DiffSynth-Studio/diffsynth/models/flux2_vae.py

# 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 math
from typing import Dict, Optional, Tuple, Union

import torch
import torch.nn as nn
from einops import rearrange

from diffusers.models.autoencoders.autoencoder_kl_flux2 import Decoder, Encoder


class Flux2VAE(torch.nn.Module):
    r"""
    A VAE model with KL loss for encoding images into latents and decoding latent representations into images.

    This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
    for all models (such as downloading or saving).

    Parameters:
        in_channels (int, *optional*, defaults to 3): Number of channels in the input image.
        out_channels (int,  *optional*, defaults to 3): Number of channels in the output.
        down_block_types (`Tuple[str]`, *optional*, defaults to `("DownEncoderBlock2D",)`):
            Tuple of downsample block types.
        up_block_types (`Tuple[str]`, *optional*, defaults to `("UpDecoderBlock2D",)`):
            Tuple of upsample block types.
        block_out_channels (`Tuple[int]`, *optional*, defaults to `(64,)`):
            Tuple of block output channels.
        act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
        latent_channels (`int`, *optional*, defaults to 4): Number of channels in the latent space.
        sample_size (`int`, *optional*, defaults to `32`): Sample input size.
        force_upcast (`bool`, *optional*, default to `True`):
            If enabled it will force the VAE to run in float32 for high image resolution pipelines, such as SD-XL. VAE
            can be fine-tuned / trained to a lower range without losing too much precision in which case `force_upcast`
            can be set to `False` - see: https://huggingface.co/madebyollin/sdxl-vae-fp16-fix
        mid_block_add_attention (`bool`, *optional*, default to `True`):
            If enabled, the mid_block of the Encoder and Decoder will have attention blocks. If set to false, the
            mid_block will only have resnet blocks
    """

    _supports_gradient_checkpointing = True
    _no_split_modules = ["BasicTransformerBlock", "ResnetBlock2D"]

    def __init__(
        self,
        in_channels: int = 3,
        out_channels: int = 3,
        down_block_types: Tuple[str, ...] = (
            "DownEncoderBlock2D",
            "DownEncoderBlock2D",
            "DownEncoderBlock2D",
            "DownEncoderBlock2D",
        ),
        up_block_types: Tuple[str, ...] = (
            "UpDecoderBlock2D",
            "UpDecoderBlock2D",
            "UpDecoderBlock2D",
            "UpDecoderBlock2D",
        ),
        block_out_channels: Tuple[int, ...] = (
            128,
            256,
            512,
            512,
        ),
        layers_per_block: int = 2,
        act_fn: str = "silu",
        latent_channels: int = 32,
        norm_num_groups: int = 32,
        sample_size: int = 1024,  # YiYi notes: not sure
        force_upcast: bool = True,
        use_quant_conv: bool = True,
        use_post_quant_conv: bool = True,
        mid_block_add_attention: bool = True,
        batch_norm_eps: float = 1e-4,
        batch_norm_momentum: float = 0.1,
        patch_size: Tuple[int, int] = (2, 2),
    ):
        super().__init__()

        # pass init params to Encoder
        self.encoder = Encoder(
            in_channels=in_channels,
            out_channels=latent_channels,
            down_block_types=down_block_types,
            block_out_channels=block_out_channels,
            layers_per_block=layers_per_block,
            act_fn=act_fn,
            norm_num_groups=norm_num_groups,
            double_z=True,
            mid_block_add_attention=mid_block_add_attention,
        )

        # pass init params to Decoder
        self.decoder = Decoder(
            in_channels=latent_channels,
            out_channels=out_channels,
            up_block_types=up_block_types,
            block_out_channels=block_out_channels,
            layers_per_block=layers_per_block,
            norm_num_groups=norm_num_groups,
            act_fn=act_fn,
            mid_block_add_attention=mid_block_add_attention,
        )

        self.quant_conv = nn.Conv2d(2 * latent_channels, 2 * latent_channels, 1) if use_quant_conv else None
        self.post_quant_conv = nn.Conv2d(latent_channels, latent_channels, 1) if use_post_quant_conv else None

        self.bn = nn.BatchNorm2d(
            math.prod(patch_size) * latent_channels,
            eps=batch_norm_eps,
            momentum=batch_norm_momentum,
            affine=False,
            track_running_stats=True,
        )

        self.use_slicing = False
        self.use_tiling = False

    @property
    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
    def attn_processors(self):
        r"""
        Returns:
            `dict` of attention processors: A dictionary containing all attention processors used in the model with
            indexed by its weight name.
        """
        # set recursively
        processors = {}

        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors):
            if hasattr(module, "get_processor"):
                processors[f"{name}.processor"] = module.get_processor()

            for sub_name, child in module.named_children():
                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)

            return processors

        for name, module in self.named_children():
            fn_recursive_add_processors(name, module, processors)

        return processors

    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
    def set_attn_processor(self, processor):
        r"""
        Sets the attention processor to use to compute attention.

        Parameters:
            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
                The instantiated processor class or a dictionary of processor classes that will be set as the processor
                for **all** `Attention` layers.

                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
                processor. This is strongly recommended when setting trainable attention processors.

        """
        count = len(self.attn_processors.keys())

        if isinstance(processor, dict) and len(processor) != count:
            raise ValueError(
                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
            )

        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
            if hasattr(module, "set_processor"):
                if not isinstance(processor, dict):
                    module.set_processor(processor)
                else:
                    module.set_processor(processor.pop(f"{name}.processor"))

            for sub_name, child in module.named_children():
                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)

        for name, module in self.named_children():
            fn_recursive_attn_processor(name, module, processor)

    def _encode(self, x: torch.Tensor) -> torch.Tensor:
        batch_size, num_channels, height, width = x.shape

        if self.use_tiling and (width > self.tile_sample_min_size or height > self.tile_sample_min_size):
            return self._tiled_encode(x)

        enc = self.encoder(x)
        if self.quant_conv is not None:
            enc = self.quant_conv(enc)

        return enc

    def encode(
        self, x: torch.Tensor, return_dict: bool = True
    ):
        """
        Encode a batch of images into latents.

        Args:
            x (`torch.Tensor`): Input batch of images.
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple.

        Returns:
                The latent representations of the encoded images. If `return_dict` is True, a
                [`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain `tuple` is returned.
        """
        if self.use_slicing and x.shape[0] > 1:
            encoded_slices = [self._encode(x_slice) for x_slice in x.split(1)]
            h = torch.cat(encoded_slices)
        else:
            h = self._encode(x)


        h = rearrange(h, "B C (H P) (W Q) -> B (C P Q) H W", P=2, Q=2)
        h = h[:, :128]
        latents_bn_mean = self.bn.running_mean.view(1, -1, 1, 1).to(h.device, h.dtype)
        latents_bn_std = torch.sqrt(self.bn.running_var.view(1, -1, 1, 1) + 0.0001).to(
            h.device, h.dtype
        )
        h = (h - latents_bn_mean) / latents_bn_std
        return h

    def _decode(self, z: torch.Tensor, return_dict: bool = True):
        if self.use_tiling and (z.shape[-1] > self.tile_latent_min_size or z.shape[-2] > self.tile_latent_min_size):
            return self.tiled_decode(z, return_dict=return_dict)

        if self.post_quant_conv is not None:
            z = self.post_quant_conv(z)

        dec = self.decoder(z)

        if not return_dict:
            return (dec,)

        return dec

    def decode(
        self, z: torch.FloatTensor, return_dict: bool = True, generator=None
    ):
        latents_bn_mean = self.bn.running_mean.view(1, -1, 1, 1).to(z.device, z.dtype)
        latents_bn_std = torch.sqrt(self.bn.running_var.view(1, -1, 1, 1) + 0.0001).to(
            z.device, z.dtype
        )
        z = z * latents_bn_std + latents_bn_mean
        z = rearrange(z, "B (C P Q) H W -> B C (H P) (W Q)", P=2, Q=2)
        """
        Decode a batch of images.

        Args:
            z (`torch.Tensor`): Input batch of latent vectors.
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether to return a [`~models.vae.DecoderOutput`] instead of a plain tuple.

        Returns:
            [`~models.vae.DecoderOutput`] or `tuple`:
                If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is
                returned.

        """
        if self.use_slicing and z.shape[0] > 1:
            decoded_slices = [self._decode(z_slice) for z_slice in z.split(1)]
            decoded = torch.cat(decoded_slices)
        else:
            decoded = self._decode(z)

        if not return_dict:
            return (decoded,)

        return decoded

    def blend_v(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor:
        blend_extent = min(a.shape[2], b.shape[2], blend_extent)
        for y in range(blend_extent):
            b[:, :, y, :] = a[:, :, -blend_extent + y, :] * (1 - y / blend_extent) + b[:, :, y, :] * (y / blend_extent)
        return b

    def blend_h(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor:
        blend_extent = min(a.shape[3], b.shape[3], blend_extent)
        for x in range(blend_extent):
            b[:, :, :, x] = a[:, :, :, -blend_extent + x] * (1 - x / blend_extent) + b[:, :, :, x] * (x / blend_extent)
        return b

    def _tiled_encode(self, x: torch.Tensor) -> torch.Tensor:
        r"""Encode a batch of images using a tiled encoder.

        When this option is enabled, the VAE will split the input tensor into tiles to compute encoding in several
        steps. This is useful to keep memory use constant regardless of image size. The end result of tiled encoding is
        different from non-tiled encoding because each tile uses a different encoder. To avoid tiling artifacts, the
        tiles overlap and are blended together to form a smooth output. You may still see tile-sized changes in the
        output, but they should be much less noticeable.

        Args:
            x (`torch.Tensor`): Input batch of images.

        Returns:
            `torch.Tensor`:
                The latent representation of the encoded videos.
        """

        overlap_size = int(self.tile_sample_min_size * (1 - self.tile_overlap_factor))
        blend_extent = int(self.tile_latent_min_size * self.tile_overlap_factor)
        row_limit = self.tile_latent_min_size - blend_extent

        # Split the image into 512x512 tiles and encode them separately.
        rows = []
        for i in range(0, x.shape[2], overlap_size):
            row = []
            for j in range(0, x.shape[3], overlap_size):
                tile = x[:, :, i : i + self.tile_sample_min_size, j : j + self.tile_sample_min_size]
                tile = self.encoder(tile)
                if self.config.use_quant_conv:
                    tile = self.quant_conv(tile)
                row.append(tile)
            rows.append(row)
        result_rows = []
        for i, row in enumerate(rows):
            result_row = []
            for j, tile in enumerate(row):
                # blend the above tile and the left tile
                # to the current tile and add the current tile to the result row
                if i > 0:
                    tile = self.blend_v(rows[i - 1][j], tile, blend_extent)
                if j > 0:
                    tile = self.blend_h(row[j - 1], tile, blend_extent)
                result_row.append(tile[:, :, :row_limit, :row_limit])
            result_rows.append(torch.cat(result_row, dim=3))

        enc = torch.cat(result_rows, dim=2)
        return enc

    def tiled_encode(self, x: torch.Tensor, return_dict: bool = True):
        r"""Encode a batch of images using a tiled encoder.

        When this option is enabled, the VAE will split the input tensor into tiles to compute encoding in several
        steps. This is useful to keep memory use constant regardless of image size. The end result of tiled encoding is
        different from non-tiled encoding because each tile uses a different encoder. To avoid tiling artifacts, the
        tiles overlap and are blended together to form a smooth output. You may still see tile-sized changes in the
        output, but they should be much less noticeable.

        Args:
            x (`torch.Tensor`): Input batch of images.
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple.

        Returns:
            [`~models.autoencoder_kl.AutoencoderKLOutput`] or `tuple`:
                If return_dict is True, a [`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain
                `tuple` is returned.
        """

        overlap_size = int(self.tile_sample_min_size * (1 - self.tile_overlap_factor))
        blend_extent = int(self.tile_latent_min_size * self.tile_overlap_factor)
        row_limit = self.tile_latent_min_size - blend_extent

        # Split the image into 512x512 tiles and encode them separately.
        rows = []
        for i in range(0, x.shape[2], overlap_size):
            row = []
            for j in range(0, x.shape[3], overlap_size):
                tile = x[:, :, i : i + self.tile_sample_min_size, j : j + self.tile_sample_min_size]
                tile = self.encoder(tile)
                if self.config.use_quant_conv:
                    tile = self.quant_conv(tile)
                row.append(tile)
            rows.append(row)
        result_rows = []
        for i, row in enumerate(rows):
            result_row = []
            for j, tile in enumerate(row):
                # blend the above tile and the left tile
                # to the current tile and add the current tile to the result row
                if i > 0:
                    tile = self.blend_v(rows[i - 1][j], tile, blend_extent)
                if j > 0:
                    tile = self.blend_h(row[j - 1], tile, blend_extent)
                result_row.append(tile[:, :, :row_limit, :row_limit])
            result_rows.append(torch.cat(result_row, dim=3))

        moments = torch.cat(result_rows, dim=2)
        return moments

    def tiled_decode(self, z: torch.Tensor, return_dict: bool = True):
        r"""
        Decode a batch of images using a tiled decoder.

        Args:
            z (`torch.Tensor`): Input batch of latent vectors.
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`~models.vae.DecoderOutput`] instead of a plain tuple.

        Returns:
            [`~models.vae.DecoderOutput`] or `tuple`:
                If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is
                returned.
        """
        overlap_size = int(self.tile_latent_min_size * (1 - self.tile_overlap_factor))
        blend_extent = int(self.tile_sample_min_size * self.tile_overlap_factor)
        row_limit = self.tile_sample_min_size - blend_extent

        # Split z into overlapping 64x64 tiles and decode them separately.
        # The tiles have an overlap to avoid seams between tiles.
        rows = []
        for i in range(0, z.shape[2], overlap_size):
            row = []
            for j in range(0, z.shape[3], overlap_size):
                tile = z[:, :, i : i + self.tile_latent_min_size, j : j + self.tile_latent_min_size]
                if self.config.use_post_quant_conv:
                    tile = self.post_quant_conv(tile)
                decoded = self.decoder(tile)
                row.append(decoded)
            rows.append(row)
        result_rows = []
        for i, row in enumerate(rows):
            result_row = []
            for j, tile in enumerate(row):
                # blend the above tile and the left tile
                # to the current tile and add the current tile to the result row
                if i > 0:
                    tile = self.blend_v(rows[i - 1][j], tile, blend_extent)
                if j > 0:
                    tile = self.blend_h(row[j - 1], tile, blend_extent)
                result_row.append(tile[:, :, :row_limit, :row_limit])
            result_rows.append(torch.cat(result_row, dim=3))

        dec = torch.cat(result_rows, dim=2)
        if not return_dict:
            return (dec,)

        return dec

    def forward(
        self,
        sample: torch.Tensor,
        sample_posterior: bool = False,
        return_dict: bool = True,
        generator: Optional[torch.Generator] = None,
    ):
        r"""
        Args:
            sample (`torch.Tensor`): Input sample.
            sample_posterior (`bool`, *optional*, defaults to `False`):
                Whether to sample from the posterior.
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`DecoderOutput`] instead of a plain tuple.
        """
        x = sample
        posterior = self.encode(x).latent_dist
        if sample_posterior:
            z = posterior.sample(generator=generator)
        else:
            z = posterior.mode()
        dec = self.decode(z).sample

        if not return_dict:
            return (dec,)

        return dec