app.py

import gradio as gr
import torch
from transformers import logging
import random
from PIL import Image
from Utils import MingleModel

logging.set_verbosity_error()


def get_concat_h(images):
    widths, heights = zip(*(i.size for i in images))

    total_width = sum(widths)
    max_height = max(heights)

    dst = Image.new('RGB', (total_width, max_height))
    x_offset = 0
    for im in images:
      dst.paste(im, (x_offset,0))
      x_offset += im.size[0]
    return dst


mingle_model = MingleModel()


def mingle_prompts(first_prompt, second_prompt):
    imgs = []
    text_input1 = mingle_model.tokenizer(first_prompt)
    text_input2 = mingle_model.tokenizer(second_prompt)
    with torch.no_grad():
        text_embeddings1 = mingle_model.text_encoder(text_input1)
        text_embeddings2 = mingle_model.text_encoder(text_input2)

    rand_generator = random.randint(1, 2048)
    # Mix them together
    mix_factors = [0.1, 0.3, 0.5, 0.7, 0.9]
    for mix_factor in mix_factors:
        mixed_embeddings = (text_embeddings1 * mix_factor + text_embeddings2 * (1 - mix_factor))

        # Generate!
        steps = 30
        guidence_scale = 8.0
        img = mingle_model.generate_with_embs(mixed_embeddings, rand_generator, num_inference_steps=steps,
                                 guidance_scale=guidence_scale)
        imgs.append(img)

    return get_concat_h(imgs)


with gr.Blocks() as demo:
    gr.Markdown(
        '''
        <p style="text-align: center;">Marketing and design made simple using AI.</p>
        ''')
    gr.Image('batman_venum.png', shape=(1024, 205))

    first_prompt = gr.Textbox(label="first_prompt")
    second_prompt = gr.Textbox(label="second_prompt")
    greet_btn = gr.Button("Submit")
    gr.Markdown("## Text Examples")
    gr.Examples([['batman, dynamic lighting, photorealistic fantasy concept art, trending on art station, stunning visuals, terrifying, creative, cinematic',
                  'venom, dynamic lighting, photorealistic fantasy concept art, trending on art station, stunning visuals, terrifying, creative, cinematic'],
                 ['A mouse', 'A leopard']], [first_prompt, second_prompt])

    gr.Markdown("# Output Results")
    output = gr.Image()

    greet_btn.click(fn=mingle_prompts, inputs=[first_prompt, second_prompt], outputs=[output])

demo.launch()