File size: 3,051 Bytes
8e01959 dd52bb4 8e01959 dd52bb4 8e01959 dd52bb4 8ede60a dd52bb4 0cc37b4 dd52bb4 0cc37b4 dd52bb4 4133b15 dd52bb4 3b96387 dd52bb4 4133b15 dd52bb4 0cc37b4 dd52bb4 4133b15 dd52bb4 e17c029 8ede60a e17c029 dd52bb4 8ede60a dd52bb4 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 |
---
language: zh
license: creativeml-openrail-m
tags:
- diffusion
- zh
- Chinese
---
# Chinese-Style-Stable-Diffusion-2-v0.1
|  |  |  |
| ------------------------------------- | ----------------------------- | ------------------------------- |
|  |  |  |
|  |  |  |
大概是Huggingface 🤗社区首个开源的Stable diffusion 2 中文模型。该模型基于[stable diffusion V2.1](https://huggingface.co/stabilityai/stable-diffusion-2-1)模型，在约500万条的中国风格筛选过的中文数据上进行微调，数据来源于多个开源数据集如[LAION-5B](https://laion.ai/blog/laion-5b/), [Noah-Wukong](https://wukong-dataset.github.io/wukong-dataset/), [Zero](https://zero.so.com/)和一些网络数据。
Probably the first open sourced Chinese Stable Diffusion 2 model in Huggingface🤗 community. This model is finetuned based on [stable diffusion V2.1](https://huggingface.co/stabilityai/stable-diffusion-2-1) with 5M chinese style filtered data. Dataset is composed of several different chinese open source dataset such as [LAION-5B](https://laion.ai/blog/laion-5b/), [Noah-Wukong](https://wukong-dataset.github.io/wukong-dataset/), [Zero](https://zero.so.com/) and some web data.
# <u>Model Details</u>
#### Text Encoder
文本编码器使用冻结参数的[lyua1225/clip-huge-zh-75k-steps-bs4096](https://huggingface.co/lyua1225/clip-huge-zh-75k-steps-bs4096)。
Text encoder is frozen [lyua1225/clip-huge-zh-75k-steps-bs4096](https://huggingface.co/lyua1225/clip-huge-zh-75k-steps-bs4096) .
#### Unet
在筛选过的的500万中文数据集上训练了150K steps，使用指数移动平均值(EMA)做原绘画能力保留，使模型能够在中文风格和原绘画能力之间获得权衡。
Training on 5M chinese style filtered data for 150k steps. Exponential moving average(EMA) is applied to keep the original Stable Diffusion 2 drawing capability and reach a balance between chinese style and original drawing capability.
## <u>Usage</u>
因为使用了customized tokenizer, 所以需要优先加载一下tokenizer, 并传入trust_remote_code=True
Customized Tokenizer should be loaded first with 'trust_remote_code=True'.
```py
import torch
from diffusers import StableDiffusionPipeline
from transformers import AutoTokenizer
tokenizer_id = "lyua1225/clip-huge-zh-75k-steps-bs4096"
sd2_id = "Midu/chinese-style-stable-diffusion-2-v0.1"
tokenizer = AutoTokenizer.from_pretrained(tokenizer_id, trust_remote_code=True)
pipe = StableDiffusionPipeline.from_pretrained(sd2_id, torch_dtype=torch.float16, tokenizer=tokenizer)
pipe.to("cuda")
image = pipe("赛博朋克风格的城市街道，8K分辨率，CG渲染", guidance_scale=10, num_inference_steps=20).images[0]
image.save("cyberpunk.jpeg")
```
|