diffusers/pr_14921/en/api/models/autoencoderkl_qwenimage21.md

AutoencoderKLQwenImage21

The 64-channel variational auto-encoder used by Qwen-Image 2.1. It compresses 16x spatially, and its per-channel latents_mean / latents_std are part of the config rather than a single scaling factor.

import torch
from diffusers import AutoencoderKLQwenImage21

vae = AutoencoderKLQwenImage21.from_pretrained("Qwen/Qwen-Image-2.1", subfolder="vae", dtype=torch.bfloat16)

AutoencoderKLQwenImage21[[diffusers.AutoencoderKLQwenImage21]]

diffusers.AutoencoderKLQwenImage21[[diffusers.AutoencoderKLQwenImage21]]

diffusers.AutoencoderKLQwenImage21(base_dim: int = 96, decoder_base_dim: int | None = 144, z_dim: int = 64, dim_mult: list = [1, 2, 4, 8, 8], num_res_blocks: int = 2, attn_scales: list = [], temperal_downsample: list = [False, True, True, True], dropout: float = 0.0, latents_mean: list = [0.5126, 0.7721, -0.0631, 1.3506, -0.7855, -2.1025, -0.3458, 1.3722, 1.8873, -1.7177, -0.651, 0.2732, 0.7562, -0.6163, -1.0277, 3.8363, 2.021, 0.0472, 0.932, 2.0087, 2.4954, -0.1391, -1.4249, 1.8464, -0.5236, 1.2826, 3.7046, -1.3035, 2.7286, -1.4518, -1.9036, -1.9955, -0.0342, -1.0265, -0.7636, 3.0555, 0.0746, -3.0751, -0.1076, 1.7376, -1.0914, -1.9435, -0.2784, -1.368, 0.4809, -0.4433, 0.3764, 0.5729, -2.0595, 1.096, -1.326, -2.0211, -5.0179, 0.5275, 4.0162, 1.8505, 0.3026, 1.9373, 1.4937, 0.2632, 0.5547, -1.7121, -0.1562, 0.0304], latents_std: list = [3.2001, 3.2936, 3.4321, 3.0091, 3.1061, 4.0379, 4.0705, 3.791, 3.0785, 3.65, 3.9308, 3.0904, 2.8778, 3.7675, 3.732, 5.0756, 3.2864, 4.0397, 3.1317, 4.0443, 2.9249, 3.9454, 3.0988, 4.2489, 3.4896, 3.8513, 3.9323, 3.4719, 3.7498, 4.283, 3.5694, 4.2467, 3.9037, 3.2947, 5.077, 3.5075, 3.27, 3.4767, 2.8063, 5.1125, 3.5327, 4.7833, 3.1286, 4.1819, 3.8527, 3.8312, 3.5605, 4.3875, 3.9624, 4.0168, 3.5643, 4.055, 5.5614, 4.2963, 4.408, 3.4959, 3.8747, 3.7608, 3.5735, 3.149, 3.7662, 3.6746, 3.4563, 3.8161], is_residual: bool = True, in_channels: int = 4, out_channels: int = 4, patch_size: int | None = None, scale_factor_temporal: int | None = 8, scale_factor_spatial: int | None = 16)

Source

A VAE model with KL loss for encoding videos into latents and decoding latent representations into videos. Introduced in [Qwen Image 2].

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

decode[[diffusers.AutoencoderKLQwenImage21.decode]]

decode(z: Tensor, return_dict: bool = True)

Source

Parameters:

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.

Decode a batch of images.

encode[[diffusers.AutoencoderKLQwenImage21.encode]]

encode(x: Tensor, return_dict: bool = True)

Source

Parameters:

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 videos. If return_dict is True, a ~models.autoencoder_kl.AutoencoderKLOutput is returned, otherwise a plain tuple is returned.

Encode a batch of images into latents.

enable_tiling[[diffusers.AutoencoderKLQwenImage21.enable_tiling]]

enable_tiling(tile_sample_min_height: int | None = None, tile_sample_min_width: int | None = None, tile_sample_stride_height: float | None = None, tile_sample_stride_width: float | None = None)

Source

Parameters:

tile_sample_min_height (int, optional) : The minimum height required for a sample to be separated into tiles across the height dimension.

tile_sample_min_width (int, optional) : The minimum width required for a sample to be separated into tiles across the width dimension.

tile_sample_stride_height (int, optional) : The minimum amount of overlap between two consecutive vertical tiles. This is to ensure that there are no tiling artifacts produced across the height dimension.

tile_sample_stride_width (int, optional) : The stride between two consecutive horizontal tiles. This is to ensure that there are no tiling artifacts produced across the width dimension.

Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow processing larger images.

forward[[diffusers.AutoencoderKLQwenImage21.forward]]

forward(sample: Tensor, sample_posterior: bool = False, return_dict: bool = True, generator: typing.Optional[torch.Generator] = None)

Source

Parameters:

sample (torch.Tensor) : Input sample.

sample_posterior (bool, optional, defaults to False) : Whether to sample from the posterior instead of taking its mode.

return_dict (bool, optional, defaults to True) : Whether or not to return a DecoderOutput instead of a plain tuple.

generator (torch.Generator, optional) : Generator used when sample_posterior is True.

Returns: DecoderOutput or tuple

DecoderOutput if return_dict is True, otherwise a plain tuple.

tiled_decode[[diffusers.AutoencoderKLQwenImage21.tiled_decode]]

tiled_decode(z: Tensor, return_dict: bool = True)

Source

Parameters:

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.

Decode a batch of images using a tiled decoder.

tiled_encode[[diffusers.AutoencoderKLQwenImage21.tiled_encode]]

tiled_encode(x: Tensor)

Source

Parameters:

x (torch.Tensor) : Input batch of videos.

Returns: torch.Tensor

The latent representation of the encoded videos.

Encode a batch of images using a tiled encoder.

AutoencoderKLOutput[[diffusers.models.modeling_outputs.AutoencoderKLOutput]]

diffusers.models.modeling_outputs.AutoencoderKLOutput[[diffusers.models.modeling_outputs.AutoencoderKLOutput]]

diffusers.models.modeling_outputs.AutoencoderKLOutput(latent_dist: DiagonalGaussianDistribution)

Source

Parameters:

latent_dist (DiagonalGaussianDistribution) : Encoded outputs of Encoder represented as the mean and logvar of DiagonalGaussianDistribution. DiagonalGaussianDistribution allows for sampling latents from the distribution.

Output of AutoencoderKL encoding method.

DecoderOutput[[diffusers.models.autoencoders.vae.DecoderOutput]]

diffusers.models.autoencoders.vae.DecoderOutput[[diffusers.models.autoencoders.vae.DecoderOutput]]

diffusers.models.autoencoders.vae.DecoderOutput(sample: Tensor, commit_loss: typing.Optional[torch.FloatTensor] = None)

Source

Parameters:

sample (torch.Tensor of shape (batch_size, num_channels, height, width)) : The decoded output sample from the last layer of the model.

Output of decoding method.