Track bundled vendor runtime sources

vendor/ComfyUI/comfy/ldm/lightricks/vae/audio_vae.py

@@ -0,0 +1,239 @@
+import json
+from dataclasses import dataclass
+import math
+import torch
+import torchaudio
+
+from comfy.ldm.mmaudio.vae.distributions import DiagonalGaussianDistribution
+from comfy.ldm.lightricks.symmetric_patchifier import AudioPatchifier
+from comfy.ldm.lightricks.vae.causal_audio_autoencoder import (
+    CausalityAxis,
+    CausalAudioAutoencoder,
+)
+from comfy.ldm.lightricks.vocoders.vocoder import Vocoder, VocoderWithBWE
+
+LATENT_DOWNSAMPLE_FACTOR = 4
+
+
+@dataclass(frozen=True)
+class AudioVAEComponentConfig:
+    """Container for model component configuration extracted from metadata."""
+
+    autoencoder: dict
+    vocoder: dict
+
+    @classmethod
+    def from_metadata(cls, metadata: dict) -> "AudioVAEComponentConfig":
+        assert metadata is not None and "config" in metadata, "Metadata is required for audio VAE"
+
+        raw_config = metadata["config"]
+        if isinstance(raw_config, str):
+            parsed_config = json.loads(raw_config)
+        else:
+            parsed_config = raw_config
+
+        audio_config = parsed_config.get("audio_vae")
+        vocoder_config = parsed_config.get("vocoder")
+
+        assert audio_config is not None, "Audio VAE config is required for audio VAE"
+        assert vocoder_config is not None, "Vocoder config is required for audio VAE"
+
+        return cls(autoencoder=audio_config, vocoder=vocoder_config)
+
+class AudioLatentNormalizer:
+    """Applies per-channel statistics in patch space and restores original layout."""
+
+    def __init__(self, patchfier: AudioPatchifier, statistics_processor: torch.nn.Module):
+        self.patchifier = patchfier
+        self.statistics = statistics_processor
+
+    def normalize(self, latents: torch.Tensor) -> torch.Tensor:
+        channels = latents.shape[1]
+        freq = latents.shape[3]
+        patched, _ = self.patchifier.patchify(latents)
+        normalized = self.statistics.normalize(patched)
+        return self.patchifier.unpatchify(normalized, channels=channels, freq=freq)
+
+    def denormalize(self, latents: torch.Tensor) -> torch.Tensor:
+        channels = latents.shape[1]
+        freq = latents.shape[3]
+        patched, _ = self.patchifier.patchify(latents)
+        denormalized = self.statistics.un_normalize(patched)
+        return self.patchifier.unpatchify(denormalized, channels=channels, freq=freq)
+
+
+class AudioPreprocessor:
+    """Prepares raw waveforms for the autoencoder by matching training conditions."""
+
+    def __init__(self, target_sample_rate: int, mel_bins: int, mel_hop_length: int, n_fft: int):
+        self.target_sample_rate = target_sample_rate
+        self.mel_bins = mel_bins
+        self.mel_hop_length = mel_hop_length
+        self.n_fft = n_fft
+
+    def resample(self, waveform: torch.Tensor, source_rate: int) -> torch.Tensor:
+        if source_rate == self.target_sample_rate:
+            return waveform
+        return torchaudio.functional.resample(waveform, source_rate, self.target_sample_rate)
+
+    def waveform_to_mel(
+        self, waveform: torch.Tensor, waveform_sample_rate: int, device
+    ) -> torch.Tensor:
+        waveform = self.resample(waveform, waveform_sample_rate)
+
+        mel_transform = torchaudio.transforms.MelSpectrogram(
+            sample_rate=self.target_sample_rate,
+            n_fft=self.n_fft,
+            win_length=self.n_fft,
+            hop_length=self.mel_hop_length,
+            f_min=0.0,
+            f_max=self.target_sample_rate / 2.0,
+            n_mels=self.mel_bins,
+            window_fn=torch.hann_window,
+            center=True,
+            pad_mode="reflect",
+            power=1.0,
+            mel_scale="slaney",
+            norm="slaney",
+        ).to(device)
+
+        mel = mel_transform(waveform)
+        mel = torch.log(torch.clamp(mel, min=1e-5))
+        return mel.permute(0, 1, 3, 2).contiguous()
+
+
+class AudioVAE(torch.nn.Module):
+    """High-level Audio VAE wrapper exposing encode and decode entry points."""
+
+    def __init__(self, metadata: dict):
+        super().__init__()
+
+        component_config = AudioVAEComponentConfig.from_metadata(metadata)
+
+        self.autoencoder = CausalAudioAutoencoder(config=component_config.autoencoder)
+        if "bwe" in component_config.vocoder:
+            self.vocoder = VocoderWithBWE(config=component_config.vocoder)
+        else:
+            self.vocoder = Vocoder(config=component_config.vocoder)
+
+        autoencoder_config = self.autoencoder.get_config()
+        self.normalizer = AudioLatentNormalizer(
+            AudioPatchifier(
+                patch_size=1,
+                audio_latent_downsample_factor=LATENT_DOWNSAMPLE_FACTOR,
+                sample_rate=autoencoder_config["sampling_rate"],
+                hop_length=autoencoder_config["mel_hop_length"],
+                is_causal=autoencoder_config["is_causal"],
+            ),
+            self.autoencoder.per_channel_statistics,
+        )
+
+        self.preprocessor = AudioPreprocessor(
+            target_sample_rate=autoencoder_config["sampling_rate"],
+            mel_bins=autoencoder_config["mel_bins"],
+            mel_hop_length=autoencoder_config["mel_hop_length"],
+            n_fft=autoencoder_config["n_fft"],
+        )
+
+    def encode(self, audio, sample_rate=44100) -> torch.Tensor:
+        """Encode a waveform dictionary into normalized latent tensors."""
+
+        waveform = audio
+        waveform_sample_rate = sample_rate
+        input_device = waveform.device
+        expected_channels = self.autoencoder.encoder.in_channels
+        if waveform.shape[1] != expected_channels:
+            if waveform.shape[1] == 1:
+                waveform = waveform.expand(-1, expected_channels, *waveform.shape[2:])
+            else:
+                raise ValueError(
+                    f"Input audio must have {expected_channels} channels, got {waveform.shape[1]}"
+                )
+
+        mel_spec = self.preprocessor.waveform_to_mel(
+            waveform, waveform_sample_rate, device=waveform.device
+        )
+
+        latents = self.autoencoder.encode(mel_spec)
+        posterior = DiagonalGaussianDistribution(latents)
+        latent_mode = posterior.mode()
+
+        normalized = self.normalizer.normalize(latent_mode)
+        return normalized.to(input_device)
+
+    def decode(self, latents: torch.Tensor) -> torch.Tensor:
+        """Decode normalized latent tensors into an audio waveform."""
+        original_shape = latents.shape
+
+        latents = self.normalizer.denormalize(latents)
+
+        target_shape = self.target_shape_from_latents(original_shape)
+        mel_spec = self.autoencoder.decode(latents, target_shape=target_shape)
+
+        waveform = self.run_vocoder(mel_spec)
+        return waveform
+
+    def target_shape_from_latents(self, latents_shape):
+        batch, _, time, _ = latents_shape
+        target_length = time * LATENT_DOWNSAMPLE_FACTOR
+        if self.autoencoder.causality_axis != CausalityAxis.NONE:
+            target_length -= LATENT_DOWNSAMPLE_FACTOR - 1
+        return (
+            batch,
+            self.autoencoder.decoder.out_ch,
+            target_length,
+            self.autoencoder.mel_bins,
+        )
+
+    def num_of_latents_from_frames(self, frames_number: int, frame_rate: int) -> int:
+        return math.ceil((float(frames_number) / frame_rate) * self.latents_per_second)
+
+    def run_vocoder(self, mel_spec: torch.Tensor) -> torch.Tensor:
+        audio_channels = self.autoencoder.decoder.out_ch
+        vocoder_input = mel_spec.transpose(2, 3)
+
+        if audio_channels == 1:
+            vocoder_input = vocoder_input.squeeze(1)
+        elif audio_channels != 2:
+            raise ValueError(f"Unsupported audio_channels: {audio_channels}")
+
+        return self.vocoder(vocoder_input)
+
+    @property
+    def sample_rate(self) -> int:
+        return int(self.autoencoder.sampling_rate)
+
+    @property
+    def mel_hop_length(self) -> int:
+        return int(self.autoencoder.mel_hop_length)
+
+    @property
+    def mel_bins(self) -> int:
+        return int(self.autoencoder.mel_bins)
+
+    @property
+    def latent_channels(self) -> int:
+        return int(self.autoencoder.decoder.z_channels)
+
+    @property
+    def latent_frequency_bins(self) -> int:
+        return int(self.mel_bins // LATENT_DOWNSAMPLE_FACTOR)
+
+    @property
+    def latents_per_second(self) -> float:
+        return self.sample_rate / self.mel_hop_length / LATENT_DOWNSAMPLE_FACTOR
+
+    @property
+    def output_sample_rate(self) -> int:
+        output_rate = getattr(self.vocoder, "output_sample_rate", None)
+        if output_rate is not None:
+            return int(output_rate)
+        upsample_factor = getattr(self.vocoder, "upsample_factor", None)
+        if upsample_factor is None:
+            raise AttributeError(
+                "Vocoder is missing upsample_factor; cannot infer output sample rate"
+            )
+        return int(self.sample_rate * upsample_factor / self.mel_hop_length)
+
+    def memory_required(self, input_shape):
+        return self.device_manager.patcher.model_size()

vendor/ComfyUI/comfy/ldm/lightricks/vae/causal_audio_autoencoder.py

@@ -0,0 +1,899 @@
+import torch
+from torch import nn
+from torch.nn import functional as F
+from typing import Optional
+from enum import Enum
+from .pixel_norm import PixelNorm
+import comfy.ops
+import logging
+
+ops = comfy.ops.disable_weight_init
+
+
+class StringConvertibleEnum(Enum):
+    """
+    Base enum class that provides string-to-enum conversion functionality.
+
+    This mixin adds a str_to_enum() class method that handles conversion from
+    strings, None, or existing enum instances with case-insensitive matching.
+    """
+
+    @classmethod
+    def str_to_enum(cls, value):
+        """
+        Convert a string, enum instance, or None to the appropriate enum member.
+
+        Args:
+            value: Can be an enum instance of this class, a string, or None
+
+        Returns:
+            Enum member of this class
+
+        Raises:
+            ValueError: If the value cannot be converted to a valid enum member
+        """
+        # Already an enum instance of this class
+        if isinstance(value, cls):
+            return value
+
+        # None maps to NONE member if it exists
+        if value is None:
+            if hasattr(cls, "NONE"):
+                return cls.NONE
+            raise ValueError(f"{cls.__name__} does not have a NONE member to map None to")
+
+        # String conversion (case-insensitive)
+        if isinstance(value, str):
+            value_lower = value.lower()
+
+            # Try to match against enum values
+            for member in cls:
+                # Handle members with None values
+                if member.value is None:
+                    if value_lower == "none":
+                        return member
+                # Handle members with string values
+                elif isinstance(member.value, str) and member.value.lower() == value_lower:
+                    return member
+
+            # Build helpful error message with valid values
+            valid_values = []
+            for member in cls:
+                if member.value is None:
+                    valid_values.append("none")
+                elif isinstance(member.value, str):
+                    valid_values.append(member.value)
+
+            raise ValueError(f"Invalid {cls.__name__} string: '{value}'. " f"Valid values are: {valid_values}")
+
+        raise ValueError(
+            f"Cannot convert type {type(value).__name__} to {cls.__name__} enum. "
+            f"Expected string, None, or {cls.__name__} instance."
+        )
+
+
+class AttentionType(StringConvertibleEnum):
+    """Enum for specifying the attention mechanism type."""
+
+    VANILLA = "vanilla"
+    LINEAR = "linear"
+    NONE = "none"
+
+
+class CausalityAxis(StringConvertibleEnum):
+    """Enum for specifying the causality axis in causal convolutions."""
+
+    NONE = None
+    WIDTH = "width"
+    HEIGHT = "height"
+    WIDTH_COMPATIBILITY = "width-compatibility"
+
+
+def Normalize(in_channels, *, num_groups=32, normtype="group"):
+    if normtype == "group":
+        return torch.nn.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True)
+    elif normtype == "pixel":
+        return PixelNorm(dim=1, eps=1e-6)
+    else:
+        raise ValueError(f"Invalid normalization type: {normtype}")
+
+
+class CausalConv2d(nn.Module):
+    """
+    A causal 2D convolution.
+
+    This layer ensures that the output at time `t` only depends on inputs
+    at time `t` and earlier. It achieves this by applying asymmetric padding
+    to the time dimension (width) before the convolution.
+    """
+
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        kernel_size,
+        stride=1,
+        dilation=1,
+        groups=1,
+        bias=True,
+        causality_axis: CausalityAxis = CausalityAxis.HEIGHT,
+    ):
+        super().__init__()
+
+        self.causality_axis = causality_axis
+
+        # Ensure kernel_size and dilation are tuples
+        kernel_size = nn.modules.utils._pair(kernel_size)
+        dilation = nn.modules.utils._pair(dilation)
+
+        # Calculate padding dimensions
+        pad_h = (kernel_size[0] - 1) * dilation[0]
+        pad_w = (kernel_size[1] - 1) * dilation[1]
+
+        # The padding tuple for F.pad is (pad_left, pad_right, pad_top, pad_bottom)
+        match self.causality_axis:
+            case CausalityAxis.NONE:
+                self.padding = (pad_w // 2, pad_w - pad_w // 2, pad_h // 2, pad_h - pad_h // 2)
+            case CausalityAxis.WIDTH | CausalityAxis.WIDTH_COMPATIBILITY:
+                self.padding = (pad_w, 0, pad_h // 2, pad_h - pad_h // 2)
+            case CausalityAxis.HEIGHT:
+                self.padding = (pad_w // 2, pad_w - pad_w // 2, pad_h, 0)
+            case _:
+                raise ValueError(f"Invalid causality_axis: {causality_axis}")
+
+        # The internal convolution layer uses no padding, as we handle it manually
+        self.conv = ops.Conv2d(
+            in_channels,
+            out_channels,
+            kernel_size,
+            stride=stride,
+            padding=0,
+            dilation=dilation,
+            groups=groups,
+            bias=bias,
+        )
+
+    def forward(self, x):
+        # Apply causal padding before convolution
+        x = F.pad(x, self.padding)
+        return self.conv(x)
+
+
+def make_conv2d(
+    in_channels,
+    out_channels,
+    kernel_size,
+    stride=1,
+    padding=None,
+    dilation=1,
+    groups=1,
+    bias=True,
+    causality_axis: Optional[CausalityAxis] = None,
+):
+    """
+    Create a 2D convolution layer that can be either causal or non-causal.
+
+    Args:
+        in_channels: Number of input channels
+        out_channels: Number of output channels
+        kernel_size: Size of the convolution kernel
+        stride: Convolution stride
+        padding: Padding (if None, will be calculated based on causal flag)
+        dilation: Dilation rate
+        groups: Number of groups for grouped convolution
+        bias: Whether to use bias
+        causality_axis: Dimension along which to apply causality.
+
+    Returns:
+        Either a regular Conv2d or CausalConv2d layer
+    """
+    if causality_axis is not None:
+        # For causal convolution, padding is handled internally by CausalConv2d
+        return CausalConv2d(in_channels, out_channels, kernel_size, stride, dilation, groups, bias, causality_axis)
+    else:
+        # For non-causal convolution, use symmetric padding if not specified
+        if padding is None:
+            if isinstance(kernel_size, int):
+                padding = kernel_size // 2
+            else:
+                padding = tuple(k // 2 for k in kernel_size)
+        return ops.Conv2d(
+            in_channels,
+            out_channels,
+            kernel_size,
+            stride,
+            padding,
+            dilation,
+            groups,
+            bias,
+        )
+
+
+class Upsample(nn.Module):
+    def __init__(self, in_channels, with_conv, causality_axis: CausalityAxis = CausalityAxis.HEIGHT):
+        super().__init__()
+        self.with_conv = with_conv
+        self.causality_axis = causality_axis
+        if self.with_conv:
+            self.conv = make_conv2d(in_channels, in_channels, kernel_size=3, stride=1, causality_axis=causality_axis)
+
+    def forward(self, x):
+        x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest")
+        if self.with_conv:
+            x = self.conv(x)
+            # Drop FIRST element in the causal axis to undo encoder's padding, while keeping the length 1 + 2 * n.
+            # For example, if the input is [0, 1, 2], after interpolation, the output is [0, 0, 1, 1, 2, 2].
+            # The causal convolution will pad the first element as [-, -, 0, 0, 1, 1, 2, 2],
+            # So the output elements rely on the following windows:
+            # 0: [-,-,0]
+            # 1: [-,0,0]
+            # 2: [0,0,1]
+            # 3: [0,1,1]
+            # 4: [1,1,2]
+            # 5: [1,2,2]
+            # Notice that the first and second elements in the output rely only on the first element in the input,
+            # while all other elements rely on two elements in the input.
+            # So we can drop the first element to undo the padding (rather than the last element).
+            # This is a no-op for non-causal convolutions.
+            match self.causality_axis:
+                case CausalityAxis.NONE:
+                    pass  # x remains unchanged
+                case CausalityAxis.HEIGHT:
+                    x = x[:, :, 1:, :]
+                case CausalityAxis.WIDTH:
+                    x = x[:, :, :, 1:]
+                case CausalityAxis.WIDTH_COMPATIBILITY:
+                    pass  # x remains unchanged
+                case _:
+                    raise ValueError(f"Invalid causality_axis: {self.causality_axis}")
+
+        return x
+
+
+class Downsample(nn.Module):
+    """
+    A downsampling layer that can use either a strided convolution
+    or average pooling. Supports standard and causal padding for the
+    convolutional mode.
+    """
+
+    def __init__(self, in_channels, with_conv, causality_axis: CausalityAxis = CausalityAxis.WIDTH):
+        super().__init__()
+        self.with_conv = with_conv
+        self.causality_axis = causality_axis
+
+        if self.causality_axis != CausalityAxis.NONE and not self.with_conv:
+            raise ValueError("causality is only supported when `with_conv=True`.")
+
+        if self.with_conv:
+            # Do time downsampling here
+            # no asymmetric padding in torch conv, must do it ourselves
+            self.conv = ops.Conv2d(in_channels, in_channels, kernel_size=3, stride=2, padding=0)
+
+    def forward(self, x):
+        if self.with_conv:
+            # (pad_left, pad_right, pad_top, pad_bottom)
+            match self.causality_axis:
+                case CausalityAxis.NONE:
+                    pad = (0, 1, 0, 1)
+                case CausalityAxis.WIDTH:
+                    pad = (2, 0, 0, 1)
+                case CausalityAxis.HEIGHT:
+                    pad = (0, 1, 2, 0)
+                case CausalityAxis.WIDTH_COMPATIBILITY:
+                    pad = (1, 0, 0, 1)
+                case _:
+                    raise ValueError(f"Invalid causality_axis: {self.causality_axis}")
+
+            x = torch.nn.functional.pad(x, pad, mode="constant", value=0)
+            x = self.conv(x)
+        else:
+            # This branch is only taken if with_conv=False, which implies causality_axis is NONE.
+            x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2)
+
+        return x
+
+
+class ResnetBlock(nn.Module):
+    def __init__(
+        self,
+        *,
+        in_channels,
+        out_channels=None,
+        conv_shortcut=False,
+        dropout,
+        temb_channels=512,
+        norm_type="group",
+        causality_axis: CausalityAxis = CausalityAxis.HEIGHT,
+    ):
+        super().__init__()
+        self.causality_axis = causality_axis
+
+        if self.causality_axis != CausalityAxis.NONE and norm_type == "group":
+            raise ValueError("Causal ResnetBlock with GroupNorm is not supported.")
+        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.norm1 = Normalize(in_channels, normtype=norm_type)
+        self.non_linearity = nn.SiLU()
+        self.conv1 = make_conv2d(in_channels, out_channels, kernel_size=3, stride=1, causality_axis=causality_axis)
+        if temb_channels > 0:
+            self.temb_proj = ops.Linear(temb_channels, out_channels)
+        self.norm2 = Normalize(out_channels, normtype=norm_type)
+        self.dropout = torch.nn.Dropout(dropout)
+        self.conv2 = make_conv2d(out_channels, out_channels, kernel_size=3, stride=1, causality_axis=causality_axis)
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                self.conv_shortcut = make_conv2d(
+                    in_channels, out_channels, kernel_size=3, stride=1, causality_axis=causality_axis
+                )
+            else:
+                self.nin_shortcut = make_conv2d(
+                    in_channels, out_channels, kernel_size=1, stride=1, causality_axis=causality_axis
+                )
+
+    def forward(self, x, temb):
+        h = x
+        h = self.norm1(h)
+        h = self.non_linearity(h)
+        h = self.conv1(h)
+
+        if temb is not None:
+            h = h + self.temb_proj(self.non_linearity(temb))[:, :, None, None]
+
+        h = self.norm2(h)
+        h = self.non_linearity(h)
+        h = self.dropout(h)
+        h = self.conv2(h)
+
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                x = self.conv_shortcut(x)
+            else:
+                x = self.nin_shortcut(x)
+
+        return x + h
+
+
+class AttnBlock(nn.Module):
+    def __init__(self, in_channels, norm_type="group"):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = Normalize(in_channels, normtype=norm_type)
+        self.q = ops.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0)
+        self.k = ops.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0)
+        self.v = ops.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0)
+        self.proj_out = ops.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0)
+
+    def forward(self, x):
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        b, c, h, w = q.shape
+        q = q.reshape(b, c, h * w).contiguous()
+        q = q.permute(0, 2, 1).contiguous()  # b,hw,c
+        k = k.reshape(b, c, h * w).contiguous()  # b,c,hw
+        w_ = torch.bmm(q, k).contiguous()  # b,hw,hw    w[b,i,j]=sum_c q[b,i,c]k[b,c,j]
+        w_ = w_ * (int(c) ** (-0.5))
+        w_ = torch.nn.functional.softmax(w_, dim=2)
+
+        # attend to values
+        v = v.reshape(b, c, h * w).contiguous()
+        w_ = w_.permute(0, 2, 1).contiguous()  # b,hw,hw (first hw of k, second of q)
+        h_ = torch.bmm(v, w_).contiguous()  # b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j]
+        h_ = h_.reshape(b, c, h, w).contiguous()
+
+        h_ = self.proj_out(h_)
+
+        return x + h_
+
+
+def make_attn(in_channels, attn_type="vanilla", norm_type="group"):
+    # Convert string to enum if needed
+    attn_type = AttentionType.str_to_enum(attn_type)
+
+    if attn_type != AttentionType.NONE:
+        logging.info(f"making attention of type '{attn_type.value}' with {in_channels} in_channels")
+    else:
+        logging.info(f"making identity attention with {in_channels} in_channels")
+
+    match attn_type:
+        case AttentionType.VANILLA:
+            return AttnBlock(in_channels, norm_type=norm_type)
+        case AttentionType.NONE:
+            return nn.Identity(in_channels)
+        case AttentionType.LINEAR:
+            raise NotImplementedError(f"Attention type {attn_type.value} is not supported yet.")
+        case _:
+            raise ValueError(f"Unknown attention type: {attn_type}")
+
+
+class Encoder(nn.Module):
+    def __init__(
+        self,
+        *,
+        ch,
+        out_ch,
+        ch_mult=(1, 2, 4, 8),
+        num_res_blocks,
+        attn_resolutions,
+        dropout=0.0,
+        resamp_with_conv=True,
+        in_channels,
+        resolution,
+        z_channels,
+        double_z=True,
+        attn_type="vanilla",
+        mid_block_add_attention=True,
+        norm_type="group",
+        causality_axis=CausalityAxis.WIDTH.value,
+        **ignore_kwargs,
+    ):
+        super().__init__()
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.z_channels = z_channels
+        self.double_z = double_z
+        self.norm_type = norm_type
+        # Convert string to enum if needed (for config loading)
+        causality_axis = CausalityAxis.str_to_enum(causality_axis)
+        self.attn_type = AttentionType.str_to_enum(attn_type)
+
+        # downsampling
+        self.conv_in = make_conv2d(
+            in_channels,
+            self.ch,
+            kernel_size=3,
+            stride=1,
+            causality_axis=causality_axis,
+        )
+
+        self.non_linearity = nn.SiLU()
+
+        curr_res = resolution
+        in_ch_mult = (1,) + tuple(ch_mult)
+        self.in_ch_mult = in_ch_mult
+        self.down = nn.ModuleList()
+
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = ch * in_ch_mult[i_level]
+            block_out = ch * ch_mult[i_level]
+
+            for _ in range(self.num_res_blocks):
+                block.append(
+                    ResnetBlock(
+                        in_channels=block_in,
+                        out_channels=block_out,
+                        temb_channels=self.temb_ch,
+                        dropout=dropout,
+                        norm_type=self.norm_type,
+                        causality_axis=causality_axis,
+                    )
+                )
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=self.attn_type, norm_type=self.norm_type))
+
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions - 1:
+                down.downsample = Downsample(block_in, resamp_with_conv, causality_axis=causality_axis)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(
+            in_channels=block_in,
+            out_channels=block_in,
+            temb_channels=self.temb_ch,
+            dropout=dropout,
+            norm_type=self.norm_type,
+            causality_axis=causality_axis,
+        )
+        if mid_block_add_attention:
+            self.mid.attn_1 = make_attn(block_in, attn_type=self.attn_type, norm_type=self.norm_type)
+        else:
+            self.mid.attn_1 = nn.Identity()
+        self.mid.block_2 = ResnetBlock(
+            in_channels=block_in,
+            out_channels=block_in,
+            temb_channels=self.temb_ch,
+            dropout=dropout,
+            norm_type=self.norm_type,
+            causality_axis=causality_axis,
+        )
+
+        # end
+        self.norm_out = Normalize(block_in, normtype=self.norm_type)
+        self.conv_out = make_conv2d(
+            block_in,
+            2 * z_channels if double_z else z_channels,
+            kernel_size=3,
+            stride=1,
+            causality_axis=causality_axis,
+        )
+
+    def forward(self, x):
+        """
+        Forward pass through the encoder.
+
+        Args:
+            x: Input tensor of shape [batch, channels, time, n_mels]
+
+        Returns:
+            Encoded latent representation
+        """
+        feature_maps = [self.conv_in(x)]
+
+        # Process each resolution level (from high to low resolution)
+        for resolution_level in range(self.num_resolutions):
+            # Apply residual blocks at current resolution level
+            for block_idx in range(self.num_res_blocks):
+                # Apply ResNet block with optional timestep embedding
+                current_features = self.down[resolution_level].block[block_idx](feature_maps[-1], temb=None)
+
+                # Apply attention if configured for this resolution level
+                if len(self.down[resolution_level].attn) > 0:
+                    current_features = self.down[resolution_level].attn[block_idx](current_features)
+
+                # Store processed features
+                feature_maps.append(current_features)
+
+            # Downsample spatial dimensions (except at the final resolution level)
+            if resolution_level != self.num_resolutions - 1:
+                downsampled_features = self.down[resolution_level].downsample(feature_maps[-1])
+                feature_maps.append(downsampled_features)
+
+        # === MIDDLE PROCESSING PHASE ===
+        # Take the lowest resolution features for middle processing
+        bottleneck_features = feature_maps[-1]
+
+        # Apply first middle ResNet block
+        bottleneck_features = self.mid.block_1(bottleneck_features, temb=None)
+
+        # Apply middle attention block
+        bottleneck_features = self.mid.attn_1(bottleneck_features)
+
+        # Apply second middle ResNet block
+        bottleneck_features = self.mid.block_2(bottleneck_features, temb=None)
+
+        # === OUTPUT PHASE ===
+        # Normalize the bottleneck features
+        output_features = self.norm_out(bottleneck_features)
+
+        # Apply non-linearity (SiLU activation)
+        output_features = self.non_linearity(output_features)
+
+        # Final convolution to produce latent representation
+        # [batch, channels, time, n_mels] -> [batch, 2 * z_channels if double_z else z_channels, time, n_mels]
+        return self.conv_out(output_features)
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self,
+        *,
+        ch,
+        out_ch,
+        ch_mult=(1, 2, 4, 8),
+        num_res_blocks,
+        attn_resolutions,
+        dropout=0.0,
+        resamp_with_conv=True,
+        in_channels,
+        resolution,
+        z_channels,
+        give_pre_end=False,
+        tanh_out=False,
+        attn_type="vanilla",
+        mid_block_add_attention=True,
+        norm_type="group",
+        causality_axis=CausalityAxis.WIDTH.value,
+        **ignorekwargs,
+    ):
+        super().__init__()
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.out_ch = out_ch
+        self.give_pre_end = give_pre_end
+        self.tanh_out = tanh_out
+        self.norm_type = norm_type
+        self.z_channels = z_channels
+        # Convert string to enum if needed (for config loading)
+        causality_axis = CausalityAxis.str_to_enum(causality_axis)
+        self.attn_type = AttentionType.str_to_enum(attn_type)
+
+        # compute block_in and curr_res at lowest res
+        block_in = ch * ch_mult[self.num_resolutions - 1]
+        curr_res = resolution // 2 ** (self.num_resolutions - 1)
+        self.z_shape = (1, z_channels, curr_res, curr_res)
+
+        # z to block_in
+        self.conv_in = make_conv2d(z_channels, block_in, kernel_size=3, stride=1, causality_axis=causality_axis)
+
+        self.non_linearity = nn.SiLU()
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(
+            in_channels=block_in,
+            out_channels=block_in,
+            temb_channels=self.temb_ch,
+            dropout=dropout,
+            norm_type=self.norm_type,
+            causality_axis=causality_axis,
+        )
+        if mid_block_add_attention:
+            self.mid.attn_1 = make_attn(block_in, attn_type=self.attn_type, norm_type=self.norm_type)
+        else:
+            self.mid.attn_1 = nn.Identity()
+        self.mid.block_2 = ResnetBlock(
+            in_channels=block_in,
+            out_channels=block_in,
+            temb_channels=self.temb_ch,
+            dropout=dropout,
+            norm_type=self.norm_type,
+            causality_axis=causality_axis,
+        )
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch * ch_mult[i_level]
+            for _ in range(self.num_res_blocks + 1):
+                block.append(
+                    ResnetBlock(
+                        in_channels=block_in,
+                        out_channels=block_out,
+                        temb_channels=self.temb_ch,
+                        dropout=dropout,
+                        norm_type=self.norm_type,
+                        causality_axis=causality_axis,
+                    )
+                )
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=self.attn_type, norm_type=self.norm_type))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = Upsample(block_in, resamp_with_conv, causality_axis=causality_axis)
+                curr_res = curr_res * 2
+            self.up.insert(0, up)  # prepend to get consistent order
+
+        # end
+        self.norm_out = Normalize(block_in, normtype=self.norm_type)
+        self.conv_out = make_conv2d(block_in, out_ch, kernel_size=3, stride=1, causality_axis=causality_axis)
+
+    def _adjust_output_shape(self, decoded_output, target_shape):
+        """
+        Adjust output shape to match target dimensions for variable-length audio.
+
+        This function handles the common case where decoded audio spectrograms need to be
+        resized to match a specific target shape.
+
+        Args:
+            decoded_output: Tensor of shape (batch, channels, time, frequency)
+            target_shape: Target shape tuple (batch, channels, time, frequency)
+
+        Returns:
+            Tensor adjusted to match target_shape exactly
+        """
+        # Current output shape: (batch, channels, time, frequency)
+        _, _, current_time, current_freq = decoded_output.shape
+        _, target_channels, target_time, target_freq = target_shape
+
+        # Step 1: Crop first to avoid exceeding target dimensions
+        decoded_output = decoded_output[
+            :, :target_channels, : min(current_time, target_time), : min(current_freq, target_freq)
+        ]
+
+        # Step 2: Calculate padding needed for time and frequency dimensions
+        time_padding_needed = target_time - decoded_output.shape[2]
+        freq_padding_needed = target_freq - decoded_output.shape[3]
+
+        # Step 3: Apply padding if needed
+        if time_padding_needed > 0 or freq_padding_needed > 0:
+            # PyTorch padding format: (pad_left, pad_right, pad_top, pad_bottom)
+            # For audio: pad_left/right = frequency, pad_top/bottom = time
+            padding = (
+                0,
+                max(freq_padding_needed, 0),  # frequency padding (left, right)
+                0,
+                max(time_padding_needed, 0),  # time padding (top, bottom)
+            )
+            decoded_output = F.pad(decoded_output, padding)
+
+        # Step 4: Final safety crop to ensure exact target shape
+        decoded_output = decoded_output[:, :target_channels, :target_time, :target_freq]
+
+        return decoded_output
+
+    def get_config(self):
+        return {
+            "ch": self.ch,
+            "out_ch": self.out_ch,
+            "ch_mult": self.ch_mult,
+            "num_res_blocks": self.num_res_blocks,
+            "in_channels": self.in_channels,
+            "resolution": self.resolution,
+            "z_channels": self.z_channels,
+        }
+
+    def forward(self, latent_features, target_shape=None):
+        """
+        Decode latent features back to audio spectrograms.
+
+        Args:
+            latent_features: Encoded latent representation of shape (batch, channels, height, width)
+            target_shape: Optional target output shape (batch, channels, time, frequency)
+                         If provided, output will be cropped/padded to match this shape
+
+        Returns:
+            Reconstructed audio spectrogram of shape (batch, channels, time, frequency)
+        """
+        assert target_shape is not None, "Target shape is required for CausalAudioAutoencoder Decoder"
+
+        # Transform latent features to decoder's internal feature dimension
+        hidden_features = self.conv_in(latent_features)
+
+        # Middle processing
+        hidden_features = self.mid.block_1(hidden_features, temb=None)
+        hidden_features = self.mid.attn_1(hidden_features)
+        hidden_features = self.mid.block_2(hidden_features, temb=None)
+
+        # Upsampling
+        # Progressively increase spatial resolution from lowest to highest
+        for resolution_level in reversed(range(self.num_resolutions)):
+            # Apply residual blocks at current resolution level
+            for block_index in range(self.num_res_blocks + 1):
+                hidden_features = self.up[resolution_level].block[block_index](hidden_features, temb=None)
+
+                if len(self.up[resolution_level].attn) > 0:
+                    hidden_features = self.up[resolution_level].attn[block_index](hidden_features)
+
+            if resolution_level != 0:
+                hidden_features = self.up[resolution_level].upsample(hidden_features)
+
+        # Output
+        if self.give_pre_end:
+            # Return intermediate features before final processing (for debugging/analysis)
+            decoded_output = hidden_features
+        else:
+            # Standard output path: normalize, activate, and convert to output channels
+            # Final normalization layer
+            hidden_features = self.norm_out(hidden_features)
+
+            # Apply SiLU (Swish) activation function
+            hidden_features = self.non_linearity(hidden_features)
+
+            # Final convolution to map to output channels (typically 2 for stereo audio)
+            decoded_output = self.conv_out(hidden_features)
+
+            # Optional tanh activation to bound output values to [-1, 1] range
+            if self.tanh_out:
+                decoded_output = torch.tanh(decoded_output)
+
+        # Adjust shape for audio data
+        if target_shape is not None:
+            decoded_output = self._adjust_output_shape(decoded_output, target_shape)
+
+        return decoded_output
+
+
+class processor(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.register_buffer("std-of-means", torch.empty(128))
+        self.register_buffer("mean-of-means", torch.empty(128))
+
+    def un_normalize(self, x):
+        return (x * self.get_buffer("std-of-means").to(x)) + self.get_buffer("mean-of-means").to(x)
+
+    def normalize(self, x):
+        return (x - self.get_buffer("mean-of-means").to(x)) / self.get_buffer("std-of-means").to(x)
+
+
+class CausalAudioAutoencoder(nn.Module):
+    def __init__(self, config=None):
+        super().__init__()
+
+        if config is None:
+            config = self.get_default_config()
+
+        model_config = config.get("model", {}).get("params", {})
+
+        self.sampling_rate = model_config.get(
+            "sampling_rate", config.get("sampling_rate", 16000)
+        )
+        encoder_config = model_config.get("encoder", model_config.get("ddconfig", {}))
+        decoder_config = model_config.get("decoder", encoder_config)
+
+        # Load mel spectrogram parameters
+        self.mel_bins = encoder_config.get("mel_bins", 64)
+        self.mel_hop_length = config.get("preprocessing", {}).get("stft", {}).get("hop_length", 160)
+        self.n_fft = config.get("preprocessing", {}).get("stft", {}).get("filter_length", 1024)
+
+        # Store causality configuration at VAE level (not just in encoder internals)
+        causality_axis_value = encoder_config.get("causality_axis", CausalityAxis.HEIGHT.value)
+        self.causality_axis = CausalityAxis.str_to_enum(causality_axis_value)
+        self.is_causal = self.causality_axis == CausalityAxis.HEIGHT
+
+        self.encoder = Encoder(**encoder_config)
+        self.decoder = Decoder(**decoder_config)
+
+        self.per_channel_statistics = processor()
+
+    def get_default_config(self):
+        ddconfig = {
+            "double_z": True,
+            "mel_bins": 64,
+            "z_channels": 8,
+            "resolution": 256,
+            "downsample_time": False,
+            "in_channels": 2,
+            "out_ch": 2,
+            "ch": 128,
+            "ch_mult": [1, 2, 4],
+            "num_res_blocks": 2,
+            "attn_resolutions": [],
+            "dropout": 0.0,
+            "mid_block_add_attention": False,
+            "norm_type": "pixel",
+            "causality_axis": "height",
+        }
+
+        config = {
+            "model": {
+                "params": {
+                    "ddconfig": ddconfig,
+                    "sampling_rate": 16000,
+                }
+            },
+            "preprocessing": {
+                "stft": {
+                    "filter_length": 1024,
+                    "hop_length": 160,
+                },
+            },
+        }
+
+        return config
+
+    def get_config(self):
+        return {
+            "sampling_rate": self.sampling_rate,
+            "mel_bins": self.mel_bins,
+            "mel_hop_length": self.mel_hop_length,
+            "n_fft": self.n_fft,
+            "causality_axis": self.causality_axis.value,
+            "is_causal": self.is_causal,
+        }
+
+    def encode(self, x):
+        return self.encoder(x)
+
+    def decode(self, x, target_shape=None):
+        return self.decoder(x, target_shape=target_shape)

vendor/ComfyUI/comfy/ldm/lightricks/vae/causal_conv3d.py

@@ -0,0 +1,90 @@
+from typing import Tuple, Union
+
+import threading
+import torch
+import torch.nn as nn
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+class CausalConv3d(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        kernel_size: int = 3,
+        stride: Union[int, Tuple[int]] = 1,
+        dilation: int = 1,
+        groups: int = 1,
+        spatial_padding_mode: str = "zeros",
+        **kwargs,
+    ):
+        super().__init__()
+
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+
+        if isinstance(stride, int):
+            self.time_stride = stride
+        else:
+            self.time_stride = stride[0]
+
+        kernel_size = (kernel_size, kernel_size, kernel_size)
+        self.time_kernel_size = kernel_size[0]
+
+        dilation = (dilation, 1, 1)
+
+        height_pad = kernel_size[1] // 2
+        width_pad = kernel_size[2] // 2
+        padding = (0, height_pad, width_pad)
+
+        self.conv = ops.Conv3d(
+            in_channels,
+            out_channels,
+            kernel_size,
+            stride=stride,
+            dilation=dilation,
+            padding=padding,
+            padding_mode=spatial_padding_mode,
+            groups=groups,
+        )
+        self.temporal_cache_state={}
+
+    def forward(self, x, causal: bool = True):
+        tid = threading.get_ident()
+
+        cached, is_end = self.temporal_cache_state.get(tid, (None, False))
+        if cached is None:
+            padding_length = self.time_kernel_size - 1
+            if not causal:
+                padding_length = padding_length // 2
+            if x.shape[2] == 0:
+                return x
+            cached = x[:, :, :1, :, :].repeat((1, 1, padding_length, 1, 1))
+        pieces = [ cached, x ]
+        if is_end and not causal:
+            pieces.append(x[:, :, -1:, :, :].repeat((1, 1, (self.time_kernel_size - 1) // 2, 1, 1)))
+        input_length = sum([piece.shape[2] for piece in pieces])
+        cache_length = (self.time_kernel_size - self.time_stride) + ((input_length - self.time_kernel_size) % self.time_stride)
+
+        needs_caching = not is_end
+        if needs_caching and cache_length == 0:
+            self.temporal_cache_state[tid] = (x[:, :, :0, :, :], False)
+            needs_caching = False
+        if needs_caching and x.shape[2] >= cache_length:
+            needs_caching = False
+            self.temporal_cache_state[tid] = (x[:, :, -cache_length:, :, :], False)
+
+        x = torch.cat(pieces, dim=2)
+        del pieces
+        del cached
+
+        if needs_caching:
+            self.temporal_cache_state[tid] = (x[:, :, -cache_length:, :, :], False)
+        elif is_end:
+            self.temporal_cache_state[tid] = (None, True)
+
+        return self.conv(x) if x.shape[2] >= self.time_kernel_size else x[:, :, :0, :, :]
+
+    @property
+    def weight(self):
+        return self.conv.weight

vendor/ComfyUI/comfy/ldm/lightricks/vae/conv_nd_factory.py

@@ -0,0 +1,90 @@
+from typing import Tuple, Union
+
+
+from .dual_conv3d import DualConv3d
+from .causal_conv3d import CausalConv3d
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+def make_conv_nd(
+    dims: Union[int, Tuple[int, int]],
+    in_channels: int,
+    out_channels: int,
+    kernel_size: int,
+    stride=1,
+    padding=0,
+    dilation=1,
+    groups=1,
+    bias=True,
+    causal=False,
+    spatial_padding_mode="zeros",
+    temporal_padding_mode="zeros",
+):
+    if not (spatial_padding_mode == temporal_padding_mode or causal):
+        raise NotImplementedError("spatial and temporal padding modes must be equal")
+    if dims == 2:
+        return ops.Conv2d(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=padding,
+            dilation=dilation,
+            groups=groups,
+            bias=bias,
+            padding_mode=spatial_padding_mode,
+        )
+    elif dims == 3:
+        if causal:
+            return CausalConv3d(
+                in_channels=in_channels,
+                out_channels=out_channels,
+                kernel_size=kernel_size,
+                stride=stride,
+                padding=padding,
+                dilation=dilation,
+                groups=groups,
+                bias=bias,
+                spatial_padding_mode=spatial_padding_mode,
+            )
+        return ops.Conv3d(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=padding,
+            dilation=dilation,
+            groups=groups,
+            bias=bias,
+            padding_mode=spatial_padding_mode,
+        )
+    elif dims == (2, 1):
+        return DualConv3d(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=padding,
+            bias=bias,
+            padding_mode=spatial_padding_mode,
+        )
+    else:
+        raise ValueError(f"unsupported dimensions: {dims}")
+
+
+def make_linear_nd(
+    dims: int,
+    in_channels: int,
+    out_channels: int,
+    bias=True,
+):
+    if dims == 2:
+        return ops.Conv2d(
+            in_channels=in_channels, out_channels=out_channels, kernel_size=1, bias=bias
+        )
+    elif dims == 3 or dims == (2, 1):
+        return ops.Conv3d(
+            in_channels=in_channels, out_channels=out_channels, kernel_size=1, bias=bias
+        )
+    else:
+        raise ValueError(f"unsupported dimensions: {dims}")

vendor/ComfyUI/comfy/ldm/lightricks/vae/dual_conv3d.py

@@ -0,0 +1,217 @@
+import math
+from typing import Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+
+
+class DualConv3d(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        kernel_size,
+        stride: Union[int, Tuple[int, int, int]] = 1,
+        padding: Union[int, Tuple[int, int, int]] = 0,
+        dilation: Union[int, Tuple[int, int, int]] = 1,
+        groups=1,
+        bias=True,
+        padding_mode="zeros",
+    ):
+        super(DualConv3d, self).__init__()
+
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.padding_mode = padding_mode
+        # Ensure kernel_size, stride, padding, and dilation are tuples of length 3
+        if isinstance(kernel_size, int):
+            kernel_size = (kernel_size, kernel_size, kernel_size)
+        if kernel_size == (1, 1, 1):
+            raise ValueError(
+                "kernel_size must be greater than 1. Use make_linear_nd instead."
+            )
+        if isinstance(stride, int):
+            stride = (stride, stride, stride)
+        if isinstance(padding, int):
+            padding = (padding, padding, padding)
+        if isinstance(dilation, int):
+            dilation = (dilation, dilation, dilation)
+
+        # Set parameters for convolutions
+        self.groups = groups
+        self.bias = bias
+
+        # Define the size of the channels after the first convolution
+        intermediate_channels = (
+            out_channels if in_channels < out_channels else in_channels
+        )
+
+        # Define parameters for the first convolution
+        self.weight1 = nn.Parameter(
+            torch.Tensor(
+                intermediate_channels,
+                in_channels // groups,
+                1,
+                kernel_size[1],
+                kernel_size[2],
+            )
+        )
+        self.stride1 = (1, stride[1], stride[2])
+        self.padding1 = (0, padding[1], padding[2])
+        self.dilation1 = (1, dilation[1], dilation[2])
+        if bias:
+            self.bias1 = nn.Parameter(torch.Tensor(intermediate_channels))
+        else:
+            self.register_parameter("bias1", None)
+
+        # Define parameters for the second convolution
+        self.weight2 = nn.Parameter(
+            torch.Tensor(
+                out_channels, intermediate_channels // groups, kernel_size[0], 1, 1
+            )
+        )
+        self.stride2 = (stride[0], 1, 1)
+        self.padding2 = (padding[0], 0, 0)
+        self.dilation2 = (dilation[0], 1, 1)
+        if bias:
+            self.bias2 = nn.Parameter(torch.Tensor(out_channels))
+        else:
+            self.register_parameter("bias2", None)
+
+        # Initialize weights and biases
+        self.reset_parameters()
+
+    def reset_parameters(self):
+        nn.init.kaiming_uniform_(self.weight1, a=math.sqrt(5))
+        nn.init.kaiming_uniform_(self.weight2, a=math.sqrt(5))
+        if self.bias:
+            fan_in1, _ = nn.init._calculate_fan_in_and_fan_out(self.weight1)
+            bound1 = 1 / math.sqrt(fan_in1)
+            nn.init.uniform_(self.bias1, -bound1, bound1)
+            fan_in2, _ = nn.init._calculate_fan_in_and_fan_out(self.weight2)
+            bound2 = 1 / math.sqrt(fan_in2)
+            nn.init.uniform_(self.bias2, -bound2, bound2)
+
+    def forward(self, x, use_conv3d=False, skip_time_conv=False):
+        if use_conv3d:
+            return self.forward_with_3d(x=x, skip_time_conv=skip_time_conv)
+        else:
+            return self.forward_with_2d(x=x, skip_time_conv=skip_time_conv)
+
+    def forward_with_3d(self, x, skip_time_conv):
+        # First convolution
+        x = F.conv3d(
+            x,
+            self.weight1,
+            self.bias1,
+            self.stride1,
+            self.padding1,
+            self.dilation1,
+            self.groups,
+            padding_mode=self.padding_mode,
+        )
+
+        if skip_time_conv:
+            return x
+
+        # Second convolution
+        x = F.conv3d(
+            x,
+            self.weight2,
+            self.bias2,
+            self.stride2,
+            self.padding2,
+            self.dilation2,
+            self.groups,
+            padding_mode=self.padding_mode,
+        )
+
+        return x
+
+    def forward_with_2d(self, x, skip_time_conv):
+        b, c, d, h, w = x.shape
+
+        # First 2D convolution
+        x = rearrange(x, "b c d h w -> (b d) c h w")
+        # Squeeze the depth dimension out of weight1 since it's 1
+        weight1 = self.weight1.squeeze(2)
+        # Select stride, padding, and dilation for the 2D convolution
+        stride1 = (self.stride1[1], self.stride1[2])
+        padding1 = (self.padding1[1], self.padding1[2])
+        dilation1 = (self.dilation1[1], self.dilation1[2])
+        x = F.conv2d(
+            x,
+            weight1,
+            self.bias1,
+            stride1,
+            padding1,
+            dilation1,
+            self.groups,
+            padding_mode=self.padding_mode,
+        )
+
+        _, _, h, w = x.shape
+
+        if skip_time_conv:
+            x = rearrange(x, "(b d) c h w -> b c d h w", b=b)
+            return x
+
+        # Second convolution which is essentially treated as a 1D convolution across the 'd' dimension
+        x = rearrange(x, "(b d) c h w -> (b h w) c d", b=b)
+
+        # Reshape weight2 to match the expected dimensions for conv1d
+        weight2 = self.weight2.squeeze(-1).squeeze(-1)
+        # Use only the relevant dimension for stride, padding, and dilation for the 1D convolution
+        stride2 = self.stride2[0]
+        padding2 = self.padding2[0]
+        dilation2 = self.dilation2[0]
+        x = F.conv1d(
+            x,
+            weight2,
+            self.bias2,
+            stride2,
+            padding2,
+            dilation2,
+            self.groups,
+            padding_mode=self.padding_mode,
+        )
+        x = rearrange(x, "(b h w) c d -> b c d h w", b=b, h=h, w=w)
+
+        return x
+
+    @property
+    def weight(self):
+        return self.weight2
+
+
+def test_dual_conv3d_consistency():
+    # Initialize parameters
+    in_channels = 3
+    out_channels = 5
+    kernel_size = (3, 3, 3)
+    stride = (2, 2, 2)
+    padding = (1, 1, 1)
+
+    # Create an instance of the DualConv3d class
+    dual_conv3d = DualConv3d(
+        in_channels=in_channels,
+        out_channels=out_channels,
+        kernel_size=kernel_size,
+        stride=stride,
+        padding=padding,
+        bias=True,
+    )
+
+    # Example input tensor
+    test_input = torch.randn(1, 3, 10, 10, 10)
+
+    # Perform forward passes with both 3D and 2D settings
+    output_conv3d = dual_conv3d(test_input, use_conv3d=True)
+    output_2d = dual_conv3d(test_input, use_conv3d=False)
+
+    # Assert that the outputs from both methods are sufficiently close
+    assert torch.allclose(
+        output_conv3d, output_2d, atol=1e-6
+    ), "Outputs are not consistent between 3D and 2D convolutions."

vendor/ComfyUI/comfy/ldm/lightricks/vae/pixel_norm.py

@@ -0,0 +1,12 @@
+import torch
+from torch import nn
+
+
+class PixelNorm(nn.Module):
+    def __init__(self, dim=1, eps=1e-8):
+        super(PixelNorm, self).__init__()
+        self.dim = dim
+        self.eps = eps
+
+    def forward(self, x):
+        return x / torch.sqrt(torch.mean(x**2, dim=self.dim, keepdim=True) + self.eps)