扩散模型DDPM代码实践

安装diffusers

pip install diffusers

按照diffusers官方代码

from diffusers import DDPMPipelinepipe = DDPMPipeline.from_pretrained("google/ddpm-cat-256")image = pipe().images[0]image.save("/data/zhz/projects/diffusion/output/ddpm_generated_image.png")

准备 google/ddpm-cat-256/

因为服务器连不了外网所以只能手动下载：

https://huggingface.co/google/ddpm-cat-256/tree/main

文件结构：

同时代码也要修改为：

from diffusers import DDPMPipeline# pipe = DDPMPipeline.from_pretrained("google/ddpm-cat-256")
pipe = DDPMPipeline.from_pretrained("/data/zhz/projects/diffusion/ddpm-cat-256")image = pipe().images[0]image.save("/data/zhz/projects/diffusion/output/ddpm_generated_image.png")

运行

python /data/zhz/projects/diffusion/test_20240929.py

等待进度条

如果是换成gpu的话：

from diffusers import DDPMPipeline# pipe = DDPMPipeline.from_pretrained("google/ddpm-cat-256")
pipe = DDPMPipeline.from_pretrained("/data/zhz/projects/diffusion/ddpm-cat-256").to("cuda")image = pipe().images[0]image.save("/data/zhz/projects/diffusion/output/ddpm_generated_image.png")

显存开销3G左右：

最后生成图片：