import gradio as gr import requests from io import BytesIO import os import json from dotenv import load_dotenv from PIL import Image, ImageDraw, ImageFont import fitz # PyMuPDF import arxiv import tiktoken from openai import OpenAI import textwrap from datetime import datetime # Load environment variables from .env file load_dotenv() openai_client = OpenAI(api_key=os.getenv("OPENAI_API_KEY")) def download_and_extract_paper_info(arxiv_id, token_limit=120000, model="gpt-3.5-turbo"): search = arxiv.Search(id_list=[arxiv_id]) paper = next(search.results()) title = paper.title publish_date = paper.published.date() pdf_url = f"https://arxiv.org/pdf/{arxiv_id}.pdf" response = requests.get(pdf_url) if response.status_code == 200: pdf_content = response.content doc = fitz.open(stream=pdf_content, filetype="pdf") text = "" encoding = tiktoken.encoding_for_model(model) for page in doc: page_text = page.get_text() text += page_text tokens = encoding.encode(text) if len(tokens) > token_limit: text = encoding.decode(tokens[:token_limit]) break return { "title": title, "publish_date": publish_date, "full_text": text } else: print(f"Failed to download paper. Status code: {response.status_code}") return None def summarize_text(text): prompt = f""" You are getting the text version of an arxiv paper your goal is to provide a summary of the paper by providing bullet points which summarise the paper. It should be exact three bullet points which summarise the paper. Return your response in JSON format where the keys are the bullet points and the values are the summaries of the bullet points as following: {{ "bullet_point_1": "content", "bullet_point_2": "content", "bullet_point_3": "content" }} Here is the text of the paper: {text} """ completion = openai_client.chat.completions.create( model="gpt-4o-mini", response_format={ "type": "json_object" }, messages=[ {"role": "user", "content": prompt} ], temperature=0.0, ) summary = completion.choices[0].message.content return summary def add_text_to_image(background_path, title, text_content, publish_date, output_path="output.jpg", scale_factor=2, offset=20): with Image.open(background_path) as img: width, height = img.size background = img.resize((width * scale_factor, height * scale_factor), Image.LANCZOS) draw = ImageDraw.Draw(background) title_font = ImageFont.truetype("fonts/Inika-Regular.ttf", 35 * scale_factor) content_font = ImageFont.truetype("fonts/Inika-Regular.ttf", 20 * scale_factor) date_font = ImageFont.truetype("fonts/Inika-Regular.ttf", 20 * scale_factor) arxiv_font = ImageFont.truetype("fonts/Larabieb.ttf", 50 * scale_factor) margin = 50 * scale_factor max_width = background.width - (2 * margin) # Dynamically calculate the width for wrapping the title wrapped_title = textwrap.wrap(title, width=int(max_width / (35 * scale_factor * 0.6))) y_text = 50 * scale_factor for line in wrapped_title: bbox = title_font.getbbox(line) line_width = bbox[2] - bbox[0] line_height = bbox[3] - bbox[1] x_text = (background.width - line_width) // 2 draw.text((x_text, y_text), line, font=title_font, fill=(0, 0, 0)) y_text += line_height + (10 * scale_factor) bullet_points = json.loads(text_content) total_height = sum(len(textwrap.wrap(value, width=90)) * (25 * scale_factor) + (20 * scale_factor) for value in bullet_points.values()) y = (background.height - total_height) // 2 bullet_width = content_font.getbbox("• ")[2] max_content_width = max(max(content_font.getbbox(line)[2] for line in textwrap.wrap(value, width=90)) for value in bullet_points.values()) bullet_start_x = (background.width - max_content_width - bullet_width) // 2 for value in bullet_points.values(): wrapped_text = textwrap.wrap(value, width=90) for i, line in enumerate(wrapped_text): if i == 0: draw.text((bullet_start_x, y), "•", font=content_font, fill=(0, 0, 0)) draw.text((bullet_start_x + bullet_width, y), line, font=content_font, fill=(0, 0, 0)) else: draw.text((bullet_start_x + bullet_width, y + (25 * scale_factor * i)), line, font=content_font, fill=(0, 0, 0)) y += (25 * scale_factor * len(wrapped_text)) + (20 * scale_factor) date_text = f"Published: {publish_date}" date_bbox = date_font.getbbox(date_text) date_height = date_bbox[3] - date_bbox[1] draw.text((margin, background.height - margin - date_height - offset), date_text, font=date_font, fill=(0, 0, 0)) arxiv_text = "@arXivGPT" arxiv_bbox = arxiv_font.getbbox(arxiv_text) arxiv_width = arxiv_bbox[2] - arxiv_bbox[0] arxiv_height = arxiv_bbox[3] - arxiv_bbox[1] arxiv_x = background.width - margin - arxiv_width arxiv_y = background.height - margin - arxiv_height - offset pre_x_text = "@ar" pre_x_width = arxiv_font.getbbox(pre_x_text)[2] draw.text((arxiv_x, arxiv_y), pre_x_text, font=arxiv_font, fill=(0, 0, 0)) x_text = "X" x_width = arxiv_font.getbbox(x_text)[2] draw.text((arxiv_x + pre_x_width, arxiv_y), x_text, font=arxiv_font, fill="#B31B1B") post_x_text = "ivGPT" draw.text((arxiv_x + pre_x_width + x_width, arxiv_y), post_x_text, font=arxiv_font, fill=(0, 0, 0)) background.save(output_path, quality=95) print(f"High-resolution image saved as {output_path}") def create_image_from_url(arxiv_id, background_path="background.jpg", output_path="output.jpg"): paper_info = download_and_extract_paper_info(arxiv_id) if paper_info: title = paper_info.get("title") publish_date = paper_info.get("publish_date") full_text = paper_info.get("full_text") summary = summarize_text(full_text) add_text_to_image(background_path, title, summary, publish_date, output_path) return output_path def fetch_arxiv_image(arxiv_link): arxiv_id = arxiv_link.split('/')[-1] output_path = create_image_from_url(arxiv_id) return Image.open(output_path) description_text = ( "It will only work with an arXiv link. Based on the arXiv paper, a summary of the paper is generated " "and displayed in arXivGPT format (https://x.com/arXivGPT). Please input the arXiv link below." ) demo = gr.Interface( fn=fetch_arxiv_image, inputs="text", outputs="image", allow_flagging="never", description=description_text ) demo.launch()