app.py

import base64
from dataclasses import dataclass
from io import BytesIO
from pathlib import Path
from typing import Literal, cast

import gradio as gr
import jinja2
from openai import OpenAI
from PIL import Image
from pydantic import BaseModel

client = OpenAI()

TEMPLATES_DIR = Path(__file__).resolve().parent / "templates"
jinja_env = jinja2.Environment(loader=jinja2.FileSystemLoader(str(TEMPLATES_DIR)))

SYSTEM_PROMPT = "You are expert prompt engineer"

StyleName = Literal[
    "General",
    "Fashion",
    "Emotional Lifestyle",
    "Extreme Sports",
    "Captivating",
    "Image Replication",
    "Red Bar Lighting",
    "Teal Noir",
]


@dataclass(frozen=True)
class StyleDefinition:
    name: StyleName
    template_filename: str
    info: str


STYLE_DEFINITIONS: dict[StyleName, StyleDefinition] = {
    "General": StyleDefinition(
        name="General",
        template_filename="general_prompt.jinja",
        info="Versatile, balanced storytelling with cinematic detail for most scenarios.",
    ),
    "Fashion": StyleDefinition(
        name="Fashion",
        template_filename="fashion_prompt.jinja",
        info="Editorial fashion aesthetic highlighting garments, styling, and runway polish.",
    ),
    "Emotional Lifestyle": StyleDefinition(
        name="Emotional Lifestyle",
        template_filename="emotional_lifestyle_prompt.jinja",
        info="Warm, candid lifestyle imagery that focuses on mood, relationships, and feelings.",
    ),
    "Extreme Sports": StyleDefinition(
        name="Extreme Sports",
        template_filename="extreme_sports_prompt.jinja",
        info="High-adrenaline action shots that emphasize energy, motion, and athletic feats.",
    ),
    "Captivating": StyleDefinition(
        name="Captivating",
        template_filename="captivating_prompt.jinja",
        info="Visually striking compositions with dramatic flair and memorable storytelling.",
    ),
    "Image Replication": StyleDefinition(
        name="Image Replication",
        template_filename="image_replication_prompt.jinja",
        info=(
            "Mimic the reference image's composition, lighting, and styling exactly while"
            " inserting the user or their face in place of the original subject. Eg. If the reference image is a music album cover, the user's face will be embedded in the album cover."
        ),
    ),
    "Red Bar Lighting": StyleDefinition(
        name="Red Bar Lighting",
        template_filename="red_bar_lighting_prompt.jinja",
        info="Red bar lighting style for image generation.",
    ),
    "Teal Noir": StyleDefinition(
        name="Teal Noir",
        template_filename="teal_noir_prompt.jinja",
        info="Teal noir style for image generation.",
    )
}

PROMPT_TEMPLATES = {
    style: jinja_env.get_template(config.template_filename)
    for style, config in STYLE_DEFINITIONS.items()
}

DEFAULT_STYLE: StyleName = "General"
STYLE_CHOICES: tuple[StyleName, ...] = tuple(STYLE_DEFINITIONS.keys())

STYLE_INFORMATION_BLOCK = "\n".join(
    f"- {style}: {config.info}" for style, config in STYLE_DEFINITIONS.items()
)


class StyleSelectionResponse(BaseModel):
    style: StyleName
    

def process_prompt(user_image, reference_image, target_label: str, user_prompt: str, style: StyleName) -> str:
    user_image_url = None
    reference_image_url = None

    if user_image is not None:
        buffer = BytesIO()
        user_image.convert("RGB").save(buffer, format="JPEG", quality=90)
        b64_image = base64.b64encode(buffer.getvalue()).decode("utf-8")
        user_image_url = f"data:image/jpeg;base64,{b64_image}"

    if reference_image is not None:
        buffer = BytesIO()
        reference_image.convert("RGB").save(buffer, format="JPEG", quality=90)
        b64_reference_image = base64.b64encode(buffer.getvalue()).decode("utf-8")
        reference_image_url = f"data:image/jpeg;base64,{b64_reference_image}"

    try:
        template = PROMPT_TEMPLATES[style]
    except KeyError as error:
        raise ValueError(f"Unsupported style: {style}") from error

    user_content = template.render(user_prompt=user_prompt)

    content = [{"type": "input_text", "text": user_content}]

    if user_image_url is not None:
        content.append({"type": "input_image", "image_url": user_image_url})
    if reference_image_url is not None:
        content.append({"type": "input_image", "image_url": reference_image_url})

    response = client.responses.create(
        model="gpt-5",
        reasoning={"effort": "minimal"},
        input=[
            {
                "role": "system",
                "content": SYSTEM_PROMPT,
            },
            {
                "role": "user",
                "content": content,
            }
        ],
    )
    return f"{response.output_text} {target_label.strip()}"


def recommend_style(user_prompt: str, reference_image: Image.Image | None) -> StyleSelectionResponse:
    if reference_image is not None:
        buffer = BytesIO()
        reference_image.convert("RGB").save(buffer, format="JPEG", quality=90)
        b64_reference_image = base64.b64encode(buffer.getvalue()).decode("utf-8")
        reference_image_url = f"data:image/jpeg;base64,{b64_reference_image}"
    else:
        reference_image_url = None


    user_prompt = f"""You are an art director who must pick the most fitting style name for a user's prompt. 
Consider the available styles and choose the single best option. User has provided the reference image.

Style Guide:
{STYLE_INFORMATION_BLOCK}

User Prompt:
{user_prompt}
"""
    content = [{"type": "input_text", "text": user_prompt}]
    if reference_image_url is not None:
        content.append({
            "type": "input_image", "image_url": reference_image_url
        })
    completion = client.responses.parse(
        model="gpt-5-mini",
        reasoning={"effort": "low"},
        input=[{
            "role": "user",
            "content": content,
        }],
        text_format=StyleSelectionResponse,
    )
    return completion.output_parsed.style


def handle_auto_style_toggle(auto_enabled: bool) -> dict[str, object]:
    return gr.update(interactive=not auto_enabled)


def generate_prompt_handler(
    user_image,
    reference_image,
    target_label: str,
    user_prompt: str,
    current_style: str | None,
    auto_style_enabled: bool,
):

    if auto_style_enabled:
        current_style = recommend_style(user_prompt, reference_image)

    prompt_text = process_prompt(
        user_image=user_image,
        reference_image=reference_image,
        target_label=target_label,
        user_prompt=user_prompt,
        style=current_style,
    )
    display_text = f"Selected style: {current_style}\n\n{prompt_text}"
    return display_text, gr.update(value=current_style, interactive=False)


with gr.Blocks() as demo:
    with gr.Row():
        with gr.Column():
            user_image = gr.Image(
                label="Upload user photo",
                type="pil"
            )
            reference_image = gr.Image(
                label="Optional: Upload reference image (Eg. movie poster, music album cover, etc.)",
                type="pil",
            )
            target_label = gr.Textbox(
                label="Enter target label",
                placeholder="SMRA",
            )
            user_prompt = gr.Textbox(
                label="Enter your prompt",
                placeholder="picture of me while sitting in a chair in the ocean",
                lines=4,
            )
            style_dropdown = gr.Dropdown(
                choices=list(STYLE_CHOICES),
                value=DEFAULT_STYLE,
                label="Style Selection",
                info="Choose the visual style for your enhanced prompt",
                interactive=False,
            )
            auto_style_checkbox = gr.Checkbox(
                label="Auto-select best style",
                value=True,
            )
            generate_button = gr.Button("Generate Prompt")
        with gr.Column():
            prompt_output = gr.Textbox(
                label="Style Prompt",
                lines=20,
            )

    generate_button.click(
        generate_prompt_handler,
        inputs=[
            user_image,
            reference_image,
            target_label,
            user_prompt,
            style_dropdown,
            auto_style_checkbox,
        ],
        outputs=[prompt_output, style_dropdown],
    )
    auto_style_checkbox.change(
        handle_auto_style_toggle,
        inputs=[auto_style_checkbox],
        outputs=[style_dropdown],
    )

demo.launch()