What this is

Downscale image, add noise on the edge and add perlin noise in the background.

Base image should be 512x512 px.

Base Image	Input Image	Mask Image

Results after inpainting:

Installation

Create and activate a python environnement, then:

git clone https://github.com/PhilSad/stable-diffusion-outpainting
pip install -r requirements.txt

Script Usage:

python ./downscale_and_noise.py --input_image ./images/base.png --output_init ./images/init.png --output_mask ./images/mask.png

Module Usage:

from downscale_and_noise import get_init_mask_image

with open('./images/base.png', 'rb') as f:
    base_image = Image.open(f)
    base_image.load()

init_image, mask_image = get_init_mask_image(base_image)

Uncrop / Outpaint image

import PIL
import requests
import torch
from io import BytesIO
from PIL import Image
from diffusers import StableDiffusionInpaintPipeline
import numpy as np

prompt = "A dog sitting on a park bench"

init_image_arr = np.array(init_image)
mask_image_arr = np.array(mask_image)
pipe = StableDiffusionInpaintPipeline.from_pretrained(
    "runwayml/stable-diffusion-inpainting"
).to('cuda')
pipe.enable_attention_slicing()

outpainted_image = pipe(prompt=prompt, image=init_image_arr, mask_image=mask_image_arr).images[0]