app.py

import torch
from PIL import Image
from RealESRGAN import RealESRGAN
import gradio as gr
import spaces

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model2 = RealESRGAN(device, scale=2)
model2.load_weights('weights/RealESRGAN_x2.pth', download=True)
model4 = RealESRGAN(device, scale=4)
model4.load_weights('weights/RealESRGAN_x4.pth', download=True)
model8 = RealESRGAN(device, scale=8)
model8.load_weights('weights/RealESRGAN_x8.pth', download=True)

@spaces.GPU
def inference(image, size):
    global model2
    global model4
    global model8
    if image is None:
        raise gr.Error("Image not uploaded")
        

    if torch.cuda.is_available():
        torch.cuda.empty_cache()
    
    if size == '2x':
        try:
            result = model2.predict(image.convert('RGB'))
        except torch.cuda.OutOfMemoryError as e:
            print(e)
            model2 = RealESRGAN(device, scale=2)
            model2.load_weights('weights/RealESRGAN_x2.pth', download=False)
            result = model2.predict(image.convert('RGB'))
    elif size == '4x':
        try:
            result = model4.predict(image.convert('RGB'))
        except torch.cuda.OutOfMemoryError as e:
            print(e)
            model4 = RealESRGAN(device, scale=4)
            model4.load_weights('weights/RealESRGAN_x4.pth', download=False)
            result = model2.predict(image.convert('RGB'))
    else:
        try:
            width, height = image.size
            if width >= 6000 or height >= 6000:
                raise gr.Error("The image is too large.")
            result = model8.predict(image.convert('RGB'))
        except torch.cuda.OutOfMemoryError as e:
            print(e)
            model8 = RealESRGAN(device, scale=8)
            model8.load_weights('weights/RealESRGAN_x8.pth', download=False)
            result = model2.predict(image.convert('RGB'))
            
    print(f"Image size ({device}): {size} ... OK")
    return result


title = "ProFaker"

gr.Interface(inference,
    [gr.Image(type="pil"), 
    gr.Radio(["2x", "4x", "8x"], 
    type="value",
    value="2x",
    label="Resolution model")], 
    gr.Image(type="pil", label="Output"),
    title=title,
    flagging_mode="never",
    cache_mode="lazy",
    delete_cache=(44000, 44000),
    ).queue(api_open=True).launch(show_error=True, show_api=True)