Pass-O-Guava / StableDiffusionUniPipeline

Stable diffusion txt2img and img2img pipeline merging and GPU parallel acceleration based on diffusers

StableDiffusionUniPipeline

👏Thanks for the work of https://github.com/hellojixian/StableDiffusionParallelPipeline

Introduction

1. This code is mainly based on the diffusers Python library 🤗
2. Merging the txt2img(StableDiffusionPipeline) and img2img(StableDiffusionImg2ImgPipeline) in one pipeline.
3. Support Double-GPUs and Single-GPU parallel acceleration

How it works

1. Prepare

# 1. Install diffusers
pip install diffusers==0.20.2

# 2. Move .py to diffusers install path
mv pipeline_stable_diffusion_uni.py xxx/diffusers/pipelines/stable_diffusion/
mv pipeline_stable_diffusion_uni_parallel.py xxx/diffusers/pipelines/stable_diffusion/

## 3. Edit __init__.py
vim xxx/diffusers/pipelines/stable_diffusion/__init__.py
#line:210 + StableDiffusionUniPipeline
#line:211 + StableDiffusionUniParallelPipeline

vim xxx/diffusers/pipelines/stable_diffusion/pipelines/__init__.py
#line:110 + StableDiffusionUniPipeline
#line:111 + StableDiffusionUniParallelPipeline

vim xxx/diffusers/pipelines/stable_diffusion/pipelines/stable_diffusion/__init__.py
#line64 + from .pipeline_stable_diffusion_uni import StableDiffusionUniPipeline
#line64 + from .pipeline_stable_diffusion_uni_parallel import StableDiffusionUniParallelPipeline

2. Use

🚀 Run benchmark_uni.ipynb

StableDiffusionUniPipeline

import torch
from PIL import Image
from diffusers import StableDiffusionUniPipeline

prompt = "a photo of an astronaut riding a horse on mars"
input_imge = Image.open("./input.png")

pipe = StableDiffusionUniPipeline.from_pretrained("runwayml/stable-diffusion-v1-5").to("cuda")
        
# txt2img
image = pipe(prompt).images[0]
        
# img2img
image = pipe(prompt, input_imge).images[0]

StableDiffusionUniParallelPipeline

import torch
from PIL import Image
from diffusers import StableDiffusionUniParallelPipeline


prompt = "a photo of an astronaut riding a horse on mars"
input_imge = Image.open("./input.png")
        
# Double-GPUs parallel
pipe = StableDiffusionUniParallelPipeline.from_pretrained("runwayml/stable-diffusion-v1-5")
image = pipe(prompt).images[0] # txt2img
image = pipe(prompt, input_imge).images[0] # img2img

# Single-GPU parallel
pipe = StableDiffusionUniParallelPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", single_gpu_parallel=True)
image = pipe(prompt).images[0] # txt2img
image = pipe(prompt, input_imge).images[0] # img2img

Feature

1. Improve code and GPUs reusability
2. Accelerate inference performance

Benchmark (GPU: 12GB x2)

Pipeline	Time
StableDiffusionPipeline	6.8s
StableDiffusionUniPipeline(Only-Merging)	6.8s
StableDiffusionUniParallelPipeline(Single-GPU)	3.4s
StableDiffusionUniParallelPipeline(Double-GPUs)	1.8s