QIE_prompt_generator.py

import base64
import mimetypes
from pathlib import Path
import gradio as gr
from typing import Optional, Tuple

# ===== ユーティリティ =====

def file_to_data_url(path: str) -> Optional[str]:
    """画像ファイルを data URL に変換"""
    if not path:
        return None
    p = Path(path)
    if not p.exists():
        return None
    mime, _ = mimetypes.guess_type(str(p))
    if not mime:
        mime = "image/png"
    with open(p, "rb") as f:
        b64 = base64.b64encode(f.read()).decode("utf-8")
    return f"data:{mime};base64,{b64}"


# ===== Meta Prompt =====
META_PROMPT = """You are an AI prompt generator for image-to-image transformation tasks in art and illustration pipelines.
Given two images — A (input) and B (output) — and an optional description or notes, your goal is to write a precise, structured English prompt that fully explains how A transforms into B, including both conceptual and visual rules of the transformation.

Instructions:
1) Write 3–6 concise sentences describing the transformation.
2) Begin with “The [type of A] transforms into …” or “Convert the [A] into …”.
3) Clearly describe:
   - What disappears or remains (e.g., hair, clothes, shadows, lines).
   - What structural or stylistic simplifications occur (e.g., box, sphere, guide lines).
   - How lines, colors, or lighting change (e.g., colored → lineart, flat → shaded).
   - Any rules for anatomy, proportion, or rendering (e.g., keep pose, maintain base colors).
   - Background or presentation constraints (e.g., white background, no shading).
4) Integrate any user-provided notes naturally into the text.

Output Format:
- English Prompt: A full paragraph (3–6 sentences) describing the transformation from A to B.
- Name Suggestions: Propose 3 short, descriptive task-name candidates (e.g., Image2Body, Body2Box, Sketch2Line, Flat2Shade, Guide2Color, Light2Render, etc.).
- Optional Japanese Translation: Provide a brief Japanese version of the English prompt for understanding if requested.

Make the wording professional, objective, and consistent with technical art pipeline prompts.
"""


# ===== OpenAI 呼び出し =====
def call_openai_chat(api_key: str, a_data_url: Optional[str], b_data_url: Optional[str], notes: str, want_japanese: bool):
    """OpenAI GPT-5 API 呼び出し"""
    if not api_key:
        return ("", "", "（エラー：API Key が未入力です）")

    try:
        from openai import OpenAI
    except Exception:
        return ("", "", "（エラー：openai パッケージが見つかりません。`pip install openai` を実行してください）")

    client = OpenAI(api_key=api_key)
    model = "gpt-5"

    user_content = []
    if a_data_url:
        user_content.append({"type": "text", "text": "Image A (input):"})
        user_content.append({"type": "image_url", "image_url": {"url": a_data_url}})
    if b_data_url:
        user_content.append({"type": "text", "text": "Image B (output):"})
        user_content.append({"type": "image_url", "image_url": {"url": b_data_url}})
    user_content.append({
        "type": "text",
        "text": f"Additional notes: {notes or '(none)'}\n{'Include Japanese translation.' if want_japanese else 'No Japanese translation.'}"
    })

    try:
        resp = client.chat.completions.create(
            model=model,
            messages=[
                {"role": "system", "content": META_PROMPT},
                {"role": "user", "content": user_content},
            ]
        )
        text = resp.choices[0].message.content.strip()
    except Exception as e:
        return ("", "", f"（APIエラー：{e}）")

    # 簡易パース
    english, names, japanese = "", "", ""
    lower = text.lower()
    if "name suggestions:" in lower:
        split = text.split("English Prompt:")[1].split("Name Suggestions:")
        english = split[0].strip()
        rest = split[1]
        if "Japanese Translation:" in rest:
            name_part, jp_part = rest.split("Japanese Translation:", 1)
            names = name_part.strip()
            japanese = jp_part.strip()
        else:
            names = rest.strip()
    else:
        english = text

    if not want_japanese:
        japanese = ""

    return english, names, japanese


# ===== Gradio UI (standalone) =====

def build_ui() -> gr.Blocks:
    with gr.Blocks(title="A→B 変換プロンプト自動生成（GPT-5固定）") as demo:
        gr.Markdown("""
# 🎨 A→B 変換プロンプト自動生成
画像A（入力）と画像B（出力）、補足説明を入力すると、  
A→B の変換内容を**英語プロンプト**として自動生成し、  
さらに**タスク名候補（3件）**を提案します。  
モデルは `gpt-5` を使用します。
""")

        api_key = gr.Textbox(label="OpenAI API Key", type="password", placeholder="sk-...")
        with gr.Row():
            img_a = gr.Image(type="filepath", label="Image A (Input)", height=300)
            img_b = gr.Image(type="filepath", label="Image B (Output)", height=300)

        notes = gr.Textbox(label="補足説明（日本語可）", lines=4, placeholder="例）髪・服・背景は消す。目は四角。鎖骨はピンク線。など", value="この画像は例であって、汎用的なプロンプトにする")
        want_japanese = gr.Checkbox(label="日本語訳を含める", value=True)
        run_btn = gr.Button("生成する", variant="primary")

        english_out = gr.Textbox(label="English Prompt", lines=8)
        names_out = gr.Textbox(label="Name Suggestions", lines=4)
        japanese_out = gr.Textbox(label="日本語訳（任意）", lines=8)

        def on_click(api_key_in, a_path, b_path, notes_in, ja_flag):
            a_url = file_to_data_url(a_path) if a_path else None
            b_url = file_to_data_url(b_path) if b_path else None
            return call_openai_chat(api_key_in, a_url, b_url, notes_in, ja_flag)

        run_btn.click(
            fn=on_click,
            inputs=[api_key, img_a, img_b, notes, want_japanese],
            outputs=[english_out, names_out, japanese_out],
        )

    return demo


if __name__ == "__main__":
    build_ui().launch()