Kaggle中使用SDXL

最新推荐文章于 2024-09-11 22:33:50 发布

Covirtue

最新推荐文章于 2024-09-11 22:33:50 发布

阅读量396

点赞数 8

分类专栏：人工智能文章标签：深度学习人工智能

本文链接：https://blog.csdn.net/weixin_47695827/article/details/136724009

版权

人工智能专栏收录该内容

15 篇文章 0 订阅

订阅专栏

Kaggle中使用SDXL方法

如图先搜索sdxl模型，然后创建代码code或者直接使用他人的代码

在这里插入图片描述

安装依赖库

!pip install diffusers --upgrade
!pip install invisible_watermark transformers accelerate safetensors

查看依赖库版本

import diffusers
diffusers.__version__

# 获取对象帮助信息。ps：一切皆可help（有时候不知道参数的情况下）
from diffusers import DiffusionPipeline
help(DiffusionPipeline.from_pretrained)

一、文生图

# 加载模型
from diffusers import DiffusionPipeline
import torch

pipe = DiffusionPipeline.from_pretrained(
    "/kaggle/input/stable-diffusion-xl/pytorch/base-1-0/1", 
    torch_dtype=torch.float16, 
    use_safetensors=True, 
    variant="fp16"
)
pipe.to("cuda")
# pipe.enable_model_cpu_offload() # 允许将模型分担到cpu

# if using torch < 2.0
# pipe.enable_xformers_memory_efficient_attention()

# 开始生成
prompt = "Imagine a charming Korean woman, HanFu."
image = pipe(
              prompt=prompt,
              width=1024,
              height=1024,
              num_inference_steps=50
             ).images[0]

# 结果展示
image

二、调用refiner进一步处理图片

# 加载模型权重
from diffusers import (
        DiffusionPipeline,
        StableDiffusionXLPipeline,
        StableDiffusionXLImg2ImgPipeline
    )
import torch


# load both base & refiner
base = StableDiffusionXLPipeline.from_pretrained(
    "/kaggle/input/stable-diffusion-xl/pytorch/base-1-0/1", 
    torch_dtype=torch.float16, 
    variant="fp16", 
    use_safetensors=True
)
base.to("cuda")
# base.enable_model_cpu_offload()

# refiner = DiffusionPipeline.from_pretrained(
#     "/kaggle/input/stable-diffusion-xl/pytorch/base-1-0/1",
#     text_encoder_2=base.text_encoder_2,
#     vae=base.vae,
#     torch_dtype=torch.float16,
#     use_safetensors=True,
#     variant="fp16",
# ) # 等效下方代码

refiner=StableDiffusionXLImg2ImgPipeline(**base.components) # 共用权重节省内存
refiner.text_encoder_2=base.text_encoder_2
refiner.to("cuda")
# refiner.enable_model_cpu_offload()

# 生成图片
# Define how many steps and what % of steps to be run on each experts (80/20) here
n_steps = 40
high_noise_frac = 0.8

prompt = "A charming Korean woman wearing pink dress"

# run both experts
image = base(
    prompt=prompt,
    num_inference_steps=n_steps,
    denoising_end=high_noise_frac,
    output_type="latent",
).images[0]
image = refiner(
    prompt=prompt,
    num_inference_steps=n_steps,
    denoising_start=high_noise_frac,
    image=image,
).images[0]

# 查看结果
image

三、图生图

# 加载模型
from diffusers.utils import load_image
from diffusers import StableDiffusionXLImg2ImgPipeline
import torch

# load both base & refiner
refiner = StableDiffusionXLImg2ImgPipeline.from_pretrained(
    "/kaggle/input/stable-diffusion-xl/pytorch/base-1-0/1", 
    torch_dtype=torch.float16, 
    variant="fp16", 
    use_safetensors=True
)
refiner.to("cuda")

image = load_image("/kaggle/input/examples/download1.png")
prompt="a pretty Korean woman wearing green dress"
num_inference_steps=50
strength=0.8
guidance_scale=7

image=refiner(
    prompt=prompt,
    num_inference_steps=num_inference_steps,
    strength=strength,
    guidance_scale=guidance_scale,
    image=image
).images[0]