app.py

import os
import textwrap
from PIL import Image, ImageDraw, ImageFont

import gradio as gr
import numpy as np
import torch
from lavis.models import load_model_and_preprocess
import openai

device = torch.device("cuda") if torch.cuda.is_available() else "cpu"

model, vis_processors, _ = load_model_and_preprocess(
    name="blip2_opt", model_type="pretrain_opt2.7b", is_eval=True, device=device
)

openai.api_key = os.environ["OPENAI_API_KEY"]

def generate_caption(image):
    pil_image = image.copy()  # Create a copy of the input PIL image
    image = vis_processors["eval"](image).unsqueeze(0).to(device)
    caption = model.generate({"image": image})

    caption = "\n".join(caption)
    #use gpt-4 to generate a meme based on the caption
    response = openai.ChatCompletion.create(
                model="gpt-4",
                messages=[
                    {"role": "system", "content": "Escribe un meme chistoso en español a partir de la descripción de una imagen dada por el usuario. No uses emojis, ni comillas, ni saltos de línea. No es necesario que empieces con 'cuando'. El output del asistente solo debe ser el texto del meme. Debe ser corto pero chistoso."},
                    {"role": "user", "content": caption}
                ],
                temperature=0.6
            )
    
    meme_text = response.choices[0].message.content
    print(meme_text)

    # Put the meme text on the image
    draw = ImageDraw.Draw(pil_image)

    # Set the fixed font size to 80
    font_size = 60
    font = ImageFont.truetype("impact.ttf", font_size)

    # Calculate the average character width for the font
    alphabet = "ABCEMOPQRSTWXZ"
    total_char_width = sum(draw.textlength(char, font=font) for char in alphabet)
    average_char_width = total_char_width / len(alphabet)
    
    # Calculate the number of characters that fit within the image width
    chars_per_line = int(pil_image.width / average_char_width)
    
    # Wrap the text to fit within the image width
    wrapped_text = textwrap.fill(meme_text, width=chars_per_line)
    
    # Calculate the position to place the text at the top and center horizontally
    text_lines = wrapped_text.split('\n')
    y = 10  # Adjust this value to add more or less padding from the top
    for line in text_lines:
        line_width = draw.textlength(line, font=font)
        line_mask = font.getmask(line)
        _, line_height = line_mask.size
        x = (pil_image.width - line_width) // 2
    
        draw.text((x, y), line, fill=(255, 255, 255), font=font)
        y += line_height + int(line_height * 0.1)

    pil_image = pil_image.convert('RGB')

    if torch.cuda.is_available():
        torch.cuda.empty_cache()

    return pil_image


with gr.Blocks() as demo:
    gr.Markdown(
        "### Memero - Generador de Memes"
    )
    gr.Markdown(
        "Genera un meme en español a partir de una imagen."
    )

    with gr.Row():
        with gr.Column():
            input_image = gr.Image(label="Imagen", type="pil")
            btn_caption = gr.Button("Generar meme")
            output_text = gr.Image(label="Meme", lines=5)

    btn_caption.click(
        generate_caption, inputs=[input_image], outputs=[output_text]
    )

demo.launch()