app.py

import gradio as gr
import os
import re
#from tempfile import NamedTemporaryFile
import numpy as np
import spaces
import random
import string
from diffusers import AutoPipelineForText2Image
import torch
from pathos.multiprocessing import ProcessingPool as ProcessPoolExecutor
import requests
from lxml.html import fromstring

pool = ProcessPoolExecutor(16)
pool.__enter__()

#model_id = "runwayml/stable-diffusion-v1-5"
#model_id = "kandinsky-community/kandinsky-3"
model_id = "stabilityai/stable-diffusion-3-medium-diffusers"

device = "cuda" if torch.cuda.is_available() else "cpu"
if torch.cuda.is_available():
    torch.cuda.max_memory_allocated(device=device)
    pipe = AutoPipelineForText2Image.from_pretrained(model_id, torch_dtype=torch.float16, variant="fp16", use_safetensors=True, token=os.getenv('hf_token'))
    pipe = pipe.to(device)
else: 
    pipe = AutoPipelineForText2Image.from_pretrained(model_id, use_safetensors=True, token=os.getenv('hf_token'))
    pipe = pipe.to(device)

def translate(text,lang):
    text = re.sub(f'[{string.punctuation}]', '', re.sub('[\s+]', ' ', text)).lower().strip()
    lang = re.sub(f'[{string.punctuation}]', '', re.sub('[\s+]', ' ', lang)).lower().strip()
    user_agents = [
        'Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/126.0.0.0 Safari/537.36',
        'Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/126.0.0.0 Safari/537.36',
        'Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/126.0.0.0 Safari/537.36',
        'Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/126.0.0.0 Safari/537.36',
        'Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/126.0.0.0 Safari/537.36',
        'Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/605.1.15 (KHTML, like Gecko) Version/16.1 Safari/605.1.15',
        'Mozilla/5.0 (Macintosh; Intel Mac OS X 13_1) AppleWebKit/605.1.15 (KHTML, like Gecko) Version/16.1 Safari/605.1.15'
    ]
    url = 'https://www.google.com/search'
    resp = requests.get(
        url = url,
        params = {'q': f'{lang} translate {text}'},
        headers = {
            'User-Agent': random.choice(user_agents)
        }
    )
    content = resp.content
    html = fromstring(content)
    
    #src = html.xpath('//pre[@data-placeholder="Enter text"]/textarea')[0].text.strip()

    translated = text
    
    try:   
        trgt = html.xpath('//span[@class="target-language"]')[0].text.strip()
        rslt = html.xpath('//pre[@aria-label="Translated text"]/span')[0].text.strip()
        if trgt.lower() == lang.lower():
            translated = rslt
    except:
        raise Exception("Translation Error!")

    ret = re.sub(f'[{string.punctuation}]', '', re.sub('[\s+]', ' ', translated)).lower().strip()
    print(ret)
    return ret

def generate_random_string(length):
    characters = string.ascii_letters + string.digits
    return ''.join(random.choice(characters) for _ in range(length))

@spaces.GPU(duration=100)
def Piper(_do,_dont):
    return pipe(
        _do,
        height=528,
        width=704,
        negative_prompt=_dont,
        num_inference_steps=400,
        guidance_scale=9
    )

def infer(prompt,prompt2):
    name = generate_random_string(12)+".png"
    prompt_en = translate(prompt,"english")
    prompt2_en = translate(prompt2,"english")
    if prompt == None or prompt.strip() == "":
        _do = 'natural colors, rough texture, proportional content, dynamic poze, realistic details, award winning photograph, soft natural lighting, deep field, high definition, highly detailed, 8k'
    else:
        _do = f'{ prompt_en }, natural colors, rough texture, proportional content, dynamic poze, realistic details, award winning photograph, soft natural lighting, deep field, high definition, highly detailed, 8k'
    if prompt2 == None or prompt2.strip() == "":
        _dont = 'ugly, deformed, inflated, disfigured, poor details, bad anatomy, labels, texts, logos'
    else:
        _dont = f'ugly, deformed, inflated, disfigured, poor details, bad anatomy, {prompt2_en} where in {prompt_en}, logo where in {prompt_en}, text where in {prompt_en}, labels where in {prompt_en}, {prompt2_en}, labels'
    image = Piper(_do,_dont).images[0].save(name)
    return name

css="""
#col-container {
    margin: 0 auto;
    max-width: 15cm;
}
#image-container {
    aspect-ratio: 4 / 3;
}
.dropdown-arrow {
    display: none !important;
}
"""

js="""
function custom(){
    document.querySelector("div#prompt input").setAttribute("maxlength","35");
    document.querySelector("div#prompt2 input").setAttribute("maxlength","15");
}
"""

if torch.cuda.is_available():
    power_device = "GPU"
else:
    power_device = "CPU"

with gr.Blocks(theme=gr.themes.Soft(),css=css,js=js) as demo:
    with gr.Column(elem_id="col-container"):
        gr.Markdown(f"""
            # Image Generator
            Currently running on {power_device}.
        """)
        with gr.Row():
            prompt = gr.Textbox(
                elem_id="prompt",
                placeholder="Photo Description",
                container=False,
                rtl=True,
                max_lines=1
            )
        with gr.Row():
            prompt2 = gr.Textbox(
                elem_id="prompt2",
                placeholder="Forbidden Content",
                container=False,
                rtl=True,
                max_lines=1
            )
        with gr.Row():
            run_button = gr.Button("Run")
        result = gr.Image(elem_id="image-container", label="Result", show_label=False, type='filepath')
    prompt.submit(
        fn = infer,
        inputs = [prompt,prompt2],
        outputs = [result]
    )
    prompt2.submit(
        fn = infer,
        inputs = [prompt,prompt2],
        outputs = [result]
    )
    run_button.click(
        fn = infer,
        inputs = [prompt,prompt2],
        outputs = [result]
    )

demo.queue().launch()