Script for image2image using stablediffusion using incremental strength

i2it.py

from torch import autocast
import torch
import requests
from PIL import Image
import sys
from diffusers import StableDiffusionImg2ImgPipeline #, preprocess
import random
import argparse

# image to image generation using StableDiffusion 0.3.0 dev

# by default generates 29 images varying strength 0 to 1
# set --start and --step to vary the starting strength etc

# Point model to your stablediffusion pretrained folder
# or a save from textual inversion if needed



parser = argparse.ArgumentParser()

# define params and their types with defaults if needed
parser.add_argument('--model', type=str, default="/work/dtest/stable-diffusion-v1-4", help='path to model')
parser.add_argument('--text', type=str, default="", help='text prompt')
parser.add_argument('--image', type=str, default="", help='path to init image')
parser.add_argument('--output', type=str, default="", help='path to output')
parser.add_argument('--h', type=int, default=512, help='image height')
parser.add_argument('--w', type=int, default=512, help='image width')
parser.add_argument('--start', type=int, default=0, help='start aka skip')
parser.add_argument('--step', type=int, default=5, help='step to increase start point per iteration')
parser.add_argument('--g', type=float, default=7.5, help='guidance scale')
parser.add_argument('--seed', type=int, default=0, help='manual seed')

opt = parser.parse_args()


# load the pipeline
device = "cuda"
pipe = StableDiffusionImg2ImgPipeline.from_pretrained(
    #"CompVis/stable-diffusion-v1-4",
    #"file:///work/dtest/stable-diffusion-v1-4",
    "textual_inversion_st-pes6",
    revision="fp16", 
    torch_dtype=torch.float16,
    use_auth_token=True
).to(device)

pipe.enable_attention_slicing()

fn = opt.image
ofn = opt.output
start = opt.start
if opt.seed > 0:
    seed = opt.seed
else:
    seed = random.randint(0,63444)

g = opt.g
    
steps = 100

print(start, seed)

init_image = Image.open(fn).convert("RGB")
init_image = init_image.resize((opt.h, opt.w))
#init_image = preprocess(init_image)

prompt =  opt.text
step = opt.step

for n in range(start,steps, step):
    beta = n/steps

    with autocast("cuda"):
      generator = torch.Generator("cuda").manual_seed(seed)    
      images = pipe(prompt=prompt, init_image=init_image, strength=beta, guidance_scale=g, generator=generator)["sample"]

    images[0].save(ofn+"-"+str(n)+".png")