conv_out(x)
        return x



class QwenImageVAE(torch.nn.Module):
    def __init__(
        self,
        base_dim: int = 96,
        z_dim: int = 16,
        dim_mult: Tuple[int] = [1, 2, 4, 4],
        num_res_blocks: int = 2,
        attn_scales: List[float] = [],
        temperal_downsample: List[bool] = [False, True, True],
        dropout: float = 0.0,
        image_channels: int = 3,
    ) -> None:
        super().__init__()

        self.z_dim = z_dim
        self.temperal_downsample = temperal_downsample
        self.temperal_upsample = temperal_downsample[::-1]

        self.encoder = QwenImageEncoder3d(
            base_dim, z_dim * 2, dim_mult, num_res_blocks, attn_scales, self.temperal_downsample, dropout, image_channels=image_channels,
        )
        self.quant_conv = QwenImageCausalConv3d(z_dim * 2, z_dim * 2, 1)
        self.post_quant_conv = QwenImageCausalConv3d(z_dim, z_dim, 1)

        self.decoder = QwenImageDecoder3d(
            base_dim, z_dim, dim_mult, num_res_blocks, attn_scales, self.temperal_upsample, dropout, image_channels=image_channels,
        )

        mean = [
            -0.7571,
            -0.7089,
            -0.9113,
            0.1075,
            -0.1745,
            0.9653,
            -0.1517,
            1.5508,
            0.4134,
            -0.0715,
            0.5517,
            -0.3632,
            -0.1922,
            -0.9497,
            0.2503,
            -0.2921,
        ]
        std = [
            2.8184,
            1.4541,
            2.3275,
            2.6558,
            1.2196,
            1.7708,
            2.6052,
            2.0743,
            3.2687,
            2.1526,
            2.8652,
            1.5579,
            1.6382,
            1.1253,
            2.8251,
            1.9160,
        ]
        self.mean = torch.tensor(mean).view(1, 16, 1, 1, 1)
        self.std = 1 / torch.tensor(std).view(1, 16, 1, 1, 1)

    def encode(self, x, **kwargs):
        x = x.unsqueeze(2)
        x = self.encoder(x)
        x = self.quant_conv(x)
        x = x[:, :16]
        mean, std = self.mean.to(dtype=x.dtype, device=x.device), self.std.to(dtype=x.dtype, device=x.device)
        x = (x - mean) * std
        x = x.squeeze(2)
        return x
    
    def decode(self, x, **kwargs):
        x = x.unsqueeze(2)
        mean, std = self.mean.to(dtype=x.dtype, device=x.device), self.std.to(dtype=x.dtype, device=x.device)
        x = x / std + mean
        x = self.post_quant_conv(x)
        x = self.decoder(x)
        x = x.squeeze(2)
        return x