一个项目体验stable-diffusion、waifu、NovelAI三大模型

#下面的只能生成512 x 1024 的图，再大像素会报错 ，第一次运行后记得重启  ，第一次运行后记得重启 ，第一次运行后记得重启 import osfrom IPython.display import clear_outputfrom utils import check_is_model_completeprint('正在解压模型')if not check_is_model_complete("./NovelAI_latest_ab21ba3c_paddle"): # 解压 latest 模型
    !unzip -o "/home/aistudio/data/data171442/NovelAI_latest_ab21ba3c_paddle.zip" # if not check_is_model_complete("./model_pruned_paddle"): # full 模型#     !unzip -o "/home/aistudio/data/data171442/NovelAI_pruned_paddle.zip"if not os.path.exists("diffusers_paddle"): 
    !unzip -o diffusers_paddle.zipprint('正在安装库')
!pip install --upgrade pip  -i https://mirror.baidu.com/pypi/simple
!pip install -U fastcore paddlenlp ftfy regex --user  -i https://mirror.baidu.com/pypi/simple
!pip install --upgrade paddlenlp  -i https://mirror.baidu.com/pypi/simpleprint('正在安装新版飞桨')# 如果需要生成 768 x 1024 则去掉下两行开头的井号运行一次 (运行成功一次即可), 但是可能卡, 建议挑人少的时候#!mv /home/aistudio/data/data168051/paddlepaddle-develop.whl /home/aistudio/data/data168051/paddlepaddle_gpu-0.0.0.post112-cp37-cp37m-linux_x86_64.whl#!pip install "/home/aistudio/data/data168051/paddlepaddle_gpu-0.0.0.post112-cp37-cp37m-linux_x86_64.whl" --userclear_output() # 清理很长的内容print('加载完毕, 下次不用再运行这里了')

# 下载及解压SD预训练模型及waifu模型 ，可能需要比较长的时间import osif not os.path.exists("CompVis-stable-diffusion-v1-4"):
    !wget https://paddlenlp.bj.bcebos.com/models/community/CompVis/stable-diffusion-v1-4/CompVis-stable-diffusion-v1-4.tar.gz
    !tar -zxvf CompVis-stable-diffusion-v1-4.tar.gz
    !rm -rf CompVis-stable-diffusion-v1-4.tar.gzif not os.path.exists("waifu-diffusion"):
    !wget http://bj.bcebos.com/paddlenlp/models/community/hakurei/waifu-diffusion/waifu-diffusion.tar.gz #waifu-diffusion-v1-2
    !tar -zxvf waifu-diffusion.tar.gz
    !rm -rf waifu-diffusion.tar.gz    #!wget https://paddlenlp.bj.bcebos.com/models/community/hakurei/waifu-diffusion/waifu-diffusion-v1-3.tar.gz #waifu-diffusion-v1-3 注意下面移动和解压
    #!tar -zxvf waifu-diffusion-v1-3.tar.gz
    #!rm -rf waifu-diffusion-v1-3.tar.gz   
    !unzip -o "data/data168051/stable-diffusion-v1-4.zip" 
    !unzip -o "sd-concepts-library.zip"# if not os.path.exists("trinart_stable_diffusion_v2_95k"):#     !wget http://bj.bcebos.com/paddlenlp/models/community/naclbit/trinart_stable_diffusion_v2/trinart_stable_diffusion_v2_95k.tar.gz#     !tar -zxvf trinart_stable_diffusion_v2_95k.tar.gz#     !rm -rf trinart_stable_diffusion_v2_95k.tar.gz

# 模型和权重加载都在这里,pipe 加载的权重,vae_path 加载的本体,如果第一次运行下面报错, 就重启内核重新运行一次from diffusers_paddle import StableDiffusionPipeline, StableDiffusionImg2ImgPipelineimport paddlefrom utils import save_image_infoimport os
pipe = StableDiffusionPipeline.from_pretrained("./CompVis-stable-diffusion-v1-4") # 加载SD默认权重#pipe = StableDiffusionPipeline.from_pretrained("./waifu-diffusion") # 加载waifu权重#pipe = StableDiffusionPipeline.from_pretrained("./NovelAI_latest_ab21ba3c_paddle") # 加载NovelAI权重# pipe = StableDiffusionPipeline.from_pretrained("./model_pruned_paddle")vae_path = 'stable-diffusion-v1-4/model_state.pdparams'  # 加载 SD默认  模型#vae_path = 'waifu-diffusion/unet/model_state.pdparams'   # 加载 waifu  模型#vae_path = './NovelAI_latest_ab21ba3c_paddle/vae/animevae.pdparams' # 加载 NovelAI vae 模型pipe.vae.load_state_dict(paddle.load(vae_path)) # 换用更好的 vae (有效果!)# 图生图pipe_i2i = StableDiffusionImg2ImgPipeline(vae=pipe.vae,text_encoder=pipe.text_encoder,tokenizer=pipe.tokenizer,
unet=pipe.unet,scheduler=pipe.scheduler,safety_checker=pipe.safety_checker,feature_extractor=pipe.feature_extractor)print('加载完毕')

from ui import gui_txt2img # 文生图display(gui_txt2img.gui) # 生成的图片自动保存到左侧的 outputs 的文件夹里

from ui import gui_img2img # 图生图, 在左侧上传图片, 然后修改 "需要转换的图片路径"display(gui_img2img.gui) # 生成的图片自动保存到左侧的 outputs 的文件夹里

from ui import gui_superres # 超分 (图片放大一倍), 在左侧上传图片, 然后修改 "需要超分的图片路径"display(gui_superres.gui) # 生成的图片自动保存到左侧的 outputs 的文件夹里

height = 512width  = 768seed   = 3354985548steps  = 50cfg    = 7    prompt = "miku, looking at viewer, long hair, standing, 1girl, hair ornament, hair flower, cute, jacket, white flower, white dress 精灵耳朵"#negative_prompt = "{mucha},{extremely detailed 4K CG},{original}, {{masterpiece}},lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, \#cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry"image = pipe(prompt,height=height,width=width,num_inference_steps=steps,guidance_scale=cfg,seed=seed,negative_prompt=negative_prompt).images[0]print('Seed =', image.argument['seed'])
display(image)
save_image_info(image, path = './jpgout/') # 保存图片到指定位置

image.argument # 查看图片生成使用的参数

# 批量文生图number = 2height = 768 width  = 512steps  = 50cfg    = 17.5  prompt = "kitsune made out of flames, digital art, synthwave  . "negative_prompt = "{mucha},{extremely detailed 4K CG},{original}, {{masterpiece}},lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry"for i in range(number):
    image = pipe(prompt,height=height,width=width,num_inference_steps=steps,guidance_scale=cfg,negative_prompt=negative_prompt).images[0]
    save_image_info(image, path = './jpgout/')
    display(image)

# 批量文生图number = 2height = 768 width  = 512steps  = 50cfg    = 17.5  prompt = "kitsune made out of flames, digital art, synthwave  . "negative_prompt = "lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry"for i in range(number):
    image = pipe(prompt,height=height,width=width,num_inference_steps=steps,guidance_scale=cfg,negative_prompt=negative_prompt).images[0]
    save_image_info(image, path = './jpgout/')
    display(image)

path     = "image_Kurisu.png"steps    = 50seed     = 20221008strength = 0.8cfg      = 7.5prompt = "kitsune made out of flames, digital art, synthwave "negative_prompt = "lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry"
    from utils import ReadImage
init_image = ReadImage(path, height=-1, width=-1) # -1 为自动判断图片大小image2 = pipe_i2i(prompt,init_image=init_image,num_inference_steps=steps,strength=strength,guidance_scale=cfg,seed=seed,negative_prompt=negative_prompt)[0][0]print('Seed =', image2.argument['seed'])
display(image2)
save_image_info(image2, path = './outputs/') # 保存图片到指定位置

image2.argument # 查看图片生成使用的参数

number   = 10path     = "https://img.php.cn/upload/article/001/571/248/175306710847236.jpg" #替换了一个网络图片steps    = 50strength = 0.8cfg      = 13.5prompt = "miku is Sitting on a wooden chair with porcelain on a wooden multi-purpose shelf behind her while With incense burner at the top of the shelf, looking at viewer, long hair, standing, 1girl, hair ornament, hair flower, cute, jacket, white flower, white dress"negative_prompt = "lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry"
    from utils import ReadImage
init_image = ReadImage(path, height=-1, width=-1) # -1 为自动判断图片大小for i in range(number):
    image2 = pipe_i2i(prompt,init_image=init_image,num_inference_steps=steps,strength=strength,guidance_scale=cfg,negative_prompt=negative_prompt)[0][0]
    save_image_info(image2, path = './outputs/')
    display(image2)

learned_embeded_path = "sd-concepts-library/Marc_Allante-style.pdparams"for token, embeds in paddle.load(learned_embeded_path).items():
    pipe.tokenizer.add_tokens(token)
    pipe.text_encoder.resize_token_embeddings(len(pipe.tokenizer))
    token_id = pipe.tokenizer.convert_tokens_to_ids(token)    with paddle.no_grad():
        pipe.text_encoder.get_input_embeddings().weight[token_id] = embeds

prompt = f"a colorful color painting of {token} styles ,kitsune made out of flames, digital art, synthwave.  "  #如果不是图片风格是物体 使用a photo of a {token}# negative_prompt = "lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry"num_samples = 1 #图片数量 images = pipe([prompt] * num_samples, height=704, width=512, num_inference_steps=50 , guidance_scale=7.5,seed=1788937216).imagesfor img in images:
    display(img)
    save_image_info(img, path = './learnout/')

learned_embeded_path = "sd-concepts-library/hitokomoru-style.pdparams"for token, embeds in paddle.load(learned_embeded_path).items():
    pipe.tokenizer.add_tokens(token)
    pipe.text_encoder.resize_token_embeddings(len(pipe.tokenizer))
    token_id = pipe.tokenizer.convert_tokens_to_ids(token)    with paddle.no_grad():
        pipe.text_encoder.get_input_embeddings().weight[token_id] = embeds

prompt = f"a colorful color painting of {token} styles ,kitsune made out of flames, digital art, synthwave.  "# negative_prompt = "lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry"num_samples = 1 #图片数量 images = pipe([prompt] * num_samples, height=704, width=512, num_inference_steps=50, guidance_scale=17.5,seed=2788937216).imagesfor img in images:
    display(img)

# 加载权重文件pipe = StableDiffusionPipeline.from_pretrained("./waifu-diffusion")#pipe = StableDiffusionPipeline.from_pretrained("./CompVis-stable-diffusion-v1-4")#pipe = StableDiffusionPipeline.from_pretrained("./NovelAI_latest_ab21ba3c_paddle")# 加载sd-concepts-library中的训练模型 learned_embeded_path = "sd-concepts-library/huang-guang-jian-style.pdparams"

# 自训练模型生成多张图片for token, embeds in paddle.load(learned_embeded_path).items():
    pipe.tokenizer.add_tokens(token)
    pipe.text_encoder.resize_token_embeddings(len(pipe.tokenizer))
    token_id = pipe.tokenizer.convert_tokens_to_ids(token)    with paddle.no_grad():
        pipe.text_encoder.get_input_embeddings().weight[token_id] = embedsimport time
timestring = time.strftime('%Y%m%d%H%M%S') #a colorful painting of {token} styles,prompt = f"a colorful color painting of {token} styles ,kitsune made out of flames, digital art, synthwave.  "negative_prompt = "lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry"numg = 50 #图片数量 num_samples = 1for _ in range(numg):    with paddle.no_grad():
                image = pipe([prompt] * num_samples,num_inference_steps=100,width=1024,height=512,negative_prompt=[negative_prompt],guidance_scale=13.5).images[0]
                timestring = time.strftime('%Y%m%d%H%M%S')
                image.save(f"./learnout/{timestring}_.png")                #               image.save(f"./jpg/{timestring}.png")
    with open(f'./learnout/{timestring}.txt', 'w') as f: 
                print(image.argument,file = f)
    display(image.argument)     
    display(image)

# 清理文件夹使用 确保已经下载文件#删除文件夹下面的所有文件(只删除文件,不删除文件夹)import osimport shutil#python删除文件的方法 os.remove(path)path指的是文件的绝对路径,如：def del_file(path_data):
    for i in os.listdir(path_data) :# os.listdir(path_data)#返回一个列表，里面是当前目录下面的所有东西的相对路径
        file_data = path_data + "/" + i#当前文件夹的下面的所有东西的绝对路径
        if os.path.isfile(file_data) == True:#os.path.isfile判断是否为文件,如果是文件,就删除.如果是文件夹.递归给del_file.
            os.remove(file_data)        else:
            del_file(file_data)#path_data = r"jpgout" # 删除TXT 生成的文件path_data = r"outputs" #删除图生图 生成的文件#path_data = r"learnout" #删除自训练文生图 生成的文件del_file(path_data)