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

```

数据集信息：
特征字段：
- 名称：label
数据类型：
分类标签（class_label）：
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
- 名称：隐变量（latent）
序列结构：
序列结构：
序列结构：32位浮点型（float32）
数据划分：
- 名称：test
字节数：106824288
样本数：6312
- 名称：train
字节数：2029441460
样本数：119915
下载大小：2082210019
数据集总大小：2136265748

---
# 「latent_lsun_church_256px」数据集卡片
本数据集衍生自 https://huggingface.co/datasets/tglcourse/lsun_church_train

每张图像均被裁剪为256像素的正方形，并使用与Stable Diffusion同款的变分自编码器（Variational Autoencoder, VAE）编码为4×32×32的隐变量表示。

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

# 加载数据集
dataset = load_dataset('tglcourse/latent_lsun_church_256px')

# 加载变分自编码器（VAE）（需获取访问权限——详见仓库模型卡片说明）
vae = AutoencoderKL.from_pretrained("CompVis/stable-diffusion-v1-4", subfolder="vae")

latent = torch.tensor([dataset['train'][0]['latent']]) # 转换为张量（批量大小, 4, 32, 32）
latent = (1 / 0.18215) * latent # 缩放以匹配Stable Diffusion的实现逻辑
with torch.no_grad():
image = vae.decode(latent).sample[0] # 解码
image = (image / 2 + 0.5).clamp(0, 1) # 归一化至(0,1)区间
image = image.detach().cpu().permute(1, 2, 0).numpy() # 转换为numpy数组，通道置于最后一维
image = (image * 255).round().astype("uint8") # 映射至(0,255)区间并转换为uint8类型
image = Image.fromarray(image) # 转换为PIL图像对象
image # 生成的PIL图像

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