import ast import copy import glob import hashlib import logging import os import re import time from pathlib import Path from typing import List, Optional, Tuple from urllib.parse import urlparse from PIL import Image, ImageDraw, ImageFont from io import BytesIO import requests import concurrent.futures import random import gradio as gr import PIL from gradio import processing_utils from gradio_client.client import DEFAULT_TEMP_DIR from text_generation import Client from transformers import AutoProcessor MODELS = [ # "HuggingFaceM4/idefics-9b-instruct", "HuggingFaceM4/idefics-80b-instruct", ] API_PATHS = { "HuggingFaceM4/idefics-9b-instruct": ( "https://api-inference.huggingface.co/models/HuggingFaceM4/idefics-9b-instruct" ), "HuggingFaceM4/idefics-80b-instruct": ( "https://api-inference.huggingface.co/models/HuggingFaceM4/idefics-80b-instruct" ), } SYSTEM_PROMPT = [ """The following is a conversation between a highly knowledgeable and intelligent visual AI assistant, called Assistant, and a human user, called User. In the following interactions, User and Assistant will converse in natural language, and Assistant will answer in a sassy way. Assistant's main purpose is to create funny meme texts from the images User provides. Assistant should be funny, sassy, and impertinent, and sometimes Assistant roasts people. Assistant should not be mean. It should not say toxic, homophobic, sexist, racist, things or any demeaning things that can make people uncomfortable. Assistant was created by Hugging Face. Here's a conversation example:""", """\nUser:""", "https://ichef.bbci.co.uk/news/976/cpsprodpb/7727/production/_103330503_musk3.jpg", "Write a meme for that image.", """\nAssistant: When you're trying to quit smoking but the cravings are too strong.""", "\nUser:How about this image?", "https://www.boredpanda.com/blog/wp-content/uploads/2017/01/image-copy-copy-587d0e7918b57-png__700.jpg", "Write something funny about this image.", """\nAssistant: Eggcellent service!""", "\nUser: Roast this person", "https://i.pinimg.com/564x/98/34/4b/98344b2483bd7c8b71a5c0fed6fe20b6.jpg", "", """\nAssistant: Damn your handwritting is pretty awful. But I suppose it must be pretty hard to hold a pen, considering you are a hammerhead shark.""", ] BAN_TOKENS = ( # For documentation puporse. We are not using this list, it is hardcoded inside `idefics_causal_lm.py` inside TGI. ";" ) EOS_STRINGS = ["", "\nUser:"] STOP_SUSPECT_LIST = [] API_TOKEN = os.getenv("HF_AUTH_TOKEN") IDEFICS_LOGO = "https://huggingface.co/spaces/HuggingFaceM4/idefics_playground/resolve/main/IDEFICS_logo.png" PROCESSOR = AutoProcessor.from_pretrained( "HuggingFaceM4/idefics-9b-instruct", token=API_TOKEN, ) BOT_AVATAR = "IDEFICS_logo.png" IMAGE_GALLERY_PATHS = [ f"example_images/{image_dir}/{ex_image}" for image_dir in os.listdir("example_images") for ex_image in os.listdir(f"example_images/{image_dir}") ] random.shuffle(IMAGE_GALLERY_PATHS) logging.basicConfig(level=logging.INFO) logger = logging.getLogger() # Monkey patch adapted from gradio.components.image.Image - mostly to make the `save` step optional in `pil_to_temp_file` def hash_bytes(bytes: bytes): sha1 = hashlib.sha1() sha1.update(bytes) return sha1.hexdigest() def pil_to_temp_file( img: PIL.Image.Image, dir: str = DEFAULT_TEMP_DIR, format: str = "png", resize: bool = False, ) -> str: """Save a PIL image into a temp file""" if resize: img = img.resize((224, 224), Image.LANCZOS) bytes_data = processing_utils.encode_pil_to_bytes(img, format) temp_dir = Path(dir) / hash_bytes(bytes_data) temp_dir.mkdir(exist_ok=True, parents=True) filename = str(temp_dir / f"image.{format}") if not os.path.exists(filename): img.save(filename, pnginfo=processing_utils.get_pil_metadata(img)) return filename def pil_to_base64( img: PIL.Image.Image, resize: bool = False, ) -> str: """Save a PIL image into a temp file""" if resize: img = img.resize((224, 224), Image.LANCZOS) base64_img = processing_utils.encode_pil_to_base64(img) return base64_img def add_file_gallery(selected_state: gr.SelectData, gallery_list: List[str]): return ( "Write a meme about this image.", gallery_list[selected_state.index]["name"], "", ) def choose_gallery(gallery_type: str): if gallery_type == "Meme templates": image_gallery_list = [ f"example_images/meme_templates/{ex_image}" for ex_image in os.listdir("example_images/meme_templates") ] elif gallery_type == "Funny images": image_gallery_list = [ f"example_images/funny_images/{ex_image}" for ex_image in os.listdir("example_images/funny_images") ] elif gallery_type == "Politics": image_gallery_list = [ f"example_images/politics_memes/{ex_image}" for ex_image in os.listdir("example_images/politics_memes") ] else: image_gallery_list = [ f"example_images/{image_dir}/{ex_image}" for image_dir in os.listdir("example_images") for ex_image in os.listdir(f"example_images/{image_dir}") ] random.shuffle(image_gallery_list) return image_gallery_list # Utils to handle the image markdown display logic def split_str_on_im_markdown(string: str) -> List[str]: """ Extract from a string (typically the user prompt string) the potential images from markdown Examples: - `User:![](https://favurl.com/chicken_on_money.png)Describe this image.` would become `["User:", "https://favurl.com/chicken_on_money.png", "Describe this image."]` - `User:![](/file=/my_temp/chicken_on_money.png)Describe this image.` would become `["User:", "/my_temp/chicken_on_money.png", "Describe this image."]` """ IMAGES_PATTERN = re.compile(r"!\[[^\]]*\]\((.*?)\s*(\"(?:.*[^\"])\")?\s*\)") parts = [] cursor = 0 for pattern in IMAGES_PATTERN.finditer(string): start = pattern.start() if start != cursor: parts.append(string[cursor:start]) image_url = pattern.group(1) if image_url.startswith("/file="): image_url = image_url[6:] # Remove the 'file=' prefix parts.append(image_url) cursor = pattern.end() if cursor != len(string): parts.append(string[cursor:]) return parts def is_image(string: str) -> bool: """ There are two ways for images: local image path or url. """ return is_url(string) or string.startswith(DEFAULT_TEMP_DIR) def is_url(string: str) -> bool: """ Checks if the passed string contains a valid url and nothing else. e.g. if space is included it's immediately invalidated the url """ if " " in string: return False result = urlparse(string) return all([result.scheme, result.netloc]) def isolate_images_urls(prompt_list: List) -> List: """ Convert a full string prompt to the list format expected by the processor. In particular, image urls (as delimited by ) should be their own elements. From: ``` [ "bonjourhello", PIL.Image.Image, "Aurevoir", ] ``` to: ``` [ "bonjour", IMG_URL, "hello", PIL.Image.Image, "Aurevoir", ] ``` """ linearized_list = [] for prompt in prompt_list: # Prompt can be either a string, or a PIL image if isinstance(prompt, PIL.Image.Image): linearized_list.append(prompt) elif isinstance(prompt, str): if "" not in prompt: linearized_list.append(prompt) else: prompt_splitted = prompt.split("") for ps in prompt_splitted: if ps == "": continue if ps.startswith(" PIL.Image.Image: """Fetching images""" return PROCESSOR.image_processor.fetch_images(url_list) def handle_manual_images_in_user_prompt(user_prompt: str) -> List[str]: """ Handle the case of textually manually inputted images (i.e. the ``) in the user prompt by fetching them, saving them locally and replacing the whole sub-sequence the image local path. """ if "" in user_prompt: splitted_user_prompt = isolate_images_urls([user_prompt]) resulting_user_prompt = [] for u_p in splitted_user_prompt: if is_url(u_p): img = fetch_images([u_p])[0] tmp_file = pil_to_temp_file(img) resulting_user_prompt.append(tmp_file) else: resulting_user_prompt.append(u_p) return resulting_user_prompt else: return [user_prompt] def prompt_list_to_markdown(prompt_list: List[str]) -> str: """ Convert a user prompt in the list format (i.e. elements are either a PIL image or a string) into the markdown format that is used for the chatbot history and rendering. """ resulting_string = "" for elem in prompt_list: if is_image(elem): if is_url(elem): resulting_string += f"![]({elem})" else: resulting_string += f"![](/file={elem})" else: resulting_string += elem return resulting_string def prompt_list_to_tgi_input(prompt_list: List[str]) -> str: """ TGI expects a string that contains both text and images in the image markdown format (i.e. the `![]()` ). The images links are parsed on TGI side """ result_string_input = "" for elem in prompt_list: if is_image(elem): if is_url(elem): try: response = requests.get(elem) if response.status_code == 200: elem_pil = Image.open(BytesIO(response.content)) except Exception: print(f"Image is not loading") else: elem_pil = Image.open(elem) base64_img = pil_to_base64(elem_pil, resize=True) result_string_input += f"![]({base64_img})" else: result_string_input += elem return result_string_input def remove_spaces_around_token(text: str) -> str: pattern = r"\s*()\s*" replacement = r"\1" result = re.sub(pattern, replacement, text) return result # Chatbot utils def insert_backslash(string, max_length=50): # Check if the string length is less than or equal to the max_length if len(string) <= max_length: return string # Start from the max_length character and search for the last space character before it for i in range(max_length - 1, -1, -1): if string[i] == " ": # Insert a backslash before the last space character return string[:i] + "\n" + string[i:] # If no space character is found, just insert a backslash at the max_length character return string[:max_length] + "\n" + string[max_length:] def resize_with_ratio(image: PIL.Image.Image, fixed_width: int) -> PIL.Image.Image: # Get the current width and height width, height = image.size # Calculate the new width while maintaining the aspect ratio up to 2:3 ratio new_width = fixed_width new_height = min(int(height * (new_width / width)), int(1.5 * new_width)) # Resize the image resized_img = image.resize((new_width, new_height), Image.LANCZOS) return resized_img def make_new_lines(draw, image, font, text_is_too_long, lines, num_lines, num_loops): max_len_increment = 0 while text_is_too_long and max_len_increment < 10: new_lines = lines.copy() last_line_with_backslash = insert_backslash( new_lines[-1], max_length=(len(new_lines[-1]) + max_len_increment) // (num_lines - num_loops), ) penultimate_line, last_line = ( last_line_with_backslash.split("\n")[0], last_line_with_backslash.split("\n")[1], ) new_lines.pop(-1) new_lines.append(penultimate_line) new_lines.append(last_line) # If the we haven't reached the last line, we split it again if len(new_lines) < num_lines: new_lines, text_width, text_is_too_long = make_new_lines( draw=draw, image=image, font=font, text_is_too_long=text_is_too_long, lines=new_lines, num_lines=num_lines, num_loops=num_loops + 1, ) text_width = max([draw.textlength(line, font) for line in new_lines]) text_is_too_long = text_width > image.width max_len_increment += 1 if not text_is_too_long: lines = new_lines return lines, text_width, text_is_too_long def test_font_size( draw, image, text, font, font_meme_text, num_lines=1, min_font=35, font_size_reduction=5, ): text_width = draw.textlength(text, font) text_is_too_long = True lines = [text] while font.size > min_font and text_is_too_long: font = ImageFont.truetype( f"fonts/{font_meme_text}.ttf", size=font.size - font_size_reduction ) if num_lines == 1: text_width = draw.textlength(text, font) text_is_too_long = text_width > image.width else: lines, text_width, text_is_too_long = make_new_lines( draw=draw, image=image, font=font, text_is_too_long=text_is_too_long, lines=lines, num_lines=num_lines, num_loops=0, ) temp_text = "\n".join(lines) if not text_is_too_long and num_lines > 1: text = temp_text return text, font, text_width, text_is_too_long def make_meme_image( image: str, text: str, font_meme_text: str, all_caps_meme_text: bool = False, text_at_the_top: bool = False, ) -> PIL.Image.Image: """ Takes an image and a text and returns a meme image. """ text = text.replace("\nUser", " ").replace("\n", " ").strip().rstrip(".") if all_caps_meme_text: text = text.upper() # Resize image fixed_width = 700 image = Image.open(image) image = resize_with_ratio(image, fixed_width) image_width, image_height = image.size height_width_ratio = image_height / image_width draw = ImageDraw.Draw(image) min_font = 35 initial_font_size = 60 if height_width_ratio >= 1: min_font = 45 initial_font_size = 80 text_is_too_long = True num_lines = 0 while text_is_too_long and num_lines < 8: num_lines += 1 font = ImageFont.truetype(f"fonts/{font_meme_text}.ttf", size=initial_font_size) text, font, text_width, text_is_too_long = test_font_size( draw, image, text, font, font_meme_text, num_lines=num_lines, min_font=min_font, font_size_reduction=5, ) if text_is_too_long: text = f"Text is too long to fit the image" if all_caps_meme_text: text = text.upper() font = ImageFont.truetype(f"fonts/{font_meme_text}.ttf", size=font.size) text_width = draw.textlength(text, font) outline_width = 2 text_x = (image_width - text_width) / 2 text_y = image_height - num_lines * font.size - 10 - 2 * num_lines if text_at_the_top: text_y = 0 for i in range(-outline_width, outline_width + 1): for j in range(-outline_width, outline_width + 1): draw.multiline_text( (text_x + i, text_y + j), text, fill="black", align="center", font=font ) draw.multiline_text((text_x, text_y), text, fill="white", align="center", font=font) return image def format_user_prompt_with_im_history_and_system_conditioning( system_prompt: List[str], current_user_prompt_str: str, current_image: Optional[str], history: List[Tuple[str, str]], ) -> Tuple[List[str], List[str]]: """ Produces the resulting list that needs to go inside the processor. It handles the potential image box input, the history and the system conditionning. """ # resulting_list = copy.deepcopy(SYSTEM_PROMPT) resulting_list = system_prompt # Format history for turn in history: user_utterance, assistant_utterance = turn splitted_user_utterance = split_str_on_im_markdown(user_utterance) optional_space = "" if not is_image(splitted_user_utterance[0]): optional_space = " " resulting_list.append(f"\nUser:{optional_space}") resulting_list.extend(splitted_user_utterance) resulting_list.append(f"\nAssistant: {assistant_utterance}") # Format current input current_user_prompt_str = remove_spaces_around_token(current_user_prompt_str) if current_image is None: if "![](" in current_user_prompt_str: current_user_prompt_list = split_str_on_im_markdown(current_user_prompt_str) else: current_user_prompt_list = handle_manual_images_in_user_prompt( current_user_prompt_str ) optional_space = "" if not is_image(current_user_prompt_list[0]): # Check if the first element is an image (and more precisely a path to an image) optional_space = " " resulting_list.append(f"\nUser:{optional_space}") resulting_list.extend(current_user_prompt_list) resulting_list.append("\nAssistant:") else: # Choosing to put the image first when the image is inputted through the UI, but this is an arbiratrary choice. resulting_list.extend( [ "\nUser:", current_image, f"{current_user_prompt_str}\nAssistant:", ] ) current_user_prompt_list = [current_user_prompt_str] return resulting_list, current_user_prompt_list def expand_layout(): return gr.Column(scale=2), gr.Gallery(height=682) def generate_meme( client, query, image, font_meme_text, all_caps_meme_text, text_at_the_top, generation_args, ): try: text = client.generate(prompt=query, **generation_args).generated_text except Exception as e: logger.error(f"Error {e} while generating meme text") text = "" if image is not None and text != "": meme_image = make_meme_image( image=image, text=text, font_meme_text=font_meme_text, all_caps_meme_text=all_caps_meme_text, text_at_the_top=text_at_the_top, ) return meme_image else: return None def model_inference( model_selector, system_prompt, user_prompt_str, chat_history, image, decoding_strategy, temperature, max_new_tokens, repetition_penalty, top_p, all_caps_meme_text, text_at_the_top, font_meme_text, ): chat_history = [] if user_prompt_str.strip() == "" and image is None: return "", None, chat_history system_prompt = ast.literal_eval(system_prompt) ( formated_prompt_list, user_prompt_list, ) = format_user_prompt_with_im_history_and_system_conditioning( system_prompt=system_prompt, current_user_prompt_str=user_prompt_str.strip(), current_image=image, history=chat_history, ) client_endpoint = API_PATHS[model_selector] client = Client( base_url=client_endpoint, headers={"x-use-cache": "0", "Authorization": f"Bearer {API_TOKEN}"}, timeout=45, ) # Common parameters to all decoding strategies # This documentation is useful to read: https://huggingface.co/docs/transformers/main/en/generation_strategies generation_args = { "max_new_tokens": max_new_tokens, "repetition_penalty": repetition_penalty, "stop_sequences": EOS_STRINGS, } assert decoding_strategy in [ "Greedy", "Top P Sampling", ] if decoding_strategy == "Greedy": generation_args["do_sample"] = False elif decoding_strategy == "Top P Sampling": generation_args["temperature"] = temperature generation_args["do_sample"] = True generation_args["top_p"] = top_p chat_history.append([prompt_list_to_markdown(user_prompt_list), ""]) query = prompt_list_to_tgi_input(formated_prompt_list) all_meme_images = [] with concurrent.futures.ThreadPoolExecutor(max_workers=4) as executor: futures = [ executor.submit( generate_meme, client, query, image, font_meme_text, all_caps_meme_text, text_at_the_top, generation_args, ) for i in range(4) ] for future in concurrent.futures.as_completed(futures): meme_image = future.result(timeout=45) if meme_image: all_meme_images.append(meme_image) return user_prompt_str, all_meme_images, chat_history def remove_last_turn(chat_history): if len(chat_history) == 0: return chat_history, "", "" last_interaction = chat_history[-1] chat_history = chat_history[:-1] chat_update = chat_history text_update = last_interaction[0] return chat_update, text_update, "" textbox = gr.Textbox( placeholder="Upload an image and ask the AI to create a meme!", show_label=False, value="Write a meme about this image.", visible=True, container=False, label="Text input", scale=8, max_lines=5, ) chatbot = gr.Chatbot( elem_id="chatbot", label="AI Meme Generator Chatbot", visible=False, avatar_images=[None, BOT_AVATAR], ) css = """ .gradio-container{max-width: 1000px!important} h1{display: flex;align-items: center;justify-content: center;gap: .25em} *{transition: width 0.5s ease, flex-grow 0.5s ease} """ with gr.Blocks(title="AI Meme Generator", theme=gr.themes.Base(), css=css) as demo: with gr.Row(scale=0.5): gr.HTML( """

AI Meme Generator powered by IDEFICS

""" ) with gr.Row(elem_id="model_selector_row"): model_selector = gr.Dropdown( choices=MODELS, value="HuggingFaceM4/idefics-80b-instruct", interactive=True, show_label=False, container=False, label="Model", visible=False, ) with gr.Row(equal_height=True): # scale=2 when expanded with gr.Column(scale=4, min_width=250) as upload_area: imagebox = gr.Image( type="filepath", label="Image to meme", height=272, visible=True ) with gr.Group(): with gr.Row(): textbox.render() with gr.Row(): submit_btn = gr.Button( value="▶️ Submit", visible=True, min_width=120 ) clear_btn = gr.ClearButton( [textbox, imagebox, chatbot], value="🧹 Clear", min_width=120 ) regenerate_btn = gr.Button( value="🔄 Regenerate", visible=True, min_width=120 ) with gr.Accordion( "Advanced settings", open=False, visible=True ) as parameter_row: with gr.Row(): with gr.Column(): all_caps_meme_text = gr.Checkbox( value=True, label="All Caps", interactive=True, info="", ) text_at_the_top = gr.Checkbox( value=False, label="Text at the top", interactive=True, info="", ) with gr.Column(): font_meme_text = gr.Radio( [ "impact", "Roboto-Regular", ], value="impact", label="Font", interactive=True, info="", ) system_prompt = gr.Textbox( value=SYSTEM_PROMPT, visible=False, lines=20, max_lines=50, interactive=True, ) max_new_tokens = gr.Slider( minimum=8, maximum=150, value=90, step=1, interactive=True, label="Maximum number of new tokens to generate", ) repetition_penalty = gr.Slider( minimum=0.0, maximum=5.0, value=1.2, step=0.01, interactive=True, label="Repetition penalty", info="1.0 is equivalent to no penalty", ) decoding_strategy = gr.Radio( [ "Greedy", "Top P Sampling", ], value="Top P Sampling", label="Decoding strategy", interactive=True, info="Higher values is equivalent to sampling more low-probability tokens.", ) temperature = gr.Slider( minimum=0.0, maximum=5.0, value=0.6, step=0.1, interactive=True, visible=True, label="Sampling temperature", info="Higher values will produce more diverse outputs.", ) decoding_strategy.change( fn=lambda selection: gr.Slider.update( visible=( selection in [ "contrastive_sampling", "beam_sampling", "Top P Sampling", "sampling_top_k", ] ) ), inputs=decoding_strategy, outputs=temperature, ) top_p = gr.Slider( minimum=0.01, maximum=0.99, value=0.8, step=0.01, interactive=True, visible=True, label="Top P", info="Higher values is equivalent to sampling more low-probability tokens.", ) decoding_strategy.change( fn=lambda selection: gr.Slider.update( visible=(selection in ["Top P Sampling"]) ), inputs=decoding_strategy, outputs=top_p, ) with gr.Column(scale=5) as result_area: generated_memes_gallery = gr.Gallery( # value="Images generated will appear here", label="IDEFICS Generated Memes", allow_preview=True, elem_id="generated_memes_gallery", show_download_button=True, show_share_button=True, columns=[2], object_fit="contain", height=428, ) # height 600 when expanded with gr.Row(equal_height=True): with gr.Box(elem_id="gallery_box"): gallery_type_choice = gr.Radio( [ "All", "Meme templates", "Funny images", "Politics", ], value="All", label="Gallery Type", interactive=True, visible=False, info="Choose the type of gallery you want to see.", ) template_gallery = gr.Gallery( value=IMAGE_GALLERY_PATHS, label="Templates Gallery", allow_preview=False, columns=6, elem_id="gallery", show_share_button=False, height=400, ) with gr.Row(variant="panel"): with gr.Column(scale=1): gr.Image( IDEFICS_LOGO, elem_id="banner-image", show_label=False, show_download_button=False, height=200, width=250, ) with gr.Column(scale=5): gr.HTML( """

AI Meme Generator is an AI system that writes humorous content inspired by images, allowing you to make the funniest memes with little effort. Upload your image and ask the Idefics chatbot to make a tailored meme.

AI Meme Generator is a space inspired from AI Dad Jokes and powered by IDEFICS, an open-access large visual language model developped by Hugging Face. Like GPT-4, the multimodal model accepts arbitrary sequences of image and text inputs and produces text outputs. IDEFICS can answer questions about images, describe visual content, create stories grounded in multiple images, etc.

⛔️ Intended uses and limitations: This demo is provided as research artifact to the community showcasing IDEFICS'capabilities. We detail misuses and out-of-scope uses here. In particular, the system should not be used to engage in harassment, abuse and bullying. The model can produce factually incorrect texts, hallucinate facts (with or without an image) and will struggle with small details in images. While the system will tend to refuse answering questionable user requests, it can produce problematic outputs (including racist, stereotypical, and disrespectful texts), in particular when prompted to do so.

""" ) with gr.Row(): chatbot.render() gr.on( triggers=[ textbox.submit, imagebox.upload, submit_btn.click, template_gallery.select, regenerate_btn.click, ], fn=expand_layout, outputs=[upload_area, generated_memes_gallery], queue=False, ).then( fn=lambda: "", inputs=[], outputs=[generated_memes_gallery], queue=False ).then( fn=model_inference, inputs=[ model_selector, system_prompt, textbox, chatbot, imagebox, decoding_strategy, temperature, max_new_tokens, repetition_penalty, top_p, all_caps_meme_text, text_at_the_top, font_meme_text, ], outputs=[textbox, generated_memes_gallery, chatbot], ) regenerate_btn.click( fn=remove_last_turn, inputs=chatbot, outputs=[chatbot, textbox, generated_memes_gallery], queue=False, ) # gallery_type_choice.change( # fn=choose_gallery, # inputs=[gallery_type_choice], # outputs=[template_gallery], # queue=False, # ) template_gallery.select( fn=add_file_gallery, inputs=[template_gallery], outputs=[textbox, imagebox, generated_memes_gallery], queue=False, ) demo.load( # fn=choose_gallery, # inputs=[gallery_type_choice], # outputs=[template_gallery], queue=False, ) demo.queue(concurrency_count=4, max_size=40, api_open=False) demo.launch(max_threads=400)