Tiny AutoEncoder
Tiny AutoEncoder for Stable Diffusion (TAESD) was introduced in madebyollin/taesd by Ollin Boer Bohan. It is a tiny distilled version of Stable Diffusion’s VAE that can quickly decode the latents in a StableDiffusionPipeline or StableDiffusionXLPipeline almost instantly.
To use with Stable Diffusion v-2.1:
import torch
from diffusers import DiffusionPipeline, AutoencoderTiny
pipe = DiffusionPipeline.from_pretrained(
"stabilityai/stable-diffusion-2-1-base", torch_dtype=torch.float16
)
pipe.vae = AutoencoderTiny.from_pretrained("madebyollin/taesd", torch_dtype=torch.float16)
pipe = pipe.to("cuda")
prompt = "slice of delicious New York-style berry cheesecake"
image = pipe(prompt, num_inference_steps=25).images[0]
image
To use with Stable Diffusion XL 1.0
import torch
from diffusers import DiffusionPipeline, AutoencoderTiny
pipe = DiffusionPipeline.from_pretrained(
"stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
)
pipe.vae = AutoencoderTiny.from_pretrained("madebyollin/taesdxl", torch_dtype=torch.float16)
pipe = pipe.to("cuda")
prompt = "slice of delicious New York-style berry cheesecake"
image = pipe(prompt, num_inference_steps=25).images[0]
image
AutoencoderTiny
class diffusers.AutoencoderTiny
< source >( in_channels: int = 3 out_channels: int = 3 encoder_block_out_channels: Tuple = (64, 64, 64, 64) decoder_block_out_channels: Tuple = (64, 64, 64, 64) act_fn: str = 'relu' upsample_fn: str = 'nearest' latent_channels: int = 4 upsampling_scaling_factor: int = 2 num_encoder_blocks: Tuple = (1, 3, 3, 3) num_decoder_blocks: Tuple = (3, 3, 3, 1) latent_magnitude: int = 3 latent_shift: float = 0.5 force_upcast: bool = False scaling_factor: float = 1.0 shift_factor: float = 0.0 )
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. - encoder_block_out_channels (
Tuple[int]
, optional, defaults to(64, 64, 64, 64)
) — Tuple of integers representing the number of output channels for each encoder block. The length of the tuple should be equal to the number of encoder blocks. - decoder_block_out_channels (
Tuple[int]
, optional, defaults to(64, 64, 64, 64)
) — Tuple of integers representing the number of output channels for each decoder block. The length of the tuple should be equal to the number of decoder blocks. - act_fn (
str
, optional, defaults to"relu"
) — Activation function to be used throughout the model. - latent_channels (
int
, optional, defaults to 4) — Number of channels in the latent representation. The latent space acts as a compressed representation of the input image. - upsampling_scaling_factor (
int
, optional, defaults to 2) — Scaling factor for upsampling in the decoder. It determines the size of the output image during the upsampling process. - num_encoder_blocks (
Tuple[int]
, optional, defaults to(1, 3, 3, 3)
) — Tuple of integers representing the number of encoder blocks at each stage of the encoding process. The length of the tuple should be equal to the number of stages in the encoder. Each stage has a different number of encoder blocks. - num_decoder_blocks (
Tuple[int]
, optional, defaults to(3, 3, 3, 1)
) — Tuple of integers representing the number of decoder blocks at each stage of the decoding process. The length of the tuple should be equal to the number of stages in the decoder. Each stage has a different number of decoder blocks. - latent_magnitude (
float
, optional, defaults to 3.0) — Magnitude of the latent representation. This parameter scales the latent representation values to control the extent of information preservation. - latent_shift (float, optional, defaults to 0.5) — Shift applied to the latent representation. This parameter controls the center of the latent space.
- scaling_factor (
float
, optional, defaults to 1.0) — The component-wise standard deviation of the trained latent space computed using the first batch of the training set. This is used to scale the latent space to have unit variance when training the diffusion model. The latents are scaled with the formulaz = z * scaling_factor
before being passed to the diffusion model. When decoding, the latents are scaled back to the original scale with the formula:z = 1 / scaling_factor * z
. For more details, refer to sections 4.3.2 and D.1 of the High-Resolution Image Synthesis with Latent Diffusion Models paper. For this Autoencoder, however, no such scaling factor was used, hence the value of 1.0 as the default. - force_upcast (
bool
, optional, default toFalse
) — 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 caseforce_upcast
can be set toFalse
(see this fp16-friendly AutoEncoder).
A tiny distilled VAE model for encoding images into latents and decoding latent representations into images.
AutoencoderTiny is a wrapper around the original implementation of TAESD
.
This model inherits from ModelMixin. Check the superclass documentation for its generic methods implemented for all models (such as downloading or saving).
Disable sliced VAE decoding. If enable_slicing
was previously enabled, this method will go back to computing
decoding in one step.
Disable tiled VAE decoding. If enable_tiling
was previously enabled, this method will go back to computing
decoding in one step.
Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
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
< source >( sample: Tensor return_dict: bool = True )
raw latents -> [0, 1]
[0, 1] -> raw latents
AutoencoderTinyOutput
class diffusers.models.autoencoders.autoencoder_tiny.AutoencoderTinyOutput
< source >( latents: Tensor )
Output of AutoencoderTiny encoding method.