tglcourse/latent_lsun_church_256px

---
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 4x32x32 latent representation using the same VAE as that employed by Stable Diffusion

Decoding
```python
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

```

原始信息汇总

数据集概述

数据集名称

• 名称: latent_lsun_church_256px

数据集特征

• 特征1: label

  • 数据类型: 类别标签
  • 类别名称:
    • 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

• 特征2: latent

  • 数据类型: 序列，序列类型为float32

数据集分割

• 分割1: test

  • 数据大小: 106824288字节
  • 样本数量: 6312

• 分割2: train

  • 数据大小: 2029441460字节
  • 样本数量: 119915

数据集大小

• 下载大小: 2082210019字节
• 数据集总大小: 2136265748字节