一，Stable Diffusion模型原理

目前开源的最火爆的AI绘画系统是 Stable Diffusion 模型（稳定扩散模型）。

可以完成 text2img, img2img, depth2img【深度图转图像】, seg2img【语义分割图转图像】 等基于提示信息【prompt】的图画创作功能。

其核心原理简要总结如下：

1，通过Attention机制引入text /semantic_map/ input_image ...等控制信息。--> 构图

2，在Attention控制下通过UNet模型在隐空间通过反向扩散机制从初始的噪声中一步一步(通常20到50步)清洗噪声生成隐空间图片信息(Latent Diffusion Model过程)。--> 底片

3，最后通过 AutoDecoder 从隐空间的图片信息还原成高清图片。--> 成片

从易到难的详细原理参考：

《零基础读懂Stable Diffusion》 https://zhuanlan.zhihu.com/p/597247221 （zero mathematics）

《Stable Diffusion原理解读》 https://zhuanlan.zhihu.com/p/583124756 （some mathematics）

《当我们谈论Text-to-image:Stable Model》https://zhuanlan.zhihu.com/p/550967884 （heavy mathematics）

公众号后台回复关键词：StableDiffusion 获取本文源代码和B站视频演示。

二，在线体验方法

英文原始: https://huggingface.co/spaces/runwayml/stable-diffusion-v1-5

百度中文：https://wenxin.baidu.com/ernie-vilg?curtype=exp

太乙中文：https://huggingface.co/spaces/IDEA-CCNL/Taiyi-Stable-Diffusion-Chinese-Webui

各种风格：https://civitai.com/

三，在Kaggle中体验

!pip install diffusers
!pip install transformers

1，text2img

from PIL import Image
import torchfrom diffusers import (StableDiffusionPipeline,StableDiffusionImg2ImgPipeline,StableDiffusionInpaintPipeline
)device=torch.device('cuda' if torch.cuda.is_available() else 'cpu')#model_id = "IDEA-CCNL/Taiyi-Stable-Diffusion-1B-Chinese-v0.1"
#model_id = "IDEA-CCNL/Taiyi-Stable-Diffusion-1B-Chinese-EN-v0.1"
#model_id = "IDEA-CCNL/Taiyi-Stable-Diffusion-1B-Anime-Chinese-v0.1" #漫画头像
#model_id = "CompVis/stable-diffusion-v1-4"
#model_id = "runwayml/stable-diffusion-v1-5"
model_id = "stabilityai/stable-diffusion-2-1"pipe_text2img = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16).to(device)

width = 960
height = 680
guide = 7
steps = 20prompt = "a moutain full of flowers and snows"
#prompt = "Bejing Autumn"
#prompt = 'shanghai night'
#prompt = 'a dog is runing after a mouse'
#prompt = 'a little girl is smiling with a cat'negative_prompt = None
output = pipe_text2img(prompt = prompt, negative_prompt=negative_prompt,width=width, height=height,guidance_scale=guide, num_inference_steps=steps)
out_img = output.images[0]
out_img

2，img2img

from PIL import Image
import torchfrom diffusers import (StableDiffusionPipeline,StableDiffusionInpaintPipeline
)device=torch.device('cuda' if torch.cuda.is_available() else 'cpu')#model_id = "IDEA-CCNL/Taiyi-Stable-Diffusion-1B-Chinese-v0.1"
#model_id = "IDEA-CCNL/Taiyi-Stable-Diffusion-1B-Chinese-EN-v0.1"
#model_id = "IDEA-CCNL/Taiyi-Stable-Diffusion-1B-Anime-Chinese-v0.1"
#model_id = "CompVis/stable-diffusion-v1-4"
#model_id = "runwayml/stable-diffusion-v1-5"
model_id = "stabilityai/stable-diffusion-2-1"pipe_img2img = StableDiffusionImg2ImgPipeline.from_pretrained(model_id, torch_dtype=torch.float16).to(device)

strength = 0.5
revise_prompt = "there is a water fall"
revised_output = pipe_img2img(revise_prompt, image=out_img, strength=strength, guidance_scale=guide, num_inference_steps=steps)
revised_output.images[0]

3，pix2pix

import PIL
import requests
import torch
from diffusers import StableDiffusionInstructPix2PixPipeline, EulerAncestralDiscreteSchedulermodel_id = "timbrooks/instruct-pix2pix"
pipe = StableDiffusionInstructPix2PixPipeline.from_pretrained(model_id, torch_dtype=torch.float16, safety_checker=None)
pipe.to("cuda")
pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe.scheduler.config)def download_image(url):image = PIL.Image.open(requests.get(url, stream=True).raw)image = PIL.ImageOps.exif_transpose(image)image = image.convert("RGB")return image
url = "https://raw.githubusercontent.com/timothybrooks/instruct-pix2pix/main/imgs/example.jpg"image = download_image(url)

image

#prompt = "turn him into cyborg"
#prompt = "turn him into oil painting"
#prompt = "let he in the wild, green grass"
#prompt = "a flower in his left hand"
#prompt = "let he wear a coat"
prompt = "let he wear a cap on the head"
images = pipe(prompt, image=image, num_inference_steps=10, image_guidance_scale=1.0).images
images[0]

公众号后台回复关键词：StableDiffusion 获取本文源代码和B站视频演示。