app_dialogue.py

import ast
import copy
import glob
import hashlib
import logging
import os
import re
from pathlib import Path
from typing import List, Optional, Tuple
from urllib.parse import urlparse
from PIL import Image, ImageDraw, ImageFont

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.<end_of_utterance>",
    """\nAssistant: When you're trying to quit smoking but the cravings are too strong.<end_of_utterance>""",
    "\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.<end_of_utterance>",
    """\nAssistant: Eggcellent service!<end_of_utterance>""",
    "\nUser: Roast this person",
    "https://i.pinimg.com/564x/98/34/4b/98344b2483bd7c8b71a5c0fed6fe20b6.jpg",
    "<end_of_utterance>",
    """\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.<end_of_utterance>""",
]

BAN_TOKENS = (  # For documentation puporse. We are not using this list, it is hardcoded inside `idefics_causal_lm.py` inside TGI.
    "<image>;<fake_token_around_image>"
)
EOS_STRINGS = ["<end_of_utterance>", "\nUser:"]
STOP_SUSPECT_LIST = []

GRADIO_LINK = "https://huggingfacem4-ai-meme-generator.hf.space"
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"

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"
) -> str:
    """Save a PIL image into a temp file"""
    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 add_file(file):
    return file.name, gr.update(label="🖼️ Uploaded!")


def add_file_gallery(selected_state: gr.SelectData, gallery_list: List[str]):
    gr.update(label="📁 Upload image", interactive=True)
    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


# This is a hack to make pre-computing the default examples work.
# During normal inference, we pass images as url to a local file using the method `gradio_link`
# which allows the tgi server to fetch the local image from the frontend server.
# however, we are building the space (and pre-computing is part of building the space), the frontend is not available
# and won't answer. So tgi server will try to fetch an image that is not available yet, which will result in a timeout error
# because tgi will never be able to return the generation.
# To bypass that, we pass instead the images URLs from the spaces repo.
DEFAULT_IMAGES_TMP_PATH_TO_URL = {}
for image_dir in os.listdir("example_images"):
    for im_path in os.listdir(f"example_images/{image_dir}"):
        H = gr.Image(
            f"example_images/{image_dir}/{im_path}", visible=False, type="filepath"
        )
        tmp_filename = H.preprocess(H.value)
        DEFAULT_IMAGES_TMP_PATH_TO_URL[
            tmp_filename
        ] = f"https://huggingface.co/spaces/HuggingFaceM4/AI_Meme_Generator/resolve/main/example_images/{image_dir}/{im_path}"


# 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 <fake_token_around_image>) should be their own elements.
    From:
    ```
    [
        "bonjour<fake_token_around_image><image:IMG_URL><fake_token_around_image>hello",
        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 "<fake_token_around_image>" not in prompt:
                linearized_list.append(prompt)
            else:
                prompt_splitted = prompt.split("<fake_token_around_image>")
                for ps in prompt_splitted:
                    if ps == "":
                        continue
                    if ps.startswith("<image:"):
                        linearized_list.append(ps[7:-1])
                    else:
                        linearized_list.append(ps)
        else:
            raise TypeError(
                f"Unrecognized type for `prompt`. Got {type(type(prompt))}. Was expecting something in [`str`,"
                " `PIL.Image.Image`]"
            )
    return linearized_list


def fetch_images(url_list: str) -> 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 `<fake_token_around_image><image:IMG_URL><fake_token_around_image>`) in the user prompt
    by fetching them, saving them locally and replacing the whole sub-sequence the image local path.
    """
    if "<fake_token_around_image>" 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 gradio_link(img_path: str) -> str:
    url = f"{GRADIO_LINK}/file={img_path}"
    return url


def prompt_list_to_markdown(prompt_list: List[str], size: int = None) -> 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):
                if size is not None:
                    resulting_string += f"<img src={elem} width={size} height={size}>"
                else:
                    resulting_string += f"![]({elem})"
            else:
                if size is not None:
                    resulting_string += f"<img src='/file={str(elem)}' width='{size}' height={str(size)}>"
                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):
                result_string_input += f"![]({elem})"
            else:
                result_string_input += f"![]({gradio_link(img_path=elem)})"
        else:
            result_string_input += elem
    return result_string_input


def remove_spaces_around_token(text: str) -> str:
    pattern = r"\s*(<fake_token_around_image>)\s*"
    replacement = r"\1"
    result = re.sub(pattern, replacement, text)
    return result


# Chatbot utils
Radio_options_to_font = {}


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 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

    if num_lines == 1:
        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
            )
            text_width = draw.textlength(text, font)
            text_is_too_long = text_width > image.width

    elif num_lines == 2:
        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
            )
            max_len_increment = 0
            while (
                text_is_too_long
                and max_len_increment < 10
                and max_len_increment < (len(text)) // 2
            ):
                temp_text = insert_backslash(
                    text, max_length=(len(text) + max_len_increment) // 2
                )
                first_line, second_line = (
                    temp_text.split("\n")[0],
                    temp_text.split("\n")[1],
                )
                text_width = max(
                    draw.textlength(first_line, font),
                    draw.textlength(second_line, font),
                )
                text_is_too_long = text_width > image.width
                max_len_increment += 1

    elif num_lines == 3:
        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
            )
            max_len_incr_1_split = 0
            while text_is_too_long and max_len_incr_1_split < 10:
                first_temp_text = insert_backslash(
                    text, max_length=(len(text) + max_len_incr_1_split) // 3
                )
                first_line, second_line = (
                    first_temp_text.split("\n")[0],
                    first_temp_text.split("\n")[1],
                )
                max_len_incr_2_split = 0
                while text_is_too_long and max_len_incr_2_split < 10:
                    temp_text_second_line = insert_backslash(
                        second_line,
                        max_length=(len(second_line) + max_len_incr_2_split) // 2,
                    )
                    second_line_1, second_line_2 = (
                        temp_text_second_line.split("\n")[0],
                        temp_text_second_line.split("\n")[1],
                    )
                    temp_text = first_line + "\n" + second_line_1 + "\n" + second_line_2
                    text_width = max(
                        draw.textlength(first_line, font),
                        draw.textlength(second_line_1, font),
                        draw.textlength(second_line_2, font),
                    )
                    text_is_too_long = text_width > image.width
                    max_len_incr_2_split += 1
                max_len_incr_1_split += 1
    else:
        raise (ValueError("num_lines can only be 1, 2 or 3"))

    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 = 30
    initial_font_size = 60
    if height_width_ratio > 1:
        min_font = 40
        initial_font_size = 80
    text_is_too_long = True
    num_lines = 0
    while text_is_too_long and num_lines < 3:
        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 - 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"<end_of_utterance>\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("<end_of_utterance>\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}<end_of_utterance>\nAssistant:",
            ]
        )
        current_user_prompt_list = [current_user_prompt_str]

    return resulting_list, current_user_prompt_list


# dope_callback = gr.CSVLogger()
# problematic_callback = gr.CSVLogger()

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: 970px!important}
h1{display: flex;align-items: center;justify-content: center;gap: .25em}
'''
with gr.Blocks(title="AI Meme Generator", theme=gr.themes.Base(), css=css) as demo:
    with gr.Row(scale=0.5):
        gr.HTML(
            """<h1 align="center">AI Meme Generator <span style="font-size: 13px;">powered by <a href="https://huggingface.co/blog/idefics">IDEFICS</a></span><img width=40 height=40 src="https://cdn-uploads.huggingface.co/production/uploads/624bebf604abc7ebb01789af/v770xGti5vH1SYLBgyOO_.png" /></h1>"""
        )
        
    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):
        with gr.Column(equal_height=1):
            imagebox = gr.Image(
                type="filepath", label="Image to meme", height=400, visible=True
            )
            with gr.Group():
                with gr.Row():
                    textbox.render()
                with gr.Row():
                    submit_btn = gr.Button(value="▶️ Submit", visible=True)
                    clear_btn = gr.ClearButton(
                        [textbox, imagebox, chatbot], value="🧹 Clear"
                    )
                    regenerate_btn = gr.Button(value="🔄 Regenerate", visible=True)
                    upload_btn = gr.UploadButton(
                        "📁 Upload image", file_types=["image"], visible=False
                    )
            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=2):
            generated_memes_gallery = gr.Gallery(
                # value="Images generated will appear here",
                label="Generated Memes",
                allow_preview=True,
                elem_id="generated_memes_gallery",
                show_download_button=True,
                show_share_button=True,
            ).style(columns=[2], object_fit="contain", height=600)
    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= value given by gallery_type_choice,
                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(
                """
                <p><strong>AI Meme Generator</strong> 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.</p>
                <p>AI Meme Generator is a space inspired from <a href="https://huggingface.co/spaces/HuggingFaceM4/ai_dad_jokes">AI Dad Jokes</a> and powered by <a href="https://huggingface.co/blog/idefics">IDEFICS</a>, 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.</p>

                <p>⛔️ <strong>Intended uses and limitations:</strong> This demo is provided as research artifact to the community showcasing IDEFICS'capabilities. We detail misuses and out-of-scope uses <a href="https://huggingface.co/HuggingFaceM4/idefics-80b#misuse-and-out-of-scope-use">here</a>. 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.</p>
            """
            )
    with gr.Row():
        chatbot.render()

    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}"},
        )

        # 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

        if image is None:
            # Case where there is no image OR the image is passed as `<fake_token_around_image><image:IMAGE_URL><fake_token_around_image>`
            chat_history.append([prompt_list_to_markdown(user_prompt_list), ""])
        else:
            # Case where the image is passed through the Image Box.
            # Convert the image into base64 for both passing it through the chat history and
            # displaying the image inside the same bubble as the text.
            chat_history.append(
                [
                    f"{prompt_list_to_markdown([image] + user_prompt_list)}",
                    "",
                ]
            )

        query = prompt_list_to_tgi_input(formated_prompt_list)
        all_meme_images = []
        for i in range(4):
            stream = client.generate_stream(prompt=query, **generation_args)

            acc_text = ""
            full_text = ""
            for idx, response in enumerate(stream):
                text_token = response.token.text

                if response.details:
                    # That's the exit condition
                    if image is not None and full_text != "":
                        meme_image = make_meme_image(
                            image=image,
                            text=full_text,
                            font_meme_text=font_meme_text,
                            all_caps_meme_text=all_caps_meme_text,
                            text_at_the_top=text_at_the_top,
                        )
                        meme_image = pil_to_temp_file(meme_image)
                        all_meme_images.append(meme_image)
                        yield user_prompt_str, all_meme_images, chat_history
                    if i == 3:
                        return

                if text_token in STOP_SUSPECT_LIST:
                    acc_text += text_token
                    continue

                if idx == 0 and text_token.startswith(" "):
                    text_token = text_token.lstrip()

                acc_text += text_token
                # Commented to not have a chatbot history that could confuse user

                # last_turn = chat_history.pop(-1)
                # last_turn[-1] += acc_text
                # if last_turn[-1].endswith("\nUser"):
                #     # Safeguard: sometimes (rarely), the model won't generate the token `<end_of_utterance>` and will go directly to generating `\nUser:`
                #     # It will thus stop the generation on `\nUser:`. But when it exits, it will have already generated `\nUser`
                #     # This post-processing ensures that we don't have an additional `\nUser` wandering around.
                #     last_turn[-1] = last_turn[-1][:-5]
                # chat_history.append(last_turn)
                # yield "", None, chat_history
                full_text += acc_text
                acc_text = ""

    textbox.submit(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],
    )
    imagebox.upload(fn=lambda: "", inputs=[], outputs=[generated_memes_gallery], queue=False).then(
        fn=lambda: "Write a meme about this image.", inputs=[], outputs=[textbox], 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,
        ],
    )
    submit_btn.click(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,
        ],
    )

    def remove_last_turn(chat_history):
        if len(chat_history) == 0:
            return gr.Update(), gr.Update()
        last_interaction = chat_history[-1]
        chat_history = chat_history[:-1]
        last_interaction[0] = re.sub(r"!\[]\(/file=.*?\)", "", last_interaction[0])
        chat_update = gr.update(value=chat_history)
        text_update = gr.update(value=last_interaction[0])
        return chat_update, text_update, ""

    regenerate_btn.click(
        fn=remove_last_turn,
        inputs=chatbot,
        outputs=[chatbot, textbox, generated_memes_gallery],
    ).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,
        ],
    )

    upload_btn.upload(add_file, [upload_btn], [imagebox, upload_btn], queue=False)
    submit_btn.click(
        lambda: gr.update(label="📁 Upload image", interactive=True), [], upload_btn
    )
    textbox.submit(
        lambda: gr.update(label="📁 Upload image", interactive=True), [], upload_btn
    )
    clear_btn.click(
        lambda: gr.update(label="📁 Upload image", interactive=True), [], upload_btn
    )
    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],
    ).success(
        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,
        ],
    )
    demo.load(
        fn=choose_gallery, inputs=[gallery_type_choice], outputs=[template_gallery]
    )
demo.queue(concurrency_count=40, max_size=40)
demo.launch()