app.py

import gradio as gr
import requests
import zipfile
import io
from PIL import Image
import openai
import os
from scipy import ndimage
# from dotenv import load_dotenv
import numpy as np
import hashlib
import queue

# load_dotenv()


cache = dict()
que = queue.Queue(30)


def save_to_memory_cache(key, file):
    print('save mask')
    cache[key] = file
    que.put(key)
    if que.full():
        rkey = que.get()
        del cache[rkey]


def get_circle_footprint(size):
    xx, yy = np.mgrid[0:size * 2, 0:size * 2]
    fp = ((xx - size) ** 2 + (yy - size) ** 2) < (size) ** 2
    return fp


def request_mask(img_path):
    removebg_api_key = os.getenv('REMOVEBG_API_KEY')
    response = requests.post(
        'https://api.remove.bg/v1.0/removebg',
        files={'image_file': open(img_path, 'rb')},
        data={
            'size': 'auto',
            'format': 'zip'
        },
        headers={'X-Api-Key': removebg_api_key},
    )
    if response.status_code == requests.codes.ok:
        zipFile = zipfile.ZipFile(io.BytesIO(response.content))
        maskImFile = zipFile.read('alpha.png')
        return maskImFile
    else:
        print("Error:", response.status_code, response.text)
        return None


def get_file_hash(path):
    with open(path, 'rb') as inputfile:
        fh = hashlib.sha256()
        fb = inputfile.read(65536)
        while len(fb) > 0:
            fh.update(fb)
            fb = inputfile.read(65536)
    return fh.hexdigest()


def process_image(prompt, img_path, mask_margin):
    openai.api_key = os.getenv('OPENAI_API_KEY')

    hsh = get_file_hash(img_path)
    print('hash',hsh)

    maskImFile = None
    if hsh in cache:
        maskImFile = cache[hsh]
    else:
        maskImFile = request_mask(img_path)
        if maskImFile != None:
            save_to_memory_cache(hsh, maskImFile)
        else:
            print('no mask received')
            return 'https://i.imgur.com/DUd0OWN.png'

    maskIm = Image.open(io.BytesIO(maskImFile))

    alpha = maskIm.getchannel(0)
    if mask_margin > 0:
        inflated_alpha = ndimage.maximum_filter(input=np.array(
            alpha), footprint=get_circle_footprint(mask_margin))
        alpha = Image.fromarray(np.uint8(inflated_alpha))
    maskIm.paste((255), [0, 0, maskIm.size[0], maskIm.size[1]])
    maskIm.putalpha(alpha)

    maskFile = io.BytesIO()
    maskIm.save(maskFile, format='PNG')
    maskFile.seek(0)

    response = openai.Image.create_edit(
        image=open(img_path, "rb"),
        mask=maskFile,
        prompt=prompt,
        n=1,
        size="512x512"
    )
    return response['data'][0]['url']

# demo = gr.Interface(fn=greet, inputs="text", outputs="text")
demo = gr.Interface(process_image, inputs=[
    'text',
    gr.Image(type='filepath', shape=(500, 500), label='image'),
    gr.Slider(minimum=0, maximum=10, value=5, step=1, label="mask margin")
], outputs=["image"])

demo.launch()