metadata

dataset_info:
  features:
    - name: label
      dtype:
        class_label:
          names:
            '0': '0'
            '1': '1'
            '2': '2'
            '3': '3'
            '4': '4'
            '5': '5'
            '6': '6'
            '7': '7'
            '8': '8'
            '9': '9'
            '10': a
            '11': b
            '12': c
            '13': d
            '14': e
            '15': f
    - name: latent
      sequence:
        sequence:
          sequence: float32
  splits:
    - name: test
      num_bytes: 106824288
      num_examples: 6312
    - name: train
      num_bytes: 2029441460
      num_examples: 119915
  download_size: 2082210019
  dataset_size: 2136265748

Dataset Card for "latent_lsun_church_256px"

This is derived from https://huggingface.co/datasets/tglcourse/lsun_church_train

Each image is cropped to 256px square and encoded to a 32x32x4 latent representation using the same VAE as that employed by Stable Diffusion

Decoding

from diffusers import AutoencoderKL
from datasets import load_dataset
from PIL import Image
import numpy as np
import torch

# load the dataset
dataset = load_dataset('tglcourse/latent_lsun_church_256px')

# Load the VAE (requires access - see repo model card for info)
vae = AutoencoderKL.from_pretrained("CompVis/stable-diffusion-v1-4", subfolder="vae")

latent = torch.tensor([dataset['train'][0]['latent']]) # To tensor (bs, 4, 32, 32)
latent = (1 / 0.18215) * latent # Scale to match SD implementation
with torch.no_grad():
    image = vae.decode(latent).sample[0] # Decode 
image = (image / 2 + 0.5).clamp(0, 1) # To (0, 1)
image = image.detach().cpu().permute(1, 2, 0).numpy() # To numpy, channels lsat
image = (image * 255).round().astype("uint8") # (0, 255) and type uint8
image = Image.fromarray(image) # To PIL
image # The resulting PIL image